# Navōn — Filtered AI for schools and families > Full text of every guide published at https://navon.org/blog. Generated 2026-09-23. --- # What makes an AI kosher? Source: https://navon.org/blog/what-makes-an-ai-kosher Published: 2026-09-03 · Category: Explainers No one certifies software the way they certify food. Here is the standard worth applying instead, and the questions that separate a filtered AI from one that merely promises to behave. There is no hechsher for software, and anyone offering one should be treated with suspicion. Kashrus is a system of supervision applied to a physical process. Software is not that. But the instinct behind the question is exactly right: before something enters a home or a classroom, someone responsible should have checked it against a known standard, and you should be able to find out what that standard was. So the useful question is not "is this AI kosher" but "what is checked, who set the rules, and what happens when the check fails". Those three questions separate a genuinely filtered tool from one that is simply well-behaved most of the time. ## 1. Where does the filtering run? This is the single most important question, and the easiest to get a straight answer to. Filtering that runs on the device — a browser extension, a kiosk mode, a managed profile — protects that device. A student who picks up a phone, uses a friend's laptop, or uninstalls the extension is outside it entirely. Filtering that runs on the server is a property of the account, not the machine. The same rules apply on a school Chromebook at 10am and a phone at 11pm, because the check happens before the request ever reaches a model. Ask a vendor plainly: if a student signs in from a device you have never seen, what still applies? If the answer involves installing something, the protection has a hole in it by design. ## 2. Who sets the standard — and can you see it? A general-purpose AI has a safety policy written by its vendor for a global audience. It will refuse the obvious things. It will not refuse everything a frum school would want refused, because it was never asked to. What matters for a school is whether the standard is yours: whether you can state the topics that are off limits for your talmidim, whether that list is applied to every tool and not only to chat, and whether you can read back what it did. A filter you cannot inspect is a promise, not a control. ## 3. What happens when the check cannot run? This is the question almost nobody asks, and it is where the real difference shows. Every content filter depends on something — a model, an API, a network call. That dependency will occasionally be unavailable. There are only two possible behaviours. The system can fail open, letting the request through unchecked so the product keeps working. Or it can fail closed, refusing the request until the check is available again. > A filter that fails open is not a filter. It is a filter-shaped feature that switches itself off exactly when nobody is watching. Failing closed costs something real: during an outage, students cannot work. That is the correct trade for a tool aimed at children, and it is worth asking any vendor to state their answer in writing. ## 4. Does it cover every tool, or only the chat box? Filtering is usually built for text first, because text is easiest. Image generation, video, voice and story tools often get added later, and the filter does not always follow them. The gap is not theoretical. A student blocked from discussing a topic in chat, who can generate a picture of it, has not been protected — they have been redirected. When you evaluate a platform, test the same restricted subject in every tool it offers, not just the one on the front page. ## 5. Can a parent or menahel see what happened? Oversight is not surveillance of every keystroke, and a system that reads like a wiretap will not be adopted. What a responsible adult needs is narrower: was anything blocked, roughly what category was it, and is there a pattern worth a conversation. A platform that blocks silently and keeps no record leaves the adult with nothing to act on. One that stores every prompt in full creates a different problem. The useful middle is a record of decisions, with enough context to be meaningful and not so much that it becomes a copy of every child's private questions. ## A short checklist 1. Does the filter run on the server, so it follows the student off the school's devices? 2. Can the school define its own blocked topics, above the vendor's baseline? 3. Do those rules apply to images and video, not only to chat? 4. Does the system fail closed when the safety check is unavailable? 5. Can a parent or administrator see what was blocked, without reading everything? 6. Are the categories that must never be switched off actually locked, or just unchecked by default? Any vendor should be able to answer all six in a sentence each. If a question produces a long detour, that is your answer. [See how the filtering works](https://navon.org/features) ## Common questions **Is there a hechsher or kosher certification for AI software?** No. Kashrus certification applies to a physical production process and there is no equivalent body certifying software. The practical alternative is to check specific properties: where the filtering runs, who sets the standard, whether it covers every tool, and what happens when the check fails. **Is ChatGPT kosher for students?** General-purpose assistants apply a safety policy written by their vendor for a global audience. They refuse obviously harmful content but were never asked to apply a particular community's standards, and a school cannot add its own blocked topics or see what was refused. **What is the most important thing to check?** Where the filtering runs. Device-based filtering protects a device; server-side filtering is a property of the account and applies wherever the student signs in. --- # Server-side vs device filtering: why extensions do not hold Source: https://navon.org/blog/server-side-vs-device-filtering Published: 2026-09-03 · Category: Safety Most school filtering protects a machine. Students do not use one machine. The difference between filtering a device and filtering an account decides whether your rules survive contact with real life. Almost every filtering product a school has been sold works the same way: it sits on the device. A browser extension, a managed Chrome profile, a DNS filter on the school network, a kiosk-mode tablet. They differ in detail and share one property — they protect the machine they are installed on. That was a reasonable design when the machine was the point of access. It is a poor fit for AI, for a specific reason worth spelling out. ## The account is the thing being used, not the computer A student signs in to an AI tool. The conversation, the history, the generated images — all of it belongs to the account and follows it everywhere. A filter attached to the school laptop knows nothing about what that same account did on a phone at home. So a school that has filtered its devices has filtered roughly a third of a student's AI use: the third that happens in the building, on equipment the school owns, during the hours it is open. ## Three ways device filtering comes apart ### The second device The most common and least sophisticated bypass. Nothing was hacked; the student simply used a phone. Any protection that depends on which hardware is in front of the child is defeated by the child owning more than one piece of hardware. ### Removing the extension Managed profiles make this harder, not impossible, and the skill required is a web search. More importantly, a student who removes an extension usually leaves no trace that the protection stopped applying — the school's dashboard shows a device that has simply gone quiet. ### Filtering the wrong layer Network and DNS filters decide whether a site can be reached. They cannot see inside an encrypted conversation with a site they have allowed. Once a school permits an AI tool at all, a DNS filter has no further opinion about what is asked of it. The choice becomes all or nothing, which is why so many schools ended up blocking AI entirely. ## What server-side filtering changes If the check runs on the server, between the student's request and the model, then it is a property of the account. There is nothing installed to remove and no device to swap. Concretely, the request arrives, the platform loads the rules that apply to that student — the school's policy, plus anything their parents added — and evaluates the request against them before any model is asked to answer. A request that fails is refused and recorded. The same sequence happens whether the student is on a school Chromebook or a phone on their own wifi. This also makes the rules legible. Because the decision happens in one place, a school can see what was blocked without installing monitoring software on anybody's personal phone — which is both more effective and considerably less invasive than the alternative. ## What server-side filtering does not solve It is worth being precise, because overstating this is how trust gets lost. Server-side filtering governs one platform. It does not stop a student opening a different AI tool in another tab. Nothing that runs inside one product can. What it does is make one tool trustworthy enough to allow — so the school has something to point students toward, rather than only a list of things to forbid. Schools that give students a permitted option see less traffic to the unpermitted ones. A prohibition with no alternative mostly teaches students to be discreet. [How the filter works](https://navon.org/features) ## Common questions **Can a student uninstall server-side filtering?** No. There is nothing installed on the device to remove. The check runs on the server before a request reaches a model, so it applies to the account wherever the student signs in. **Does a DNS or network filter do the same job?** No. Network filters decide whether a site can be reached. Once an AI tool is allowed, they cannot see inside the encrypted conversation, so they cannot distinguish an acceptable request from an unacceptable one. **Does server-side filtering stop a student using a different AI tool?** No, and no in-product filter can. It makes one tool trustworthy enough to permit, which gives the school an option to point students toward instead of only a prohibition. --- # Should schools allow AI? A practical answer Source: https://navon.org/blog/should-schools-allow-ai Published: 2026-09-03 · Category: For schools The choice is not between AI and no AI. It is between AI you can see and AI you cannot. Here is how to think it through without pretending either extreme works. Most schools have already made this decision twice. First they blocked AI, because it appeared suddenly and the risks were obvious before the uses were. Then they discovered the block had changed where students used it rather than whether they used it. That is the honest starting point. A prohibition a school cannot enforce is not a policy; it is a statement of preference. The useful question is what a school can actually control. ## What a ban does and does not do A ban does real work. It sets an expectation, it keeps AI out of lessons where it would undermine the point of the exercise, and it gives staff a clear line to hold. None of that is nothing. What a ban does not do is prevent use. Students have phones and home computers. The measurable effect of a ban is that use moves to tools the school cannot see, on devices it does not control, with no filtering appropriate to its community — and with a strong incentive for students not to mention it. > A ban relocates the behaviour. It does not remove it, and it removes the school's visibility of it. ## The three questions worth settling first ### 1. Where is AI genuinely inappropriate? Some assessments measure whether a student can do a thing unaided. AI use there defeats the purpose, and no filtering setting fixes that — it is a matter of assessment design and supervision. Decide which work is meant to be unaided and say so plainly, per assignment rather than as a blanket rule. ### 2. Where is it obviously useful? Explaining a concept a different way, drafting and critiquing, practice questions, translation, summarising a long text. These are the cases where refusing to engage puts your talmidim behind peers who have the tool at home anyway. ### 3. What must never appear, in any context? This is the part a general-purpose assistant will not do for you. Every community has content it will not have in front of its children, and that list is more specific than any vendor's default safety policy. It only becomes enforceable if the tool lets you state it and applies it everywhere. ## A workable middle position 1. Permit one AI tool that filters on the server, so the rules follow students home rather than stopping at the school gate. 2. Write the school's own blocked topics into it, rather than relying on a vendor's global defaults. 3. Say per assignment whether AI is permitted, expected, or forbidden — not once, as a blanket rule. 4. Teach students to check what it tells them. This is a skill, and it is the one that transfers. 5. Review what was blocked periodically. Patterns are more useful than individual incidents. 6. Tell parents what you have done. A parent who knows the school provides a filtered option is far less likely to hand a child an unfiltered one. ## On honesty with students Students know when a rule is unenforceable, and a rule everyone knows is unenforceable costs the school credibility on the rules that matter. A school that says "here is the tool we permit, here is what it will not do, and here is where you must work unaided" is describing something true. That is a stronger position than a prohibition that quietly does not hold. [Navōn for schools](https://navon.org/services/schools) ## Common questions **Does banning AI in school work?** It sets expectations and keeps AI out of work meant to be unaided, but it does not prevent use. Students use phones and home computers, so a ban mainly moves use to tools the school cannot see or filter. **What should a school allow AI to be used for?** Explanation, drafting and critique, practice questions, translation and summarising are the common useful cases. Work that is specifically meant to measure unaided ability is where AI use defeats the purpose. **How do we stop AI being used to cheat?** Filtering does not solve this; assessment design does. Decide per assignment whether AI is permitted, expected or forbidden, and say so explicitly rather than relying on a blanket rule. --- # Writing an AI acceptable use policy for your school Source: https://navon.org/blog/ai-acceptable-use-policy-for-schools Published: 2026-09-03 · Category: Policy Most AI policies fail in the same two ways: they are written as prohibitions nobody can enforce, or so vaguely that no teacher can apply them on a Tuesday. Here is a structure that avoids both. A policy exists so that a teacher facing a decision at 9am, and a parent asking a question at 9pm, get the same answer. Judge every sentence you write against that. This is a structure, not a template to adopt unread. The decisions inside it are yours, and a policy copied without those decisions being made is the kind that sits in a folder. ## 1. Scope: say what this covers State which tools, which people, and which settings. Be explicit about school-owned versus personal devices and about school hours versus homework, because that is the first thing anyone will argue about. A common trap: writing the policy so it applies only on school devices. That is the easy scope to enforce and it excludes most of the actual use. ## 2. Permitted, expected, forbidden Three categories, not two. Most policies have permitted and forbidden and end up unusable, because the interesting cases are the ones where a teacher wants to require AI use — critique this draft, generate practice questions — and there is no category for it. Then push the decision down to the assignment. A blanket rule cannot distinguish an essay testing unaided argument from a revision exercise where AI is the point. Give staff a standard phrase for each category to put on the assignment itself. ## 3. Content standards This is where a school-specific policy earns its keep. Name the categories your community will not have in front of its children. Do not gesture at "inappropriate content" — a teacher cannot apply that and a filter cannot be configured from it. Then say where the standard is enforced. If it is enforced in the tool, say which, and confirm it applies to images and video as well as to chat. If it is enforced only by expectation, say that too, honestly. ## 4. Attribution and honesty Decide what a student must disclose and in what form. "Acknowledge AI use" is too vague to comply with. A single required sentence — which tool, for what part, what the student changed — is specific enough to follow and to check. Also decide what happens when a student does disclose. If disclosure is punished, students stop disclosing, and you have built an incentive to conceal. ## 5. Verification State plainly that AI output can be confidently wrong and that the student is responsible for what they submit. This single expectation does more good than most restrictions, because it is the habit that transfers to every tool they will use later. Avoid promising to detect AI writing. Detectors are unreliable in both directions, and a policy that rests on one will eventually accuse a student who did nothing wrong. ## 6. Privacy Say what students must never enter: their own or others' personal details, anything about another child, staff information. Then say what the school can see about their use, in specific terms. Be accurate about the tool's own data handling. If you cannot state whether a vendor trains on submitted content, ask them in writing before you name them in a policy. ## 7. What happens when the rules are broken Proportionate and stated in advance. Distinguish a student exploring from a student concealing; those are different behaviours and treating them identically teaches the wrong lesson. ## Two wording traps - "Students may not use AI to do their work for them." Nobody can apply this. Where is the line between a tool that explains and a tool that drafts? Name the specific activity instead. - "AI use must be appropriate." This puts the whole judgement on the teacher in the moment and gives the student no notice of what was expected. The test for any sentence: could two teachers apply it to the same piece of work and reach the same conclusion? If not, it needs to be more specific. [See the school controls](https://navon.org/services/schools) ## Common questions **What should an AI policy for schools include?