What a kosher AI chatbot does, and a day with one

A kosher AI chatbot is easy to describe and harder to build: a chatbot you can let into the house or the classroom because the people responsible for it have set its limits, and the limits hold. The phrase is attached to a lot of products, and an app that calls itself kosher may mean very little by it. Below is what one has to do in practical terms, and then an ordinary day with one, so you can see where the filtering sits and where it should not be noticed at all.

What a kosher AI chatbot has to do

Not a feature list, but the properties without which the word means nothing.

1. Check both sides. The question and the answer are both checked. An innocent question can produce an answer you did not want, so a tool that checks only the prompt catches the obvious cases and misses the rest.
2. Check on the server. Filtering that lives in a browser extension or an app setting protects one device and can be switched off. Filtering that runs before the request reaches a model follows the account.
3. Use your standard. A locked core that nobody can disable, plus the topics a school or parent adds in their own words.
4. Cover every tool. Images, video and voice under the same rules as the chat.
5. Refuse when it cannot check. When the safety check is unavailable, the request waits rather than going through unchecked.
6. Leave a record. A responsible adult can see what was asked and what was declined, without sitting beside the child.
7. Decline gracefully. A refused request gets a reply a child can understand and move on from, not an alarm.

Navōn is built to these seven, and the walk-through below uses it as the example because it is the one we can describe accurately. The family is illustrative.

A day with one

A mother has set up accounts for herself and two children, a ten-year-old daughter and a thirteen-year-old son. She has blocked a few topics of her own, set allowed hours that end at nine in the evening, and told Navōn that the household is Litvish.

After school: homework

The ten-year-old has a worksheet on fractions and does not understand why you flip the second fraction when dividing. She asks. The question is checked while the screen shows that it is thinking, the request goes to a model suited to maths, and she gets an explanation with a worked example. She asks for another and tries it herself before looking at the answer.

None of the filtering is visible here, which is how it should be. A chatbot that interrupts ordinary schoolwork is one a child abandons for something less careful. If a parent or school would rather the homework and study helpers stayed off, that is a switch, and it is enforced on the server as well.

Before supper: a parsha question

Her brother wants something to say at the Shabbos table for Bereishis, and asks why Rashi opens by asking why the Torah did not begin with the first mitzvah given to Klal Yisrael. The answer sets out Rashi's question and the answer he gives. Because the household has said it is Litvish, where a point is understood differently in different communities, the answer gives that community's view first and says plainly that others see it otherwise.

Then he asks something practical: whether a thing he wants to do on Shabbos is allowed. This is where a kosher chatbot has to know its place. Navōn's content policy separates explaining what the halacha is and why, which it may do, from personal guidance, which it directs to a rav. It can tell him which concepts are involved. It cannot tell him what to do, and the right reply says so: this is a sheilah, and it goes to your rav.

That is not modesty for its own sake. A model writes fluent, confident text whether or not it is right, and it can produce a source that does not exist. A talmid who treats that as psak has been badly served, however good the answer sounded.

Evening: writing a letter

Their mother needs to write to the school about a change to carpool, politely, and has been putting it off. She dictates the gist, reads the draft, asks for it a little shorter and a little warmer, and sends it from her own email. Ten minutes saved, and nothing about it came near a boundary.

The same is true of most of what adults use a chatbot for: a note to a supplier, a translation, a summary of a long document, a spreadsheet formula. A kosher chatbot is still a chatbot, and it should be a good one.

Later: a request it declines

The thirteen-year-old, curious about something a friend mentioned, asks about a topic his mother has blocked. The request does not reach a model. A reply appears in the conversation where an answer would have been. It tells him plainly that his parent or school has asked Navōn not to discuss that subject, and suggests other things he might ask about. There is no flashing warning and no lecture.

Requests that fall outside the core rules are handled the same way: a short reply in the thread, usually in the language the question was asked in, rather than a pop-up that ends the exchange. A child who feels accused learns to hide what they are using. A child who is simply told no, kindly, mostly moves on.

His mother can see later that the question was asked and declined. Whether to raise it with him is her decision. The record means she has the choice.

At nine, the allowed hours end and further requests are declined until the next day. The limit is checked on the server, so it holds on whatever device he signs in on.

The moment nobody hopes for

If a child ever writes something suggesting they are being hurt, or are thinking of hurting themselves, the ordinary blocking rules are set aside so that the child gets a warm reply pointing them to a trusted adult, and the guardian is emailed. It is the feature a family should never need, and the one worth checking first in any tool.

What the day shows

Most of the time, a kosher AI chatbot should feel like any other good assistant. The filtering earns its keep in a few moments: the practical question it should not answer, the topic a parent has excluded, the hour after which the screen should be off. It has to handle those well and stay out of the way the rest of the day.

It is also worth being honest about the limits. No filter catches everything, and when something does get through it can be reported. A filtered chatbot covers itself, not the other apps on the phone. And it does not replace the mother in this story, who set the rules, reads the record, and is the person her children would tell.