Tznius and AI image generation

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.