** Scope (which tools, people and devices), three use categories rather than two, a specific content standard, an attribution requirement, a verification expectation, privacy rules, and proportionate consequences. **Should schools rely on AI detection tools?** No. Detectors are unreliable in both directions and a policy resting on one will eventually accuse a student who did nothing wrong. Assessment design and required disclosure are more dependable. **Should the policy cover personal devices?** Yes. A policy scoped only to school-owned devices excludes most actual use, which is the enforcement gap that makes many AI policies ineffective. --- # Safe AI for kids: what a parent should actually check Source: https://navon.org/blog/safe-ai-for-kids-what-parents-should-check Published: 2026-09-03 · Category: For families "Safe for kids" is a marketing phrase, not a specification. These are the six things worth checking, and what a good answer looks like for each. Every AI product aimed at families describes itself as safe. The word is doing no work — it has no agreed definition and nobody audits the claim. What follows is what to check instead. None of it requires technical knowledge, and all of it can be done before you hand over an account. ## 1. Can you set the rules, or only accept theirs? Most tools ship one safety policy written by the vendor. It will refuse the obvious. It will not know what your family does not discuss. Look for whether you can add your own topics — and then test one to see whether it actually took effect. ## 2. Do the rules apply everywhere, or just to chat? Test the same subject in every tool the product offers: chat, images, video, story generation. Filters are usually built for text first and the other tools catch up later, or never. A child blocked in chat who can generate a picture of the same thing has found the gap without looking for it. ## 3. Does it work on your child's phone? If the protection is an app setting or a browser extension, it protects that device. Sign in on a different phone and try something that should be refused. If it is allowed there, you have device protection, not account protection. ## 4. What happens when their safety system is down? Worth asking support directly. If the safety check cannot run, does the product refuse to answer, or answer anyway? Both are choices a company makes deliberately. Only one of them is the right choice for a tool a child uses. ## 5. Can you see what happened without reading everything? There is a real tension here. Reading every message a teenager sends is usually the wrong parenting decision and will end the child's honesty with you. Seeing nothing at all leaves you unable to notice a pattern that matters. What you want is the middle: whether anything was blocked, roughly what kind of thing, and how often. Enough to start a conversation, not so much that you are reading a diary. ## 6. Are time limits real limits? If a product offers screen-time controls, set one and then test it after it should have expired. Time limits are one of the most commonly advertised and least commonly enforced features in this category, because they are easy to display and fiddly to implement. ## One thing worth more than all six Sit with your child the first few times they use it. Not to supervise — to see what they ask, which is genuinely interesting, and to be the person they mention it to when something odd comes back. No filter substitutes for being the adult a child tells. [Navōn for families](https://navon.org/services/family) ## Common questions **Is ChatGPT safe for kids?** It applies a general safety policy that refuses obviously harmful content, but parents cannot add their own blocked topics, the rules cannot be extended to a specific family's standards, and there is no parental view of what was refused. **What age is right for a child to use AI?** There is no single answer, but the more useful question is whether the tool lets the responsible adult set the standard and see what happened. A tool a parent can configure and review is appropriate earlier than one they cannot. **Do parental controls on AI tools actually work?** Test them rather than trusting the description. Set a blocked topic and try it, set a time limit and test after it expires, and try the same request on a different device. Advertised controls are not always enforced ones. --- # Is ChatGPT safe for kids? An honest answer Source: https://navon.org/blog/is-chatgpt-safe-for-kids Published: 2026-09-03 · Category: For families The answer is neither yes nor no, and anyone giving you a one-word answer is selling something. Here is what ChatGPT actually refuses, what it does not, and what that means for your child. Parents ask this expecting a yes or a no. The truthful answer is that ChatGPT is a general-purpose tool built for adults, with safety measures aimed at preventing serious harm, used by a very large number of children anyway. Whether that is acceptable depends on the child, the supervision, and what you expect the tool to refuse. What follows is what it genuinely does well, and the four gaps that matter most. ## What it does refuse The obvious categories are handled, and handled reasonably. Explicit sexual content, instructions for violence or weapons, self-harm encouragement, and clearly illegal activity all get refused, and attempts to get around a refusal by rephrasing usually fail. OpenAI has put real work into this and it shows. It is also, in ordinary use, a patient and accurate explainer. For a child stuck on a maths concept at 9pm, it is genuinely better than nothing and often better than a parent who last studied the topic thirty years ago. ## Gap 1: refusing harm is not the same as matching your standards A general assistant applies one safety policy written for a global audience of adults. It will not refuse a topic because your family does not discuss it, because your community holds a different standard, or because your eleven-year-old is not ready for it. Those were never inputs to the policy. So the tool is safe in the sense of not causing catastrophic harm, and unaligned in the sense that its idea of appropriate is not yours and cannot be made to be. ## Gap 2: you cannot see what happened If a child asks something concerning and the model declines, nothing tells you. There is no parental view, no summary, no signal. The conversation history sits in the child's own account, which they can delete. That matters less for the blocked request itself and more for the pattern. A single odd question is noise. The same question three times in a fortnight is a conversation a parent would want to have, and there is no mechanism that would ever surface it. ## Gap 3: it is confidently wrong, and children are poorly placed to notice Every large language model produces fluent, plausible, incorrect statements. Adults with domain knowledge catch these. A child researching a topic for the first time has no basis for catching anything — the wrong answer and the right one arrive in the same authoritative tone. This is the risk most parents underrate, because it is undramatic. It is also the one with a real remedy: teaching a child to check, which is a durable skill. ## Gap 4: age rules that are not enforced The terms require users to be 13, and 18 or parental consent under some jurisdictions. Enforcement is a date-of-birth field. Any child who can subtract can pass it. Treat the stated age limit as a statement of intent, not a control. The practical control is whichever account the child is actually using and whether an adult set it up. ## So what should a parent do? 1. Decide whether the issue is harm or alignment. If you want catastrophic content refused, a mainstream assistant largely does that. If you want your standards applied, no general tool will. 2. Set up the account yourself, so age and settings are not the child's to choose. 3. Use it alongside them for the first few sessions. What they ask is genuinely informative. 4. Teach one habit above all others: check anything that matters before relying on it. 5. If you need your own blocked topics, time limits, or any visibility into what was refused, you need a tool built to provide those. A general assistant will not grow them. None of this means ChatGPT is dangerous. It means it is a general tool, and a general tool cannot enforce a particular family's standard — that is not a flaw, it is what general means. [See how it works for families](https://navon.org/services/family) ## Common questions **Is ChatGPT safe for a 10 year old?** It refuses seriously harmful content, but it applies one global safety policy, has no parental visibility, and its stated age minimum is 13 enforced only by a date field. For a younger child, supervision and a tool the parent can configure matter more than the model's own refusals. **Can parents see what their child asked ChatGPT?** No. There is no parental dashboard or reporting. The history lives in the child's account and can be deleted by them. **What is the biggest risk of children using AI?** Not shocking content, which is largely refused, but confident wrongness. Models state incorrect things fluently, and a child learning a topic for the first time has no basis to catch the error. --- # AI in yeshivas and Jewish day schools Source: https://navon.org/blog/ai-in-yeshivas-and-jewish-day-schools Published: 2026-09-03 · Category: For schools The AI conversation in a yeshiva is not the same conversation as in a public school. The concerns are different, and so are the workable answers. Most published guidance on AI in schools was written for a context that assumes a school's only content concern is age-appropriateness. That guidance is not useless here, but it does not address the questions that actually get asked in a yeshiva or day school. Those questions are usually four. ## 1. Will it put things in front of talmidim that should not be there? This is the first question and the one that decides everything else. A general-purpose assistant applies its vendor's safety policy: it refuses what a large company considers harmful, and permits an enormous amount that a frum school would not want raised with its talmidim. The practical distinction is between a tool with a fixed policy and a tool where the school writes the policy. Only the second can be aligned with a particular kehilla's standards, because only the second is asking. It is worth checking that the standard applies to every tool. Filters are usually built for text first; a talmid blocked from discussing a subject in chat who can generate an image of it has found the gap without trying. ## 2. Is this bittul zman? Sometimes, plainly. A tool that answers instantly invites aimless use, and a phone in a pocket during seder is a problem whatever is on it. But the same tool used to clarify a difficult Tosfos, to check a translation, or to generate practice questions before a bechina is the opposite of bittul zman. The question is not the tool, it is whether its use is directed. This is where time controls earn their place — not as a punishment, but so the tool is available during the hours it is meant to be used and not during the hours it is not. A control that a school can actually set, and that follows the talmid home, is worth more than an instruction that relies on self-discipline at 11pm. ## 3. What about limmudei kodesh specifically? Here caution is warranted, and it should be said plainly: a language model is not a posek and is not a reliable source for halacha. It will produce a fluent, confident, and sometimes wrong answer, including inventing sources that do not exist. A talmid who takes a machine's answer as a psak has been badly served. The defensible uses are narrower and real: translation help, unpacking a difficult passage in one's own words, generating questions to test one's own understanding, organising notes. The line worth teaching is that the model can help you understand what you are reading, and cannot tell you what the halacha is. Any school permitting AI should say this explicitly, because students will not infer it. Fluency reads as authority. ## 4. If we ban it, are we solving anything? A ban is honest about the school's position and it is worth something. But talmidim have phones and older brothers with laptops, and a ban's measurable effect is usually that use moves out of view rather than stopping. There is also a cost that is easy to miss. Students who leave without ever having used these tools in a supervised setting will encounter them anyway, in a workplace or a seminary, with nobody having taught them where the tool is unreliable or where it is inappropriate. A school that permits a filtered tool and teaches its limits is doing more for its talmidim than one that says nothing. > The choice is rarely between AI and no AI. It is between AI a school has shaped and AI it has not seen. ## A workable position - Permit one tool where the school writes the content standard, rather than accepting a vendor's default. - Confirm the standard covers images and video, not only chat. - Use time windows so the tool is available when it should be and not during seder or late at night. - State explicitly that it is not a source for halacha, and why. - Tell parents what the school has permitted. A parent who knows there is a filtered option is less likely to hand over an unfiltered one. [See the school controls](https://navon.org/services/schools) ## Common questions **Can AI be used for limmudei kodesh?** For translation, unpacking a difficult passage, and generating self-test questions, yes. It is not a posek and is not a reliable source for halacha: it produces confident answers and can invent sources, so it must not be treated as authoritative. **Should a yeshiva ban AI entirely?** A ban states the school's position but rarely stops use, since talmidim have their own devices. It also means students never learn the tool's limits in a supervised setting. Permitting one filtered tool and teaching its boundaries is usually more effective. **How do we stop AI being used during seder?** Time windows set by the school, enforced on the server, make the tool unavailable outside permitted hours regardless of which device a talmid uses. --- # How AI content filtering actually works Source: https://navon.org/blog/how-ai-content-filtering-works Published: 2026-09-03 · Category: Explainers Vendors describe filtering as if it were one thing. It is at least three, they fail in different ways, and the differences decide what your students can actually get to. "Content filtering" covers several quite different techniques. Knowing which one a product uses tells you more about what it will miss than any marketing page will. ## Keyword and pattern matching The oldest approach: a list of forbidden words and phrases, matched against the request. It is fast, cheap, completely predictable, and easy to audit — you can read the list. It fails in both directions, badly. It misses anything phrased around the list, which takes a student roughly one attempt to discover. And it blocks legitimate work constantly: the classic example is a biology lesson that cannot mention anatomy. A filter that blocks real schoolwork trains students to route around it, which is worse than not having it. Where it genuinely belongs is a short list of specific terms a school wants caught exactly, evaluated locally and instantly — not as the main line of defence. ## Classifiers A model trained to score text against categories — violence, sexual content, harassment. Better than keywords because it responds to meaning rather than spelling, so rephrasing does not defeat it. The limits are that categories are fixed by whoever trained it, thresholds are a blunt dial, and it cannot express a rule like "this school does not discuss this particular subject". You get the categories on offer. ## Model-based policy screening The current approach: a separate model is given the policy in writing and asked to judge each request against it. Because the policy is text, it can be specific in a way a fixed category list cannot — including a particular school's own topics. It is slower and costs a call per request, and it is not perfect: judgement calls near the boundary go both ways. What it buys is expressiveness. This is the only approach where "our school does not discuss X" is a rule the system can actually apply. ## Before or after? Two separate checks, and a serious product does both. Screening the request before generation stops the obvious cases cheaply and means an unacceptable request never reaches a generating model at all. Screening the output after generation catches what the request did not reveal — an innocuous prompt can produce an unsuitable answer, particularly for images. A product that only screens input is betting that a bad answer never follows a reasonable-looking question. That is not a safe bet. ## The question that matters more than any of this Every technique above depends on something that can be unavailable — a model, an API, a network hop. So the decisive question is what the system does when the check cannot run. Fail open means the request proceeds unchecked, so the product stays up and the filtering silently stops. Fail closed means the request is refused until checking works again, so the filtering holds and the product is briefly unavailable. > A filter that fails open is not a filter. It is a filter-shaped feature that switches itself off exactly when nobody is watching. For a tool used by children, failing closed is the only defensible choice, and it costs something real. Ask any vendor which they do and expect a straight answer. ## And whose rules are they? A technically excellent filter enforcing someone else's standard still is not your filter. The two questions to settle are whether you can add your own topics, and whether they apply to every tool rather than just the chat box. [See what runs on every request](https://navon.org/features) ## Common questions **What is the difference between a keyword filter and an AI content filter?