A kosher ChatGPT alternative: what changes and what stays the same

People searching for a kosher ChatGPT are rarely asking for a weaker tool. They want what ChatGPT does well, the drafting, explaining, translating and summarising, without what it will also do for anyone who asks. A kosher ChatGPT alternative is worth considering if that describes you, and worth approaching clear-eyed, because the word kosher promises more than any software can deliver.

This is an honest account of who should look, what changes when you switch, what stays the same, and what no filtered tool can do.

Who should be looking for a kosher ChatGPT alternative

Not everyone. An adult using AI for work, comfortable with a general-purpose policy and willing to close the tab if something unwelcome appears, may be fine where they are. The case for an alternative is strongest where that description stops fitting.

• Parents of children who will use AI whether or not anyone has planned for it. ChatGPT's content policy is written by its vendor for a very broad audience, and it is not something a parent can rewrite in their own words.
• Schools that cannot sanction a tool through which a talmid can reach any subject, and whose staff need to see what was declined.
• Frum workplaces where staff use AI for drafting and research, and the office would rather it did not produce text or pictures nobody would want on a shared screen.
• Anyone who has assumed their internet filter covers it. A filter decides which addresses to allow. Once it has allowed a chat site, every conversation inside is one permitted connection, and the filter cannot see the questions or the answers.

What changes

Who writes the standard

In ChatGPT, the policy is the vendor's and applies to everyone identically. In Navōn, a core set of categories is locked for every account and cannot be disabled by anyone, and above that the school or parent adds their own topics, in their own words, and sets categories to blocked, guided or allowed. A household can say which tradition it follows, so a question in halacha or hashkafa is answered from that community's practice first, with a plain note where others hold otherwise.

Both sides are checked, on the server

Every prompt, including the text of an attached document and any uploaded picture, is checked before it reaches a model. Every answer is scanned when it is complete and replaced on screen if it fails. The checks run on Navōn's servers, so they follow the account to whatever device it is used on, and there is nothing on the device to turn off.

Images and video follow the same rules

A general image generator has no concept of tznius. In Navōn, image and video prompts pass through the same check as chat, generated pictures are scanned before they are saved, and people in them are dressed modestly.

Someone responsible can see what happened

A parent or school can see what was asked and what was declined. Guardians can set allowed hours and a daily limit, and both are enforced on the server. If a child writes something suggesting distress, the child gets a reply pointing them to a trusted adult and the guardian is emailed.

It refuses when it cannot check

If the safety check is unavailable, Navōn refuses the request and says why, rather than answering unchecked. The cost is an occasional pause. The alternative is filtering that switches itself off at exactly the moment nobody is watching.

Refusals are part of the conversation

When a request is declined, the reply appears in the thread where the answer would have been, in the language it was asked in, and suggests something else to do. There is no alarm and no lecture, which matters more than it sounds: a tool that makes a child feel accused is a tool they stop telling you about.

What stays the same

The usual worry about switching is that a filtered tool must be a weaker one. It need not be. Navōn does not run a small model of its own. On Auto, it reads each request and routes it to one of several leading models from Anthropic, OpenAI and Google, choosing one suited to long or careful writing, hard reasoning and maths, code, or a question that needs current information from the web, and it shows which model answered. You can also pick a model yourself.

So the ordinary work carries on as before. Drafting an email, tidying a letter, explaining a concept in chemistry or a difficult Rashi, translating, summarising a long document, writing a spreadsheet formula. Dictation works in English and Hebrew, and answers can be read aloud.

What you give up is narrow: the material your standard excludes, and now and then a borderline request that a stricter filter declines where a looser one would not. For most families and schools, that is the point of switching.

What any filtered tool cannot do

Worth stating plainly, because the word kosher invites more confidence than it should.

• It cannot be perfect. Filtering is judgement at a boundary, done at scale, and every system misses something sometimes. A vendor who says nothing can ever get through is describing a product that does not exist.
• It cannot reach other tools. The filter covers Navōn. It does not cover a different chatbot in another tab or another app on the phone.
• It cannot pasken. A model writes fluent, confident text and can be wrong or invent a source. It can help you learn and prepare. A practical sheilah goes to your rav.
• It cannot carry a hechsher. There is no certification for software, and Navōn does not claim one.
• It cannot replace you. A filtered tool makes one option trustworthy enough to permit. A child still needs an adult who knows what they use, and whom they would tell if something went wrong.

How to judge any alternative

Whichever tool you consider, test it against your own standard rather than its marketing. Add a topic your family genuinely does not discuss, then try to reach it, both directly and described in other words. Try the image tool, not only the chat. Sign in on a second device and see whether the rules came with you. And ask the vendor, in writing, what happens when the safety check cannot run. The answer to that last question tells you what they actually prioritise.