** A keyword filter matches forbidden words, so it misses rephrasing and blocks legitimate work like biology lessons. A model-based filter judges meaning against a written policy, so it survives rephrasing and can express a specific school's rules. **Should a filter check the answer as well as the question?** Yes. An innocuous prompt can produce an unsuitable answer, especially for images. A product that screens only the input is assuming a bad output never follows a reasonable question. **What does fail closed mean?** That when the safety check itself cannot run, the request is refused rather than allowed through unchecked. The alternative, failing open, means filtering silently stops during an outage. --- # Teaching students to check what AI tells them Source: https://navon.org/blog/teaching-students-to-check-ai-output Published: 2026-09-03 · Category: For schools Of everything schools might teach about AI, one habit outlasts the rest: checking. Here is a way to teach it that takes a single lesson and actually sticks. Most AI guidance for schools is about restriction. Restriction is necessary and it ages badly — the tools change, the rules chase them. One thing does not age: a student who checks before relying on an answer is protected from every model, including ones that do not exist yet. ## Why students do not check on their own Not laziness. Fluency reads as competence. A model's wrong answer arrives in the same measured, confident register as its right one, with no hedging and no visible seams. Every social cue we use to gauge reliability in a human — hesitation, qualification, admitted ignorance — is absent. Students are not being careless when they trust it. They are applying a heuristic that works on people to something the heuristic was never built for. Say that out loud to them; it lands better than being told to be sceptical. ## The lesson that works Ask the class a question in your subject where you know the model produces something plausible and wrong. Specific dates, obscure attributions, precise statistics and citations are reliable places to look. Project the answer. Then have them verify it against a real source, and find the error themselves. The moment a student catches a confident machine being wrong is worth more than any amount of instruction about limitations, because they now have a memory rather than a rule. Run it once a term with a fresh example. It stays true as the models improve, because the failure mode is not going away — it is getting subtler, which makes the habit more valuable, not less. ## A rule students can actually remember Long lists of guidance do not survive contact with a deadline. One sentence does: > If it matters, check it somewhere else before you use it. "If it matters" is doing deliberate work. Nobody verifies a brainstorm or a rephrasing, and demanding that guarantees the rule gets ignored entirely. Facts, figures, quotations, citations and anything going into submitted work — those matter. ## Watch for invented sources Worth teaching specifically, because it surprises people. Models generate citations that look correct in every respect — plausible author, plausible title, plausible year, plausible journal — and simply do not exist. The format is learned; the reference is not retrieved. The rule follows directly: a citation you have not opened is not a citation. If a student cannot produce the source, it does not go in the work. ## Make disclosure safe If admitting AI use is punished, students conceal it, and you lose the visibility that would let you teach anything. Ask for one line — which tool, for which part, what they changed — and treat an honest line as compliance rather than as a confession. The students you most want to reach are the ones who would otherwise never mention it. [Navōn for schools](https://navon.org/services/schools) ## Common questions **Why does AI make things up?** A language model predicts plausible continuations rather than retrieving facts. A fabricated citation is generated in the correct format because the format was learned, even though the reference was never looked up. **How do I teach students to spot AI errors?** Ask a question in your subject where the model produces something plausible and wrong, project the answer, and have students find the error against a real source. Catching a confident machine being wrong teaches more than instruction about limitations. **Should students be allowed to cite AI as a source?** No. It is not a source; it does not retrieve, it generates. Students should verify claims against real sources and cite those, and a citation they have not opened should not go into the work. --- # 12 questions to ask before you buy an AI tool for your school Source: https://navon.org/blog/questions-to-ask-an-ai-vendor Published: 2026-09-03 · Category: Policy Demos are built to look impressive. These twelve questions are built to find out what happens when nobody is watching — with the bad answers listed next to the good ones. A demo shows you the product working. What you need to know is how it behaves when it is not working, when a student is trying to get around it, and when something goes wrong at 11pm on a Sunday. These questions get at that. Ask all twelve, in order, and write the answers down. ## Filtering ### 1. Where does the filtering run? Good: on the server, before the request reaches a model. Worrying: an extension, an app setting, a managed profile — those protect a device, and a student with a second device is outside them. ### 2. What happens when your safety check is unavailable? Good: the request is refused until it recovers. Worrying: any answer that avoids the word "refused", or an explanation of how rarely it happens. Frequency is not the question. ### 3. Can we add our own blocked topics, and does that apply to images and video? Good: yes, and a demonstration in the image tool, not only in chat. Worrying: "our filter already covers everything inappropriate" — that is their definition, not yours. ### 4. Do you check the output as well as the request? Good: both. Worrying: only the input, which assumes a reasonable-looking question never produces an unsuitable answer. ## Data ### 5. Which companies receive our students' prompts? You want names, including any gateway or reseller in the middle, not a category like "leading AI providers". Ask for the current sub-processor list in writing. ### 6. Is our students' content used to train models — yours or anyone else's? Good: a clear no for their own use, and a straight statement of what the underlying providers' terms allow. Worrying: "we do not train on your data" with no mention of the providers actually running the model. That sentence can be true and still not answer the question. ### 7. How long is student content kept, and can we delete it? Good: a stated retention period and a working deletion path. Worrying: "indefinitely, for quality purposes". ## Oversight ### 8. What exactly can staff see? Good: a specific answer — blocked attempts, categories, patterns over time. Worrying at both extremes: nothing at all leaves you unable to act, and every message in full is a surveillance system your families have not agreed to. ### 9. Can a parent see their own child's activity? Worth asking even if you do not intend to enable it, because the answer reveals whether the data model has any notion of a guardian relationship. ### 10. Can an administrator switch off a category like child-safety filtering? Good: certain categories are locked and cannot be disabled by anyone, including them. Worrying: "an admin can configure all categories" — that is a single compromised or careless admin account away from a serious problem. ## Reality ### 11. How many schools use this today, and may we speak to one? A new product is not disqualifying and a straight answer is essential. Worrying: a number with no reference, or named institutions who turn out not to be customers. ### 12. What does it do when your AI provider is down or out of credit? Good: a clear error telling the user to try later. Worrying: not knowing, or a vague answer — this is the most common real-world failure of AI products and it will happen during a lesson. ## Before you sign 1. Ask for a trial account and try to break the filter yourself for twenty minutes. Try rephrasing, try another language, try the image tool. 2. Test on a personal phone that the school does not manage. 3. Set a blocked topic and a time limit, then verify both actually took effect. 4. Get the sub-processor list and the retention period in writing, not in a demo. Twenty minutes of adversarial testing tells you more than any deck. If a vendor is reluctant to hand over a trial account for exactly that, you have learned something. [Ask us the twelve](https://navon.org/contact-us) ## Common questions **What should schools ask AI vendors about data privacy?** Which companies receive student prompts including any gateway or reseller, whether that content is used for training by the vendor or the underlying providers, how long it is retained, and whether it can be deleted. Ask for the sub-processor list in writing. **How do we test an AI tool's filter before buying?** Get a trial account and attempt to defeat it: rephrase blocked requests, try another language, try the image and video tools rather than only chat, and test on a personal device the school does not manage. **What is the most revealing question to ask?** What happens when the safety check itself is unavailable. A vendor that fails open has a filter that switches off during an outage, and the answer tends to be evasive rather than a simple 'the request is refused'. --- # Why AI detectors do not work, and what to do instead Source: https://navon.org/blog/why-ai-detectors-do-not-work Published: 2026-09-03 · Category: For schools Every school that has leaned on an AI detector has eventually accused a student who did nothing wrong. The failure is not a bug in a particular product — it is the whole idea. A detector offers exactly what an overwhelmed department wants: paste the essay in, get a percentage, act on it. The trouble is that the percentage does not mean what it appears to mean, and treating it as evidence produces the single worst outcome available to a school — a diligent student accused of cheating. ## The failure is structural, not a product defect Detectors work by measuring how predictable a text is. Machine-generated writing tends to choose likely words in likely orders, so unusually smooth, unsurprising prose scores as machine-written. The problem is that plenty of human writing is also smooth and unsurprising, and it is a specific kind of human writing: careful, conventional, formal. Which is to say, the writing produced by a student who has been taught to write clearly, who is working in a second language and sticking to safe constructions, or who is following a structure the school gave them. > A detector systematically suspects the students who have most carefully done what they were told. This bias is not something a better model removes, because it is the signal itself. Meanwhile a student who asks a model to write with more variation, or who rewrites a few sentences by hand, defeats the detector without effort. So it points the wrong way in both directions at once. ## What a percentage actually is "87% AI" reads like a probability that the student cheated. It is not. It is a score describing the text's statistical texture, on a scale the vendor chose, with a threshold the vendor picked. It carries no information about who wrote it. Test this before trusting any tool: run several pieces of work you watched a student produce. Schools that do this usually stop using the detector, because the outputs do not survive contact with cases where the answer is known. ## The cost of one wrong accusation Consider what it takes to defend yourself. The student has no evidence of their own innocence beyond insisting, and the school is holding a number. It falls on a child to disprove a machine. That damages the student's trust in the school permanently, and the story travels — to their friends, to their parents, to the wider community. One such case costs more than any number of undetected AI essays. ## What actually works ### Assess the process, not only the artefact Ask for the outline, the draft with its revisions, and the sources. A student who did the work has these and a student who did not cannot fabricate them convincingly. This also happens to be better pedagogy than grading a finished product alone. ### Make some work unfakeable by design Handwritten in-class writing, a short viva on a submitted essay, or a task tied to something specific to your classroom — a discussion you had, a text you annotated together. Two minutes of "talk me through your second paragraph" separates the student who wrote it from the student who did not, with no technology at all. ### Require disclosure and make it safe One line: which tool, which part, what they changed. If honest disclosure is punished, students conceal, and you are back to guessing. Treat the line as compliance. ### Say what each task is for Most AI-in-homework anxiety comes from tasks whose purpose was never stated. If a task exists to practise an unaided skill, say so and assess it in conditions that make that real. If it does not, AI use may not matter much. ## If you already have a detector You do not need to throw it away, but change its status. It may raise a question; it may never be the answer. Nothing should follow from a score alone except a conversation, and the conversation — not the number — is what you act on. [Navōn for schools](https://navon.org/services/schools) ## Common questions **Are AI detectors accurate?** No. They measure how predictable a text is, so they flag careful, conventional writing — including that of second-language students — while a lightly edited AI draft passes. The bias is in the signal itself, so a better model does not fix it. **Can a student be disciplined based on an AI detector score?** A score describes a text's statistical texture, not who wrote it, so it is not evidence. It may prompt a conversation, but nothing should follow from the number alone. **What should schools use instead of AI detection?** Assess the process — outlines, drafts, sources — include some work that is unfakeable by design such as in-class writing or a short viva, require honest disclosure without punishing it, and state what each task is meant to measure. --- # AI and homework: a guide for parents Source: https://navon.org/blog/ai-and-homework-a-guide-for-parents Published: 2026-09-03 · Category: For families The useful question is not whether your child uses AI for homework. It is whether the thinking still happened. Here is how to tell. Most parents discover this the same way: a piece of homework comes back noticeably better than usual, or a child answers a question about their own work a beat too slowly. The instinct is to ban it, which mostly relocates the behaviour. A more useful frame is that AI can occupy one of two positions in a homework task. It can do the thinking, or it can support the thinking. The same tool, the same subject, and the same evening can go either way. ## Uses that support the learning - Explaining a concept again, differently. A patient re-explanation at 9pm is genuinely valuable, and repetition without impatience is something a tired parent often cannot offer. - Generating practice questions. Working through problems a machine produced is real practice. - Critiquing the child's own draft. The child writes, then asks what is unclear. The writing is theirs; the feedback loop got faster. - Translation and vocabulary support, which removes a barrier rather than removing the task. - Checking their own reasoning after they have reached an answer — not before. ## Uses that replace it - Producing the answer, which is then copied. - Writing the essay, lightly reworded. - Summarising a text the child was meant to read. - Solving the problem set, with the working reproduced as if it were theirs. The distinction is not about how much AI was involved. It is about whether the effortful part still happened. Learning largely lives in the struggle — the retrieval, the false start, the second attempt. A tool that removes the struggle removes the learning along with it, however good the output looks. ## The kitchen-table test You do not need software for this. Ask your child to explain one part of what they submitted, in their own words, without looking. A child who did the thinking can do this easily, imperfectly, with the hesitations of someone reconstructing rather than reciting. A child who did not will produce either a fluent summary that suspiciously matches the text, or nothing. Do it as interest rather than interrogation. "Tell me about this bit" gets you a real answer; "did you write this yourself" gets you a defensive one. ## What to say when you find it The instinct is a rule about AI. A better conversation is about what homework is for — and most children have genuinely never been told. They think it is a thing to hand in. Once it is framed as practice for something they will be tested on alone, using a machine to skip it starts to look like what it is: skipping their own practice. That reframing does more than a prohibition, because it survives you leaving the room. ## The part almost nobody mentions Models state wrong things with total confidence, including in maths and history — the subjects where homework help is most often sought. A child who cannot yet evaluate the answer has no way to notice. So the single most useful habit to teach is checking anything that matters against another source. It protects them from the tool's real failure mode, and it is the skill that stays useful as the tools change. [Navōn for families](https://navon.org/services/family) ## Common questions **Should I let my child use AI for homework?** It depends on the use, not the tool. Explaining a concept, generating practice questions and critiquing the child's own draft support learning; producing the answer or writing the essay replaces the thinking the homework existed to practise. **How do I know if my child used AI for their homework?** Ask them to explain one part in their own words without looking. A child who did the thinking reconstructs it with normal hesitation; one who did not either recites something matching the text or cannot answer. **Is using AI for homework cheating?** Using it to skip the effortful part of a task meant to build a skill is skipping their own practice, whatever the school's rule says. Using it to understand something faster generally is not. --- # When something gets through the filter Source: https://navon.org/blog/when-something-gets-through-the-filter Published: 2026-09-03 · Category: Safety Every filter has a miss rate above zero. What separates schools is not whether something gets through — it is what happens in the following hour. Any vendor implying their filter is perfect is either misinformed or hoping you are. Content moderation involves judgement calls at a boundary, made at scale, by systems that are right most of the time. Most of the time is not all of the time. So the responsible thing is to decide in advance what happens on the day something gets through, when the people involved are upset and the temptation to react quickly is strongest. ## In the first hour 1. Attend to the student first. Everything else can wait ten minutes; a child who has seen something upsetting cannot. 2. Preserve the evidence before anything changes it. Screenshot what was seen, and note the account, the approximate time and the tool used. Do not delete the conversation — it is the only record of what actually happened. 3. Do not interrogate. What was asked matters, but a child who feels accused stops being accurate, and you need accuracy more than you need a confession. 4. Tell one designated person. Whoever holds safeguarding should know before the story spreads through the staff room. ## Then work out which kind of failure it was This determines everything afterwards, and the three cases look identical at first. ### A judgement call at the boundary The system evaluated it and allowed it, and reasonable people might disagree. This is the most common case and the least alarming. The remedy is to tighten the specific rule. ### A gap in coverage The rule existed but did not apply to this tool — most often, a topic blocked in chat that was not blocked in image generation. This is a product defect and worth raising sharply with the vendor, because it means the school's policy was never actually in force where it mattered. ### The check did not run The safety system was unavailable and the request went through unchecked. If a product can do this, it will do it again, and no amount of rule-tightening helps. Ask the vendor directly whether their system fails open, and treat a vague answer as a yes. ## Telling parents Tell them, and tell them before they hear it elsewhere. A school that reports its own incident is trusted; a school that is discovered not to have is not, and the second outcome is far more expensive than the first. Say what happened, what the child saw, what you have changed, and what you are asking of them. Resist the urge to minimise — parents are far more forgiving of an imperfect filter honestly reported than of a confident reassurance that later turns out to have been wrong. ## What to change afterwards Add the specific topic to your blocked list, in the words that actually came up rather than a euphemism. Check the same request in every other tool, not only the one where it happened. Ask the vendor what changed on their side, and whether other schools saw it. Then leave the rest alone. The instinct after an incident is to restrict everything, and a filter tightened into uselessness gets worked around, which leaves you worse off than before. ## The honest expectation to set When a school adopts a filtered AI tool, say to staff and parents at the outset that the filter reduces exposure and does not eliminate it, and that there is a plan for the exception. A school that has said this in advance is in a completely different position on the day than one that promised the problem could not occur. [Ask the hard question](https://navon.org/contact-us) ## Common questions **What should a school do if inappropriate content reaches a student through an AI tool?** Attend to the student first, preserve the evidence before it is deleted, avoid interrogating, and inform the designated safeguarding person. Then determine whether it was a boundary judgement, a gap in tool coverage, or the safety check failing to run. **Can any AI content filter be 100% effective?** No. Moderation involves judgement at a boundary at scale, so the miss rate is above zero. A vendor claiming perfection is not describing a real system. **Should we tell parents when something gets through?** Yes, and before they hear it elsewhere. Parents forgive an imperfect filter honestly reported far more readily than a reassurance that later proves wrong. --- # Screen time limits for AI that actually hold Source: https://navon.org/blog/screen-time-limits-for-ai-that-hold Published: 2026-09-03 · Category: For families Set a limit, then test it after it should have expired. A surprising number of advertised time controls are decorative. Time controls are among the most commonly advertised and least commonly enforced features in family software. They are easy to put in an interface and fiddly to implement properly, and nobody checks — so a settings screen that saves a value and does nothing with it can survive a long time. The test takes two minutes. Set a limit that has already been exceeded, or a window that has closed, then try to use the product. If it works, you have a setting, not a limit. ## Three things a real limit needs ### It has to be enforced on the server If the countdown lives in the app on one device, a second device has its own countdown, and reinstalling resets it. Enforcement has to sit with the account, checked when a request is made. ### It has to know which clock you meant "Available 4pm to 7pm" is meaningless without a timezone. If a product stores the hours but not the zone, it will evaluate them against a server clock — and a family several hours away from that server gets a window at the wrong time of day, every day. ### The child must not be able to set it Obvious, and not always true. If the timezone or the limit can be changed from the child's own session, or supplied by their device, the control is advisory. ## What limits are worth setting Not many, and chosen deliberately. Two are usually enough: - A window that ends before bedtime. The specific harm of late-night use is sleep, and this addresses it directly. - A daily budget, if AI use is displacing something rather than supporting it. Set it generously at first — a limit that makes legitimate homework impossible gets circumvented, and then nothing is enforced. Resist stacking more. Every additional restriction increases the chance the child hits a wall while doing exactly what you wanted them to do, and each of those moments makes the whole arrangement feel arbitrary. ## Tell them the limit exists A silent restriction that produces a confusing failure teaches a child the tool is broken, not that a boundary exists. If the product says plainly that a parent set these hours, the limit does the work you wanted, and the child knows where to bring an objection. Also decide in advance what you will do when they ask for more. They will, and "no, because I said" spends credibility you will want for larger questions. ## What a time limit cannot do It governs one product. A child with a phone can open a different AI tool at midnight, and no limit inside one platform reaches that. Time controls manage the tool you have chosen; they are not a substitute for the conversation about the ones you have not. [See the family controls](https://navon.org/services/family) ## Common questions **Do parental time limits on AI apps actually work?** Test rather than assume. Set a window that has closed and try to use the product. A real limit is enforced on the server against the account; one enforced in an app is defeated by a second device or a reinstall. **Why do time windows need a timezone?** Hours like '4pm to 7pm' are meaningless without knowing whose clock. A product that stores hours but not the zone evaluates them against its server clock, so a family in another timezone gets the window at the wrong time of day. **What screen time limits should I set for AI?** Usually two: a window that ends before bedtime, and a generous daily budget if AI is displacing other activity. A limit tight enough to block legitimate homework gets circumvented. --- # AI image generation and children: the gap most filters miss Source: https://navon.org/blog/ai-image-generation-and-children Published: 2026-09-03 · Category: Safety A student blocked from discussing a subject who can generate a picture of it has not been protected. They have been redirected — and this is the most common gap in filtered AI products. If you test one thing in a filtered AI product before trusting it, test the image tool. Text filtering was built first and is usually the most mature part of any platform. Image, video and voice generation tend to arrive later, and the filtering does not always arrive with them. The result is a specific and predictable gap: a school configures its blocked topics, tests them in chat, sees them work, and never checks whether the same rules reach the picture generator. Often they do not. ## Why the gap happens It is rarely negligence. It is how these products get built. Filtering usually starts as a check in the chat path. When image generation is added, it is a different endpoint with a different request shape, and the check has to be added there deliberately. If the guardian rules live in one function and the image tool calls another, nothing fails loudly — the picture is simply generated. Nobody sees an error, so nobody investigates. The same applies to video, voice narration, and any story or book generator. Each is a separate path, and each has to be wired to the same rules. ## What makes images harder than text - The prompt and the result can diverge. An innocuous description can produce an unsuitable image, so screening only the request is not enough — the output has to be checked too. - Modesty standards are visual and specific. A rule about clothing or depiction has to be enforced in the image itself, not inferred from the words asked for. - Indirection is easy. A subject can be requested by describing a scene rather than naming it, which a text rule written in nouns may not catch. - Images circulate. A generated picture leaves the platform as a file and travels in a way a chat message usually does not. ## How to test it in ten minutes 1. Set a blocked topic your school genuinely cares about. 2. Confirm it is refused in chat. It almost certainly will be. 3. Ask for an image of the same subject, named directly. 4. Ask again, describing the scene without naming the subject. 5. Repeat in every generation tool the product offers — video, story, book, presentation. 6. Do it all on a device the school does not manage. If step 3 or 4 succeeds anywhere, the school's policy is not in force where it matters, whatever the settings screen shows. ## What good looks like The guardian's rules are evaluated in one place that every tool consults, rather than reimplemented per feature. The request is screened before generation and the result is screened after. A blocked attempt is recorded the same way regardless of which tool it came from, so a pattern in the image tool is as visible as one in chat. And there is a switch for the category as a whole — a guardian who wants image and story generation off entirely should be able to turn it off, rather than trying to enumerate every subject in advance. [See what runs on every request](https://navon.org/features) ## Common questions **Do AI content filters cover image generation?** Not always. Text filtering is usually built first and image, video and voice tools are separate paths that must be wired to the same rules deliberately. This is the most common gap in filtered AI products. **How do I test whether a filter covers images?** Set a blocked topic, confirm it is refused in chat, then request an image of the same subject — first naming it, then describing the scene without naming it. Repeat in every generation tool the product offers. **Why is filtering images harder than filtering text?** The prompt and the result can diverge, so the output must be screened as well as the request. Visual standards such as modesty have to be judged in the image itself rather than inferred from the words asked for. --- # Kosher AI: what it means and how to evaluate it Source: https://navon.org/blog/kosher-ai Published: 2026-09-03 · Category: Explainers The phrase is doing a lot of work and means different things to different people. Here is what it can reasonably mean, what it cannot, and how to judge a tool against it. "Kosher AI" has become shorthand in frum communities for a tool that can be brought into a home or a classroom without worrying about what it will produce. It is a useful shorthand and an imprecise one, and the imprecision matters, because vendors have noticed the phrase sells. This is a guide to what the phrase can defensibly mean, what it cannot, and how to evaluate a specific product against it. ## Start with what it cannot mean There is no certifying body for software. Kashrus supervision applies to a physical production process with an inspectable supply chain, and there is no equivalent for a model whose outputs are generated fresh each time. Anyone presenting a hechsher-like badge for an AI product is using the visual language of certification without the substance behind it, and that should lower your confidence rather than raise it. It also cannot mean a guarantee. Content moderation involves judgement at a boundary, at scale, and every system has a miss rate above zero. A vendor claiming nothing objectionable can ever appear is not describing a real system. ## What it can defensibly mean Something narrower and genuinely valuable: that the tool is built so a responsible adult sets the standard, that the standard is enforced everywhere the tool is used, and that you can find out what it did. Three properties, each checkable. ## Property one: the standard is yours A general-purpose assistant applies one safety policy, written by its vendor for a global audience of adults. It refuses what a large company considers harmful. It has no view on what your kehilla does not discuss, because it was never asked. So the first question is whether you can state your own topics and have them enforced — not choose from the vendor's categories, but write your own. A product that offers a strictness slider is offering you their judgement at three intensities. Test it concretely. Add a topic your community genuinely cares about, in the words that would actually come up, then try to reach it. If it is refused, ask for it obliquely — described rather than named. A rule that only catches the noun is not much of a rule. ## Property two: it applies everywhere Two dimensions here, and products fail at both. ### Every tool, not just chat Text filtering is built first and matures first. Image, video, voice and story generators are separate paths, and the rules have to be deliberately extended to each. A talmid blocked from discussing a subject who can generate a picture of it has been redirected, not protected. Test every tool the product offers, not the one on the front page. ### Every device, not just the school's If the filtering is an extension, an app setting or a managed profile, it protects that machine. A bochur with a phone is outside it. Filtering that runs on the server, before a request reaches a model, is a property of the account and follows the user everywhere they sign in. ## Property three: you can find out what happened A filter that blocks silently and keeps no record leaves a parent or menahel with nothing. One that stores every prompt in full creates a different problem — a surveillance archive of children's private questions that someone now has to protect. The workable middle is a record of decisions: what was blocked, roughly what category, and how often. That is enough to notice a pattern worth a conversation and not so much that it becomes a diary. ## The question almost nobody asks Every filter depends on something — a model, an API, a network call — and that dependency will sometimes be unavailable. There are exactly two possible behaviours, and the choice is made deliberately by the people who built it. Fail open: the request proceeds unchecked, so the product keeps working and the filtering silently stops. Fail closed: the request is refused until checking is available again, so the filtering holds and the product is briefly unusable. > A filter that fails open is not a filter. It is a filter-shaped feature that switches itself off exactly when nobody is watching. Failing closed costs something real — during an outage, nobody can work. For a tool used by children that is the correct trade, and a vendor should be willing to state their answer plainly. An answer about how rarely outages happen is not an answer to the question asked. ## Locked categories There is a category of content that must not be switchable by anyone: child exploitation, and depending on your standards, several others. Ask whether an administrator — theirs or yours — can turn those off in a settings screen. The right answer is that certain categories are locked in the system itself and cannot be disabled by any account. "An admin can configure all categories" means one careless or compromised admin account stands between a school and a serious problem. ## What no tool can do for you Worth stating plainly, because the phrase "kosher AI" invites the opposite belief. No filter reaches a different tool in another tab. No filter makes a machine a reliable source for halacha — a model produces fluent, confident answers and will invent sources that do not exist, and a talmid who treats that as psak has been badly served. And no configuration substitutes for an adult who knows what a child is doing and is someone they will tell. A filtered tool narrows exposure and makes one option trustworthy enough to permit. That is genuinely worth something. It is not a fence around the whole problem, and a product sold as one is overselling. ## The evaluation in short 1. Can you write your own blocked topics, in your own words? 2. Are they enforced in the image and video tools, not only in chat? 3. Does the filtering run on the server, so it follows a student off school devices? 4. Does it fail closed when the safety check is unavailable? 5. Are the categories that must never be disabled actually locked? 6. Can a responsible adult see what was blocked, without reading everything? 7. Is the vendor willing to answer all six plainly, in writing? Six questions about the product and one about the company. The seventh is often the most informative. [See how the filtering works](https://navon.org/features) ## Common questions **What is kosher AI?** There is no certification for software. Used carefully, the phrase describes an AI tool where the school or family sets the content standard rather than accepting a vendor's global default, that standard is enforced on every tool and every device, and a responsible adult can see what was blocked. **Is there a hechsher for AI?** No. Kashrus supervision applies to a physical process with an inspectable supply chain, and there is no equivalent for generated output. A badge presented as certification is borrowing the visual language without the substance. **Can AI be used for halacha questions?** It should not be treated as authoritative. A language model produces fluent, confident answers and can invent sources that do not exist. It can help unpack a difficult passage or translate; it cannot tell you what the halacha is. **What is the single most important thing to check?** What the system does when its own safety check cannot run. A product that fails open has filtering that silently switches off during an outage, which is exactly when nobody is watching. --- # Tznius and AI image generation Source: https://navon.org/blog/tznius-and-ai-image-generation Published: 2026-09-03 · Category: Safety A text rule is written in words and a modesty standard is about what appears in a picture. Those are different problems, and most products only solved the first. Of all the places a filtered AI product tends to fall short, image generation is the most common and the least tested. It is worth understanding why, because the reason predicts exactly where to look. ## Why a visual standard is a different problem A topic rule is a statement about subject matter: this subject is not discussed. It can be checked against the words of a request, and against the words of an answer. A modesty standard is not about subject matter at all. A request to draw a person is entirely acceptable; whether the result is acceptable depends on how the person is depicted — clothing, pose, framing, setting. None of that is in the request. It is decided by the model as it generates, and it can only be judged by examining the image itself. So a product that screens only the prompt cannot enforce a tznius standard at all, however good its text filtering is. The prompt was fine. The picture was the problem. ## The three failures, in order of frequency ### The rules never reached the image tool The most common by far, and it is an architectural accident rather than negligence. Filtering starts in the chat path. Image generation arrives later as a separate endpoint, and unless the guardian rules are deliberately wired into it, nothing fails loudly — the picture is simply produced. No error appears, so nobody investigates. This is why a school can configure topics, verify them in chat, and still have a policy that is not in force where it matters most. ### Only the request is screened The rules reached the image tool, but they are applied to the prompt alone. This catches a request that names something objectionable and misses everything decided during generation — which, for modesty, is nearly everything. ### Indirection A subject can be requested by describing a scene rather than naming it. A rule written in nouns misses a request written in description. This is where a model-based check that reads the whole request beats a keyword list, which will only ever catch the words it was given. ## What to test, and it takes ten minutes 1. Set a blocked topic your school or family genuinely cares about. 2. Confirm it is refused in chat. It almost certainly will be. 3. Ask the image tool for the same subject, named plainly. 4. Ask again, describing the scene without naming the subject. 5. Ask for a picture of a person in a way that tests your modesty standard specifically. 6. Repeat in every generating tool — video, story, comic, presentation. 7. Do the whole thing on a device the school does not manage. Any success at steps 3 to 5 tells you the settings screen and the enforced reality have come apart. ## What good looks like The guardian's rules live in one place that every tool consults, rather than being reimplemented per feature — that is what stops the image tool drifting from the chat tool as the product grows. The request is screened before generation and the image is screened after. A blocked attempt is recorded identically wherever it came from, so a pattern in the image tool is as visible as one in chat. And there should be a switch for the whole category. A parent or school that wants image and story generation off entirely should be able to turn it off, rather than trying to enumerate in advance every way a picture could go wrong. Enumerating is a losing game; switching off is not. [See what runs on every request](https://navon.org/features) ## Common questions **Why do AI image filters miss modesty issues?** A modesty standard is about how something is depicted — clothing, pose, framing — none of which appears in the request. A product that screens only the prompt cannot enforce it, because the prompt was acceptable and the generated image was the problem. **How do I test whether an AI tool's filter covers images?** Set a blocked topic, confirm it is refused in chat, then ask the image tool for the same subject by name and again by describing the scene without naming it. Repeat in every generating tool and on an unmanaged device. **Can image generation be turned off entirely?** In a well-built product, yes, as a single category switch. That is more reliable than trying to enumerate in advance every way a generated picture could be unacceptable. --- # AI for schools: a complete guide Source: https://navon.org/blog/ai-for-schools-complete-guide Published: 2026-09-03 · Category: For schools Most AI guidance for schools is either abstract principles or a product pitch. This is the sequence of decisions an actual school faces, in the order they come up. A school adopting AI faces about eight decisions. They arrive in a predictable order, most of them are settled badly by default rather than deliberately, and the expensive mistakes are all made early. This walks through them in sequence. ## Decision 1: permit, ban, or ignore Ignoring is the default and the worst of the three: students use AI, the school has no position, and staff each improvise a different rule. A ban is honest and does real work — it keeps AI out of assessments meant to be unaided and gives staff a line to hold. What it does not do is stop use, because students have phones. Its measurable effect is that use moves out of view, on tools the school has not filtered. Permitting one filtered tool gives the school something to point at, which is worth more than a prohibition it cannot enforce. It is also the only option that lets the school teach anything about how the tools fail. ## Decision 2: which tool The properties that matter are narrow. Can the school write its own blocked topics? Do those apply to image and video, not only chat? Does filtering run on the server, so it follows students home? What happens when the safety check is unavailable? Can staff see what was blocked? Everything else — model quality, interface, integrations — is real but secondary, because a tool that fails these cannot be adopted at all. ## Decision 3: the content standard This is the school's, not the vendor's, and it takes a meeting rather than a settings screen. Name the topics in the words that would actually come up, not euphemisms — a filter configured from "inappropriate material" enforces nothing. Expect to revise it. The first version is always missing something, and the useful signal is what students actually attempted in the first term. ## Decision 4: the policy Three categories, not two: permitted, expected, forbidden. Most policies fail because they only have permitted and forbidden, and cannot express a task where the teacher wants AI used. Push the decision to the assignment rather than issuing a blanket rule, and give staff a standard phrase for each category to put on the task itself. Also settle attribution — one sentence naming the tool, the part, and what the student changed — and make honest disclosure safe rather than punishable. ## Decision 5: assessment This is where most of the actual risk lives, and no filter addresses it. Decide which tasks measure unaided ability and make those unfakeable by design: in-class writing, a short viva on a submitted piece, work tied to something specific to your classroom. Do not build the response around AI detectors. They measure how predictable writing is, which means they suspect careful and second-language students while a lightly edited AI draft passes. A school that disciplines on a detector score will eventually accuse a student who did nothing wrong, and that costs more than every undetected essay combined. ## Decision 6: staff Staff need three things, and a demo of the product is not among them. - Twenty minutes actually using the tool, including trying to break the filter. Nobody trusts a control they have not tested. - The three policy categories and the exact phrase to put on an assignment. - One person to escalate to, and permission to say "I do not know, I will find out" to a student. The failure mode is a policy circulated by email that nobody reads, after which each teacher improvises and students correctly conclude the rules are arbitrary. ## Decision 7: parents Tell them before rollout, not after. Say which tool, what it filters, what the school can see, and what you are asking of them at home. A parent who knows the school provides a filtered option is markedly less likely to hand over an unfiltered one. Say plainly that the filter reduces exposure rather than eliminating it, and that there is a plan for the exception. A school that has said this in advance is in a completely different position on the day something gets through than one that promised it could not happen. ## Decision 8: what you will review Set this before launch or it never happens. Once a term is enough: what was blocked and in what categories, whether any pattern suggests a topic to add, which staff are using it and which are not, and whether the assessment changes held. Patterns are more useful than incidents. A single blocked request is noise; the same category appearing repeatedly is information. ## A realistic first term 1. Weeks 1-2: choose the tool, write the content standard, draft the policy. 2. Weeks 3-4: staff session, including hands-on time trying to defeat the filter. 3. Week 5: letter to parents, before any student account exists. 4. Weeks 6-8: one department or year group only. Small enough to fix things quietly. 5. Weeks 9-12: widen, using what the pilot taught you about the topic list. 6. End of term: first review. Expect the content standard to need revision. The common failure is launching school-wide in week one, discovering the topic list has a gap in week three, and having no way to fix it quietly. [Navōn for schools](https://navon.org/services/schools) ## Common questions **How should a school introduce AI?** Pilot with one department or year group for several weeks before widening. Choose the tool, write your own content standard, run a hands-on staff session, and write to parents before any student account exists. **What should schools look for in an AI tool?** Whether the school can write its own blocked topics, whether those apply to image and video as well as chat, whether filtering runs on the server so it follows students home, what happens when the safety check is unavailable, and what staff can see. **Do schools need an AI policy?** Yes, with three categories rather than two — permitted, expected and forbidden — decided per assignment rather than as a blanket rule, plus an attribution requirement that is safe to comply with. **Should schools use AI detection software?** No. Detectors flag careful and second-language writing while lightly edited AI passes, so a school disciplining on a score will eventually accuse an innocent student. Assessment design is the dependable answer. --- # Rolling out AI in your school, week by week Source: https://navon.org/blog/rolling-out-ai-in-your-school Published: 2026-09-03 · Category: For schools Most failed rollouts fail the same way: everything at once, in week one, with no way to fix a problem quietly. Here is the sequence that works instead. A rollout is not a purchase followed by an announcement. The schools where this goes badly almost all did the same thing — enabled everyone at once, then discovered a gap in the topic list with six hundred students already using it. This plan assumes a term. It can be compressed, but not the pilot. ## Weeks 1-2: decide and configure Choose the tool against the properties that matter: your own blocked topics, applying to image and video, filtering on the server, failing closed, visible to staff. Then write the content standard. This is a meeting, not a settings screen — get the people whose judgement the community trusts in a room and name the topics in the words that would actually come up. Configure it, then spend twenty minutes trying to defeat what you just configured, including in the image tool and on a personal phone. Draft the policy in parallel: three categories, an attribution sentence, and a note on what each assignment type expects. ## Weeks 3-4: staff One session, hands-on. Give every member of staff an account and twenty minutes to use it and try to break the filter. This is the step most often cut and the one that determines whether staff trust the tool. Then the practical part: the three categories, the exact phrase to put on an assignment, who to escalate to, and what to say to a student who asks something you cannot answer. Expect resistance, and expect it to be reasonable. The teacher who says this will make cheating easier is not being obstructive; they are naming the real risk. Answer it with the assessment changes, not with reassurance. ## Week 5: parents Write before any student account exists. Which tool, what it filters, what the school can see, what you are asking at home, and who to contact. Say plainly that the filter reduces exposure rather than eliminating it. Doing this before rollout rather than after is the difference between informing parents and being caught. ## Weeks 6-8: pilot One department or one year group. Small enough that a problem can be fixed without an announcement, large enough to produce real usage. Watch three things: what got blocked and whether the categories look right, what staff are actually using it for versus what you expected, and whether anything reached a student that should not have. Add topics as you learn them — the first list is always incomplete, and finding that out with thirty students is much better than with six hundred. ## Weeks 9-12: widen Extend to the rest of the school with the revised topic list. Have the pilot staff run the session for the next group; a teacher who has used it for six weeks is more persuasive than any administrator. Keep the escalation route obvious. Most problems in this phase are people not knowing who to tell. ## End of term: review What was blocked, in what categories, and does that suggest topics to add. Which staff adopted it and which did not, and why — non-adoption is usually a signal about training or about assessment anxiety, not about the tool. Whether the assessment changes held. Whether anything got through, and which of the three kinds of failure it was. ## The four ways rollouts fail - Everyone at once. No way to fix a gap quietly, and the first problem becomes a school-wide story. - Policy by email. Nobody reads it, every teacher improvises, students conclude the rules are arbitrary. - No assessment change. The tool arrives, cheating anxiety rises, staff blame the tool for a gap in task design. - Parents told afterwards. Whatever the school did becomes secondary to the fact that it did not say. [Talk to us](https://navon.org/contact-us) ## Common questions **How long does it take to roll out AI in a school?** About a term. Two weeks to choose and configure, two for staff training, one for parent communication, three for a pilot with one department or year group, then widen — with a review at the end of term. **Should we pilot AI with one class first?** Yes. A pilot is small enough to fix a gap in the topic list quietly and large enough to produce real usage. The first content standard is always incomplete, and discovering that with thirty students is far better than with six hundred. **What is the most common rollout mistake?** Enabling everyone at once, so the first problem becomes a school-wide story with no way to fix it quietly. The second most common is circulating the policy by email and never training staff. --- # Talking to parents about AI in school (with a letter you can adapt) Source: https://navon.org/blog/talking-to-parents-about-ai Published: 2026-09-03 · Category: For schools Parents are not opposed to AI in school. They are opposed to finding out about it afterwards. Here is what to say, and a letter you can adapt. Schools consistently underestimate how well this goes when it is done first, and how badly it goes when it is done second. A parent told in advance is a parent with a question. A parent who finds out from their child has a grievance, and the grievance is rarely about AI — it is about not being told. ## What parents actually want to know Not the technology. Four things: 1. What will my child see? Meaning: what does it refuse, and who decided. 2. Can you see what they did? Both directions matter — some parents want oversight, some are uneasy about surveillance, and both want to know which it is. 3. Will this replace their thinking? The real fear behind most objections, and it deserves a direct answer rather than reassurance. 4. What do I have to do? Most parents want a small, clear ask, not a project. ## Say the honest thing about limits The temptation is to reassure. Resist it. Say that the filter reduces exposure and does not eliminate it, that there is a plan when something gets through, and that you will tell them if it does. This feels like it invites worry. In practice it does the opposite: it is what a competent institution sounds like, and it means that on the day something happens, you are reporting a known possibility rather than explaining why your assurance was wrong. ## The questions that will come back ### "Why not just ban it?" Because a ban stops school use and not use. Their child has a phone. The choice is between a tool the school has filtered and shaped, and whatever they find on their own. ### "Will it make cheating easier?" Yes, for tasks that were already vulnerable — which is why the assessment changes matter more than the filter. Name the specific changes. This answer only works if you have actually made them. ### "My child is not allowed AI at home." Respect it, say how it will work in class, and check whether the objection is to AI generally or to unfiltered AI. It is very often the second, and this parent frequently becomes the most supportive once they understand the filtering. ### "Is our data safe?" Answer specifically: which companies receive the content, whether it is used for training, how long it is kept. If you cannot answer, get it from the vendor in writing before you send the letter. ## A letter you can adapt > Dear parents, > > From [date], [year group] will use [tool] in school. I want to tell you what it is, what it will not do, and what we are asking of you. > > What it is. [Tool] is an AI assistant with content filtering we configure. We have set the topics it will not discuss, and those rules apply to writing, images and video. The filtering runs on [tool]'s servers rather than on a device, so it applies wherever your child signs in, including at home. > > What we can see. Staff can see when a request was blocked and what kind of topic it involved. We do not read every conversation, and we do not intend to. > > What it will not do. No filter catches everything. If something reaches a student that should not have, we will address it and we will tell you. We would rather say that now than promise otherwise. > > On schoolwork. Some tasks will permit AI, some will expect it, and some will forbid it. Each assignment will say which. Where work is meant to show what your child can do unaided, we have changed how it is assessed so that remains true. > > What we ask of you. Two things. Ask your child to show you what they use it for — most are pleased to. And if anything troubles you, tell us rather than assuming we know. > > [Name] is available on [contact] for any question, including whether your child takes part. > > [Signature] Adapt it, keep it under a page, and send it before the first account exists. The paragraph on limits is the one to keep. [Ask us in writing](https://navon.org/contact-us) ## Common questions **How should a school tell parents about using AI?** In writing, before any student account exists. Cover what the tool is, what it filters and who decided, what staff can see, that the filter is not perfect and there is a plan, how assessment has changed, and one small clear ask. **What do parents worry about most with AI in school?** In practice: what their child will see, whether the school can see what they did, whether it replaces their child's thinking, and what they are being asked to do. The technology itself is rarely the concern. **What if a parent does not want their child using AI?** Respect it and offer an alternative, but check whether the objection is to AI generally or to unfiltered AI. It is often the second, and that parent frequently becomes supportive once the filtering is explained. --- # AI for teachers: what it is actually good for Source: https://navon.org/blog/ai-for-teachers-what-it-is-good-for Published: 2026-09-03 · Category: For schools Most advice for teachers is either breathless or defensive. This is a working list of what saves real time, what does not, and how to spot the difference. Almost everything written for teachers about AI is about students using it. Far less is about the teacher using it, which is where most of the immediate value sits and where the risks are much smaller. The pattern that predicts whether a task will work: AI is good where you can judge the output instantly, and poor where you cannot. You know within seconds whether a comprehension question is well-formed. You cannot tell at a glance whether a historical date is right. ## Where it genuinely saves time ### Differentiating material you already have Take a text or worksheet and ask for a version two reading levels down, or with the vocabulary glossed, or broken into smaller steps. This is the strongest use by a distance: the content is yours and already correct, the model is only reshaping it, and you can see immediately whether the reshaping worked. ### Generating practice items Twenty more questions in the style of the six you wrote. Quality is easy to check by reading them, and the failure mode is a dull question rather than a wrong one. Say the year group and the exact skill; vague requests produce generic filler. ### First drafts of routine writing Letters home, a description for a trip, the same information rewritten for a different audience. You will rewrite most of it, and starting from something is faster than starting from nothing. ### Explaining a concept another way When a student has not understood your explanation twice, a third framing helps and is genuinely hard to produce on demand. Ask for three analogies for the same idea and take the one that fits your class. ### Feedback scaffolding Not marking. Give it your rubric and a piece of work and ask what a student might be told about structure. Use it as a prompt for your own comments — you know the child, and it does not. ## Where it wastes your time - Facts you cannot verify at a glance. Dates, attributions, statistics, anything requiring a citation. Checking takes longer than looking it up yourself. - Anything requiring knowledge of your specific class. It does not know which three students will find this hard or that last week's lesson went badly. - Final grades. Beyond the reliability problem, a grade a teacher cannot explain is one they should not have given. - Long documents where an error hides. A plausible schemes-of-work document with one wrong sequencing decision costs more to audit than to write. ## The three-minute rule If checking the output would take longer than doing the task yourself, do the task yourself. This resolves nearly every case. Rewriting a worksheet for a lower reading level: checking takes a minute, doing it takes twenty. Worth it. Producing a timeline of a historical period: checking every date takes longer than writing the timeline from your own knowledge. Not worth it. ## Prompting, briefly Most disappointing output comes from a request that omits the context you were carrying in your head. - Say the year group and the ability range. "Year 8, mixed, several working below" changes the output completely. - Say what it is for. Starter, homework, assessment and revision want different things. - Give it your example. One question you wrote is worth a paragraph of description. - Ask for more than you need and select. Ten questions to choose four from beats asking for four. - Push back rather than restarting. "Too wordy for this group, halve it" gets there faster than a new prompt. ## On the honesty of it Some teachers feel uneasy using a tool students are restricted from. That is worth taking seriously rather than dismissing. The distinction that holds up: the student's task is usually to build a capability, so a tool that skips the effort defeats the point. The teacher's task is usually to produce material, and how it was produced matters less than whether it is right and suitable — which you are qualified to judge and the student is not. That said, if a school expects students to disclose AI use, staff modelling the same disclosure is worth more than any assembly about honesty. [Navōn for schools](https://navon.org/services/schools) ## Common questions **What can teachers use AI for?** Differentiating existing material, generating practice questions, first drafts of routine writing, alternative explanations of a concept, and scaffolding for feedback. It works best where you can judge the output instantly. **Should teachers use AI to mark work?** Not for grades. It can suggest what a student might be told about structure against your rubric, but a grade a teacher cannot explain is one they should not have given, and the model does not know the child. **How do I know if a task is worth using AI for?** If checking the output would take longer than doing the task yourself, do it yourself. Reshaping a worksheet passes that test; producing a timeline of dates you would have to verify does not. --- # AI and students with learning differences Source: https://navon.org/blog/ai-and-students-with-learning-differences Published: 2026-09-03 · Category: For schools This is where AI has the most to offer and the least written about it. The question is always the same: is the tool removing a barrier, or removing the task? For a student who reads slowly, writes with great effort, or is working in their second language, AI can be transformative in a way it simply is not for a confident learner. It can also, used differently, remove the exact work that would have built the skill. The distinction is the same one that governs any access arrangement: does this remove a barrier that is not what we are assessing, or does it remove the thing we are assessing? A reader for a maths paper removes a barrier. A reader for a reading comprehension paper removes the test. ## Where it clearly removes a barrier ### A student who reads slowly, in a subject that is not reading A dense history source can be simplified without touching the historical thinking the lesson is about. The student still analyses, compares and argues; they are no longer failing the lesson because of decoding speed. ### A student writing in a second language Vocabulary support and checking whether a sentence says what they meant lets a student demonstrate understanding of the subject rather than their current English. In a science lesson, English is the barrier and not the assessment. ### A student who cannot get started For many students with ADHD or anxiety around writing, the blank page is the whole obstacle. A model that asks three questions about what they want to say, and reflects the answers back as a rough structure, gets past an obstacle that has nothing to do with capability. ### A student who needs the same thing explained repeatedly Some students need six explanations, and human patience is finite in a class of thirty. A tool that gives the seventh without any change in tone is genuinely valuable, and the student knows it is not sighing. ## Where it removes the learning - Summarising a text in a lesson whose purpose is learning to summarise. - Producing the writing in a task assessing written expression, including when a student struggles with exactly that. - Answering comprehension questions about a text the student has not read. - Any use that means the student never attempts the difficult step, when attempting it is the point. The hard cases are where the barrier and the assessment are the same thing — a dyslexic student in an English writing assessment. There, AI is not an access arrangement, and the right answer is the school's existing framework for support, not a chatbot. ## The question that resolves most cases > What is this task assessing, and is the thing AI would do the thing being assessed? If the answer is no, it is support. If yes, it is a substitute. The same tool, the same student, and the same paragraph can be either, depending on the lesson's purpose — which is why this cannot be settled by a blanket rule and has to be decided task by task. ## Three practical cautions ### Dependency is real, and so is its opposite A student who always has the text simplified may not build reading stamina. A student who never gets past the first paragraph builds nothing at all. The answer is to reduce the support deliberately over time rather than to withhold it. ### Simplification loses things A simplified source is not the source. In history particularly, the language is often part of the evidence. Use simplification to get a student into a text, then bring them back to the original for the part that matters. ### Ask the student Students with learning differences usually know exactly which part is hard, and are rarely asked. "Which bit slows you down?" produces a better answer than any assumption about what a diagnosis implies. ## A note on dignity A student who has spent years being visibly the one with the different worksheet gains something real from a tool everyone in the room is using. That is not a small benefit and it does not show up in any assessment data. It also argues for the tool being available to the whole class rather than assigned to particular students, with the teacher directing its use per task. Support that does not announce itself is more likely to be used. [Navōn for schools](https://navon.org/services/schools) ## Common questions **Can AI help students with dyslexia?** Yes, where reading is a barrier rather than the thing being assessed — simplifying a dense source in a history lesson lets the student do the historical thinking. In an assessment of reading itself, it removes the test rather than a barrier. **Is using AI an unfair advantage for students with learning differences?** It follows the same logic as any access arrangement: it is support when it removes a barrier that is not being assessed, and a substitute when it does the thing being assessed. That is decided task by task, not by a blanket rule. **How can AI help EAL students?** Vocabulary support and checking whether a sentence conveys the intended meaning lets a student show subject understanding rather than their current English — appropriate in a science lesson, not in a language assessment. --- # Twelve classroom activities that use AI well Source: https://navon.org/blog/classroom-activities-using-ai Published: 2026-09-03 · Category: For schools Not "use AI to write an essay". Twelve activities where the tool does something genuinely useful, several of which teach students to distrust it. The weakness of most AI lesson ideas is that they use AI to do faster what students should be doing slowly. These are activities where the tool contributes something a teacher could not easily provide alone — and where, in several cases, the model being wrong is the entire point. ## Activities that teach scepticism ### 1. Find the error Generate an explanation in your subject containing a plausible mistake. Students work in pairs to find it against a real source. Nothing teaches the limits of fluent output like catching it out, and they remember the lesson because they did the catching. ### 2. Verify the citations Ask for five sources on a topic, then have students try to find each one. Some will not exist. The discovery that a perfectly formatted reference is invented lands harder than any warning, and it stays with them. ### 3. Same question, three ways Students ask the same factual question with different framings and compare answers. Where the answers diverge, they have found the edge of what the model reliably knows — which is a much more useful lesson than a rule about when to trust it. ## Activities that build subject skill ### 4. Beat the model Students write their answer first, then compare with the model's, then produce a better third version arguing why. This puts them in the position of critic, which is a higher-order task than either writing or copying. ### 5. Interrogate a source Give students a primary source and have them use AI only to ask clarifying questions about vocabulary and context — never for interpretation. The analysis stays theirs; the barrier to entry drops. ### 6. Generate the counter-argument Students write their position, then ask for the strongest objection to it, then respond. Building the steelman is the part students skip, and having one supplied means the response has to be real. ### 7. Explain it to a ten-year-old The student writes the simple explanation; the model critiques whether it would actually land. Simplifying accurately requires understanding thoroughly, and this makes that visible. ## Activities for practice and feedback ### 8. Endless practice items Students generate more questions of a type they are struggling with, and mark their own against a worked example. This is the most straightforward use and one of the most effective. ### 9. Rubric self-check Give students the rubric and have them ask what their own draft is missing before submitting. Most improvements are ones they could have made unaided and did not, because nobody prompted them at the right moment. ### 10. Vocabulary in context Rather than a definitions list, students ask for the same term used in five different sentences and work out the meaning from use. Closer to how vocabulary is actually acquired. ## Activities about AI itself ### 11. Write the class policy Students draft their own rules for AI use in your subject and argue for them. They arrive at stricter rules than most teachers would impose, and having argued for them they follow them. ### 12. Where should this not be used? Students identify tasks in their week where AI would be inappropriate, and say why. This produces better thinking about integrity than being told, and it surfaces uses staff had not considered. ## Two things that make these work Say what the activity is for at the start. "We are doing this to see where it fails" changes how students engage entirely. And do them on a tool the school has filtered. An activity that sends thirty students to an unfiltered assistant to see what it gets wrong is a different lesson than the one you planned. [Navōn for schools](https://navon.org/services/schools) ## Common questions **What are good classroom activities using AI?** The strongest use it for something a teacher cannot easily provide: finding a deliberate error, checking whether cited sources exist, generating a counter-argument to a student's own position, or producing unlimited practice items on a weak area. **How can I teach students that AI makes mistakes?** Have them catch one. Generate an explanation containing a plausible error and let students find it against a real source, or ask for five citations and have them try to locate each. The discovery teaches more than the warning. **Should students use AI to write essays?** Not where the essay is meant to build or measure writing. More useful is having students write first, then compare with the model's version and argue for a better third — which puts them in the position of critic. --- # AI for a small school with no IT department Source: https://navon.org/blog/ai-for-small-schools-without-it-staff Published: 2026-09-03 · Category: For schools Guidance written for schools with an IT department is not much use when the IT department is whoever last fixed the printer. This is the smaller version. Almost every article about AI in schools quietly assumes a technology lead, a procurement process, and someone to configure things. Plenty of schools — small day schools, yeshivos, single-site primaries — have none of that, and the guidance does not scale down. It scales down like this. ## What you can skip Some of the standard advice exists because large organisations have problems you do not. - Single sign-on integration. Useful at two thousand students, unnecessary at two hundred. Email and password is fine. - A formal procurement process. You need the vendor's answers in writing, not a tender. - A dedicated pilot committee. One teacher who is willing and one person who can make decisions is a pilot. - Elaborate acceptable-use documentation. One page that staff actually read beats twelve that they do not. ## What you cannot skip ### Someone must own it Not a department — a person, by name, who decides what is on the blocked list and who staff go to. The failure mode in small schools is that everyone assumes someone else is handling it, and the answer to "who set these rules" is nobody. ### You must test the filter yourself Twenty minutes, once, before students have accounts. Set a blocked topic, try to reach it in chat, then try the image tool, then try on your own phone. You do not need technical knowledge for this — you need to try to break it and notice what happens. ### Parents must be told first This costs nothing and is the difference between a school that informed its community and one that was found out. In a small school, where word travels faster and relationships are closer, it matters more rather than less. ### The assessment question is unavoidable No tool solves it, and it is not a technology problem. Which tasks measure unaided ability, and how do you know they still do? A small school can answer this in one staff meeting, which is an advantage over a large one. ## Choose for the absence of an administrator The properties that matter most when nobody is watching the dashboard: - Sensible defaults. If the tool is only safe once configured by an expert, it is not the right tool for you. - Server-side filtering. Anything requiring installation on devices means someone maintaining installations, and that person does not exist. - Fails closed. Nobody is monitoring whether the safety check is running, so it needs to protect itself. - Locked categories that no administrator can switch off — including an administrator who clicks something without understanding it. - A short review page rather than a dashboard requiring interpretation. In other words, exactly the properties that matter for a big school, but load-bearing rather than merely desirable, because there is no team to compensate for their absence. ## A realistic plan for a small school 1. One person owns it. Write the name down. 2. That person spends twenty minutes trying to break the filter, including on a personal phone. 3. One staff meeting: what it is, the three use categories, who to ask. 4. One letter to parents before any student account exists. 5. One class for three weeks. 6. Widen. Review at the end of term, for half an hour. That is the whole programme. It fits in a term without anybody's job changing. ## The advantage you have Small schools do this better than large ones more often than anyone expects. You can change the blocked list the afternoon you decide to. Every teacher can be in one room. You will hear about a problem the day it happens rather than in a report next term. The thing large schools have is process, and process is mostly a substitute for the direct knowledge you already have. [Talk to us](https://navon.org/contact-us) ## Common questions **How can a small school introduce AI without an IT department?** Name one owner, have them spend twenty minutes trying to break the filter, run one staff meeting, write to parents before any account exists, pilot with one class for three weeks, then widen. It fits a term without anyone's job changing. **What should a school without technical staff look for in an AI tool?** Safe defaults rather than safety that depends on expert configuration, server-side filtering so nothing needs installing on devices, fail-closed behaviour, categories no administrator can accidentally disable, and a short review page rather than a dashboard. **Do small schools need a formal AI policy?** One page staff actually read is worth more than twelve they do not. What cannot be skipped is naming who owns the decisions and settling which tasks measure unaided ability. --- # Talking to your child about AI Source: https://navon.org/blog/talking-to-your-child-about-ai Published: 2026-09-03 · Category: For families The conversation matters more than the settings. Here is what to say, roughly when, and what to do when they ask whether it is alive. Parents ask which controls to set. The controls matter and they are the smaller half. The larger half is whether your child would mention it to you if something strange happened — and that is built in conversation, long before anything strange does. ## The one idea underneath everything Whatever their age, one idea does the most work: it sounds certain whether or not it is right. Children learn to gauge reliability from human cues — hesitation, hedging, admitting ignorance. A model has none of those. Its wrong answers arrive in the same steady, confident tone as its right ones. A child is not being careless when they believe it; they are applying a rule that works on people to something it was never built for. Said that way, it does not sound like a restriction. It sounds like a fact about how the thing works, which is what it is. ## Roughly by age ### Younger children Keep it concrete. It is a computer program that has read an enormous amount and is very good at guessing what words come next. It is not a person, it does not know you, and it is sometimes wrong even when it sounds sure. Expect "is it alive?" — most children ask. The answer that satisfies them is that it is very good at sounding like a person and there is nobody there, in the same way a character in a book sounds like a person. Do not overcomplicate it. ### Middle years This is the age for the checking habit, and the age they can genuinely understand why. Show them once: ask something you know well, find the error together. One demonstration beats ten instructions. Also the right time for the honest conversation about schoolwork — that using it to skip the hard part is skipping their own practice. Children find this more persuasive than a rule, because it is about them rather than about permission. ### Teenagers They are likely more fluent with it than you and will not tolerate being lectured on it. Better ground: what it does with what they tell it, why a confident answer about something that matters deserves checking, and where using it would be dishonest by their own standards rather than yours. Ask what they use it for. Genuinely ask — the answer is usually more mundane and more interesting than parents expect, and asking is what keeps the channel open. ## Questions children actually ask - "Is it alive?" — No. Very good at sounding like someone; nobody is there. - "Does it know me?" — Depends on the product, and worth knowing the real answer before you give one. - "Why did it refuse?" — Because someone set rules about what it discusses. If you set them, say so; a child who knows a person decided can argue with the person, which is healthier than believing the machine simply is that way. - "Is it smarter than you?" — It has read more. It does not understand, and it cannot tell when it is wrong. Those are different things. - "Can I tell it things?" — The genuinely important one. Say plainly that it is not private, it is not a friend, and anything that matters should come to a person. ## The one to watch for Some children, particularly lonely ones, start treating a model as a confidant. It is endlessly available, never bored, never judgemental — and it is not a relationship. It cannot notice a pattern, cannot act, and cannot care. This is not a reason to prohibit; prohibition would push it out of sight. It is a reason to be curious about what they talk to it about, and to make sure there is a person they would tell. If a child is bringing serious things to a machine, that is information about what else is available to them. ## Setting rules together Children follow rules they helped write. Ask what they think is fair for when and how long, and negotiate from there. You will often get stricter proposals than you would have imposed. And tell them what you have set. A silent limit that produces a confusing failure teaches a child the tool is broken. A stated limit teaches them a boundary exists and that there is someone to ask about it. [Navōn for families](https://navon.org/services/family) ## Common questions **How do I explain AI to my child?** A computer program that has read an enormous amount and is very good at guessing what words come next. It is not a person, does not know them, and is sometimes wrong even when it sounds certain — which is the single most important idea at any age. **What should I tell my child not to share with AI?** Anything personal about themselves or others, and anything that matters emotionally. It is not private and it is not a friend; important things should go to a person who can actually notice and act. **My child treats the AI like a friend. Should I worry?** Be curious rather than prohibitive. It is available and never judgemental, but it cannot notice a pattern or care. A child bringing serious things to a machine tells you something about who else is available to them. --- # What is a language model? An explainer without the jargon Source: https://navon.org/blog/what-is-a-language-model Published: 2026-09-03 · Category: Explainers You do not need technical knowledge to make good decisions about these tools. You do need one accurate mental model, because the right one predicts every failure you will see. Most explanations of AI are either marketing or engineering. What a parent or a menahel actually needs is a working mental model — accurate enough to predict behaviour, simple enough to hold in your head. That is what this is. ## The one-sentence version A language model is a system that has read an enormous quantity of text and learned, in extraordinary statistical detail, which words tend to follow which other words. When you ask it something, it is not looking up an answer. It is producing the words that most plausibly follow your question, one at a time, each choice informed by everything it has written so far. That sounds reductive, and the surprising thing — the thing that took researchers by surprise too — is how much genuine capability falls out of doing that well. To predict the next word in a paragraph of legal reasoning, you have to have absorbed something about legal reasoning. Prediction at sufficient scale looks a great deal like understanding. ## Why this explains everything that goes wrong ### It invents things that sound right This is the single most important consequence. The model is producing plausible text, and a fabricated citation is extremely plausible text — it has the shape of thousands of real citations it has seen. It has an author, a year, a journal, and no existence. Nothing in the mechanism distinguishes recalling from inventing. Both are the same operation: produce what plausibly comes next. This is why "it makes things up" is not a bug being fixed but a property being managed. ### It is equally confident when wrong Hesitation in human speech signals uncertainty. A model produces hesitation only when hesitation is the plausible next text — which has nothing to do with whether the content is right. There is no internal confidence being expressed. This is why children in particular are poorly placed to catch errors: every cue they use to judge a speaker is absent. ### It does not know what it does not know Asked about something obscure or fictional, it will often produce a confident description rather than saying it does not know, because a confident description is more plausible text than an admission of ignorance. Newer models handle this better; none handle it reliably. ### It has no memory unless the product gives it one Each request is answered from the conversation it is shown. Anything remembered across sessions is the product storing text and feeding it back — a design decision, not something the model does. Worth knowing when you ask what a tool remembers about your child. ## What training actually was Two stages, roughly. First it read an enormous amount of text and learned to predict. That produces something knowledgeable and unhelpful — it will continue your sentence rather than answer your question. Then it was shaped by humans rating responses, teaching it that answering, being useful, and refusing certain things score well. This second stage is where its manners and its refusals come from — which is worth knowing, because it means the refusals reflect the judgement of the people who did the shaping. That is exactly why a general assistant's idea of appropriate is not your community's: different people, different brief. ## What it is not - It is not a search engine. It is not consulting sources as it answers, unless the product has specifically been built to do that. - It is not a database. There is no record being retrieved that could be checked. - It does not understand in the way a person does, and arguing about whether that word applies is less useful than knowing it cannot tell when it is wrong. - It is not a person, and it is not anyone. There is no one at the other end forming an impression of your child. ## What follows for a school or a family Three things, and they all fall out of the mechanism rather than from anyone's opinion. 1. Anything that matters must be checked elsewhere, because plausibility and accuracy are different properties and the model optimises the first. 2. Its refusals reflect whoever shaped it. If you want your own standard applied, it has to be applied around the model, by a system that checks requests and answers against rules you wrote. 3. Fluency is not evidence of anything. This is the single most useful thing to teach a child about it, and it does not expire as models improve. [See how the filtering works](https://navon.org/features) ## Common questions **How does an AI chatbot actually work?** It has learned in statistical detail which words tend to follow which, and answers by producing the most plausible continuation of your question one word at a time. It is not looking anything up. **Why does AI make up facts and sources?** Because producing plausible text and recalling true text are the same operation for it. A fabricated citation has the shape of thousands of real ones, and nothing in the mechanism distinguishes inventing from remembering. **Does AI understand what it is saying?** It cannot tell when it is wrong, which is the property that matters practically. Whether that counts as understanding is a less useful question than knowing it has no internal signal of its own accuracy. **Does an AI remember my child's conversations?** The model itself does not. Anything remembered between sessions is the product storing text and feeding it back — a design decision worth asking any vendor about directly. --- # A glossary of AI terms for schools Source: https://navon.org/blog/ai-glossary-for-schools Published: 2026-09-03 · Category: Explainers Every term you are likely to meet in a vendor conversation, defined plainly, with a note on why it matters rather than only what it means. Vendor calls and policy documents are full of terms used loosely and sometimes deliberately. These are the ones that turn up, defined plainly, each with a note on what it actually means for a school. ## The basics ### Large language model (LLM) The system underneath a chatbot. It has learned which words tend to follow which, and answers by producing plausible continuations rather than retrieving facts. Why it matters: this is why it invents citations, and why that is a property to manage rather than a bug awaiting a fix. ### Prompt What you type. Also used for the hidden instructions a product adds before your text — the "system prompt" — which is where a product's personality and some of its rules live. ### Token The chunks of text a model reads and writes, roughly three quarters of a word each. Why it matters: usage limits and prices are usually stated in tokens, so a "1 million token" allowance is about 750,000 words. ### Context window How much text the model can consider at once, including the whole conversation so far. Why it matters: in a long conversation, the earliest part eventually falls out, and the model appears to forget what it was told. ### Hallucination The industry's term for the model stating something false with confidence. A poor word — it implies a malfunction, when it is the ordinary mechanism producing a plausible-but-untrue continuation. ## Safety and filtering ### Content filter A check applied to requests, answers or both. Why it matters: ask which. A product that screens only the request cannot catch an unsuitable answer to a reasonable question, which is most of the risk with images. ### Pre-filter and post-scan Checking before generation, and checking the output afterwards. A serious product does both. ### Fail open / fail closed What happens when the safety check itself is unavailable. Fail open lets the request through unchecked; fail closed refuses it. Why it matters: this is the single most revealing question to ask a vendor, and a filter that fails open switches itself off exactly when nobody is watching. ### Guardrails A loose marketing term covering anything from a filter to a phrase in the system prompt. Why it matters: it means nothing on its own. Ask what the guardrails actually are and where they run. ### Alignment How well a model's behaviour matches its makers' intentions. Why it matters: note whose intentions. A well-aligned general model is aligned to its vendor's global policy, not to your school's standards. ### Jailbreak A prompt crafted to get around a model's restrictions, usually by role-play or hypothetical framing. Why it matters: students share these. A filter that runs outside the model, on the request and the answer, is not defeated by persuading the model of anything. ## Deployment ### Server-side vs client-side Whether something runs on the provider's machines or on the user's device. Why it matters: client-side filtering — an extension, a managed profile — protects a device. Server-side filtering is a property of the account and follows the student to any device. ### API The interface one program uses to call another. Why it matters mainly for a detail: content sent through a provider's API is usually governed by different terms from its consumer chat product, which is why "we do not train on your data" needs to be asked about the specific route your content takes. ### Gateway or reseller A company sitting between a product and the model providers. Why it matters: it is another party receiving your students' prompts, and it is easy to omit from a sub-processor list. Ask explicitly whether one is involved. ### Sub-processor Any third party handling your data on a vendor's behalf. Why it matters: you are entitled to the current list, and it should name the model providers and any gateway. ## Data ### Training data What a model learned from. Why it matters: the live question is whether your students' content becomes training data for anyone — the vendor, or the providers behind them. These are different questions and a vendor answering only the first has not answered. ### Zero retention A provider setting where content is not stored after answering. Why it matters: usually available on business tiers and not consumer ones. Ask whether it is enabled on the tier your school is actually on. ### Data residency Which country your data is stored in. Why it matters for schools with legal obligations about transfers; ask before it becomes a question someone else asks you. ## Terms used loosely - "AI-powered" — describes nothing. Ask what it does. - "Safe for schools" — a claim, not a specification. Ask what is checked, by whom, and what happens when the check fails. - "Enterprise-grade security" — no fixed meaning. Ask for the specific control you care about. - "Kid-safe" — same. Ask whether you can add your own blocked topics and whether they reach the image tool. - "Custom AI" — usually a system prompt, not a trained model. Worth clarifying, because the two differ enormously in what they can guarantee. [Ask us anything here](https://navon.org/contact-us) ## Common questions **What does fail closed mean in AI filtering?** That when the safety check itself cannot run, the request is refused rather than allowed through unchecked. The alternative, failing open, means filtering silently stops during an outage. **What is a hallucination in AI?** The industry term for a model stating something false with confidence. The word implies a malfunction, but it is the ordinary mechanism producing a plausible continuation that happens not to be true. **What is an AI gateway or reseller?** A company sitting between a product and the model providers. It is another party receiving your students' prompts and is easily omitted from a sub-processor list, so ask explicitly whether one is involved. --- # Where your child's AI conversations actually go Source: https://navon.org/blog/where-your-childs-ai-conversations-go Published: 2026-09-03 · Category: For families "We take your privacy seriously" tells you nothing. Here is what actually happens to the text, and the four questions that produce real answers. When a child types a question into an AI tool, that text goes somewhere and is handled by more parties than most parents expect. None of this is sinister; it is how the products are built. But it is worth knowing accurately, because the vague reassurance in most privacy pages is not an answer. This describes the general shape. The specifics vary by product, which is exactly why the questions at the end matter. ## The journey of a single question 1. The text leaves the device, encrypted, and reaches the product's servers. 2. The product may check it against a content filter — often by sending it to another company's model to be judged. 3. It is sent to whichever company actually runs the model. This is frequently not the company whose name is on the app. 4. There may be a gateway or reseller in between, which is a further party receiving the text. 5. The answer comes back, possibly gets checked again, and is shown to the child. 6. Some or all of it is stored: the conversation history, and often a log for abuse prevention. So a single question can be handled by three or four organisations. A parent asking "who sees this" and being told "only us" has been given an answer that is unlikely to be complete. ## The training question, asked properly This is where careful wording does a lot of work, and it is worth learning to hear it. "We do not train our models on your data" can be entirely true while your child's conversations are still used for training — by the company that actually runs the model, under their terms, because the product is simply a customer of theirs. So the question to ask is not whether the vendor trains on it. It is whether anyone in the chain does, including the model providers and any gateway. A vendor who has arranged for that not to happen will say so specifically and can point to the setting or the agreement. A vendor who has not will answer the narrower question. ## How long it is kept Usually longer than parents assume, and for reasons that are often legitimate — abuse investigation, debugging, the child's own history. The reasonable expectations are a stated period rather than "as long as necessary", and a way to delete a conversation that actually deletes it rather than hiding it from the interface. Worth asking specifically whether deleting a conversation removes it from the providers downstream, because frequently it does not. ## What a school or parent can see A separate question from what the company can see, and there is a real tension in it. Full visibility of everything a teenager writes will end their honesty and is rarely the right parenting decision. No visibility at all leaves an adult unable to notice a pattern that matters. The workable middle is a record of decisions — what was blocked, roughly what category, how often — which is enough to prompt a conversation and not a transcript of a child's private questions. Whichever a product does, you should be able to find out in a sentence, and your child should be told. Monitoring a child has not been told about tends to be discovered, and what it costs is not the monitoring. ## The four questions 1. Which companies receive what my child types — including any gateway or reseller, by name? 2. Is that content used to train any model, by you or by them? Not just by you. 3. How long is it kept, and does deleting a conversation delete it everywhere? 4. What can I see, what can the school see, and does my child know? A product built with children in mind can answer all four in a few sentences. Difficulty answering is itself informative, and worth more than any privacy page. ## What to tell your child One idea, and it holds regardless of the product: it is not private, and it is not a friend. Anything that genuinely matters should go to a person who can actually notice and act. This is worth saying even about a well-designed product, because the habit protects them in every other tool they will ever use. [Ask us the four](https://navon.org/contact-us) ## Common questions **Who can see what my child types into an AI chatbot?** Typically several organisations: the product, whichever company runs the model, any gateway or reseller in between, and sometimes a separate provider doing content filtering. A vendor answering 'only us' has probably not described the full chain. **Is my child's AI conversation used for training?** Ask whether anyone in the chain uses it, not just the vendor. 'We do not train on your data' can be true while the company actually running the model does, under their own terms. **Can I delete my child's AI conversations?** Usually from the interface, but ask specifically whether that removes it from the providers downstream. Frequently it does not. --- # Scheduling AI availability around Shabbos and Yom Tov Source: https://navon.org/blog/scheduling-ai-around-shabbos Published: 2026-09-03 · Category: For families A weekly window is straightforward. Yom Tov is not, and neither is a family three timezones from the server. Here is what a scheduling feature actually has to handle. Let us be clear about scope at the outset. The halachic questions around technology and Shabbos are real, they are not simple, and they belong to your rav. Nothing here is a psak and nothing here should be read as one. What this covers is the practical layer underneath: if a family or a school has decided a tool should be unavailable over Shabbos and Yom Tov, what does a scheduling feature have to get right to actually deliver that? More products get this wrong than you would expect. ## Why a weekly window is not enough Most parental-control scheduling offers a repeating weekly pattern: these hours, these days. That handles Shabbos approximately and Yom Tov not at all. - Shabbos begins and ends at different times each week, and by a different amount depending on latitude. A fixed Friday cutoff is either early most of the year or late some of it. - Yom Tov does not fall on the same weekday each year, so no weekly repeating rule expresses it. - A two-day Yom Tov adjacent to Shabbos produces a three-day period that no weekly pattern describes. - Chol hamoed sits between, and a family may want different treatment. So a genuinely useful implementation needs either a calendar the product understands, or the ability to set a specific one-off period — and honestly, most products have neither. ## The timezone problem, which is not optional A window like "unavailable from 18:00 Friday" is meaningless without knowing whose clock. If a product stores the hours but not the zone, it evaluates them against its server's clock — and a family several hours away gets a window at the wrong time of day, every single week. This is a common and quiet bug. The setting saves, the interface looks right, and the enforcement happens at the wrong moment. It is worth testing directly: set a window that should be active now and confirm the tool is actually unavailable. The zone also has to be set by the adult and stored on the server. If it is taken from the child's device, a child who changes their device timezone changes the window. ## What enforcement has to look like For a scheduled unavailability to mean anything: - It has to be checked on the server when a request is made, not counted down by an app. A second device otherwise has its own idea of the time. - It has to apply to every tool, not only chat. A window that stops conversation but permits image generation is not a window. - It should say why. A child who hits a silent failure concludes the product is broken; one who is told a guardian set these hours knows a boundary exists and who to ask. - The child must not be able to change it, including indirectly through a device setting. ## What a schedule cannot do It governs one product. A phone with other applications on it is not addressed by any setting inside a single tool, and it would be dishonest to imply otherwise. A scheduled window makes one tool behave the way a family has decided it should. Whatever else is on the device is a separate conversation, and mostly not a technical one. [See the family controls](https://navon.org/services/family) ## Common questions **Can AI tools be scheduled to turn off for Shabbos?** Some support a repeating weekly window, which approximates Shabbos but cannot express Yom Tov, since those do not fall on fixed weekdays. What matters technically is that the window is enforced on the server in a timezone the guardian set, and applies to every tool. **Why do time windows need a timezone?** Hours are meaningless without knowing whose clock. A product storing hours but not the zone evaluates them against its server clock, so a family in another timezone gets the window at the wrong time of day every week. **Is it permitted to use AI on Shabbos?** That is a halachic question for your rav, not something an article or a vendor should answer. What a product can offer is the ability to make the tool unavailable during a period you have decided on. --- # What to do if you find your child has been using AI secretly Source: https://navon.org/blog/child-using-ai-secretly Published: 2026-09-03 · Category: For families You have found something you were not meant to find. What happens in the next twenty minutes determines whether you learn anything, or only that they get better at hiding it. The discovery is usually accidental — a browser tab, an essay that does not sound like them, a younger sibling mentioning it. The instinct is to address the concealment immediately and firmly. It is worth slowing down, because the concealment is information, and how you handle the first conversation determines whether you get any more of it. ## First, what was it actually about? Children conceal AI use for several quite different reasons, and they call for different responses. ### There was a rule and they broke it The simplest case. Ordinary boundary-testing, handled like any other broken rule. Notably, this is the least common of the four. ### There was no rule and they assumed you would disapprove Very common, and worth recognising as such. The child has inferred a rule from your general attitude to technology and hidden it pre-emptively. Punishing this teaches them the inference was right. ### They were embarrassed by what they asked The concealment is about the subject, not the tool. Children ask machines things they will not ask people, precisely because there is no one to be embarrassed in front of. This is worth knowing about, and it is a reason to stay calm — the topic is the thing that matters. ### They were struggling and it was the only help available The one most often missed. A child falling behind, who does not want to say so, finds something that answers without judging. The AI use is a symptom. You cannot tell which from the artefact you found. You can only tell by asking, which is why the first conversation should be a question rather than a verdict. ## The first twenty minutes - Do not open with the consequence. Once a child is defending themselves you will learn nothing further. - Ask what they used it for, and mean it. "Show me" works better than "what were you doing". - Say the true thing about your reaction. "I was surprised and I want to understand" is honest and does not put them on trial. - Separate the two issues. Whether AI use was acceptable, and whether hiding it was, are different conversations. Running them together guarantees the second swallows the first. - If the work was submitted as their own, that has to be addressed. But address it as its own thing, not as evidence about their character. ## What not to do Do not remove access entirely as a first response, unless what you found genuinely warrants it. It reliably produces the same behaviour on a device you do not know about, and you lose the visibility you just gained. Do not read everything. The temptation is strong and the cost is high — a child who discovers they are being read stops writing anything real anywhere you can see, and you will not get that back. Do not treat it as evidence of a character problem. Most of these situations are a child solving a problem in the way that was available, and the useful response addresses the problem. ## Afterwards Set the rule together rather than announcing it. Children follow rules they helped write, and this is exactly the moment when they are paying attention. If a filtered tool with limits you set would help, say what it does and why — including that you will be able to see when something is blocked. A control your child knows about works as a boundary. One they do not know about works only until it is discovered, and then it works against you. And if the concealment turned out to be about schoolwork they could not manage, the AI conversation is the smaller one. [See the family controls](https://navon.org/services/family) ## Common questions **My child used AI to do their homework. What should I do?** Find out what it was about before setting a consequence. Children conceal AI use because a rule was broken, because they assumed disapproval, because the subject embarrassed them, or because they were struggling and it was the only help available. The responses differ. **Should I take away AI access if my child hid it?** Rarely as a first response. It reliably moves the behaviour to a device you do not know about, and you lose the visibility you just gained. Setting rules together works better than removing access. **Should I read my child's AI conversations?** Reading everything costs more than it gains: a child who discovers it stops writing anything real where you can see. A record of what was blocked and how often is usually enough to notice a pattern worth discussing.