What AI Will Not Generate: Limits, Moderation and Common Sense

When a request fails, the reason can be of three entirely different kinds, and confusing them does not pay: some limits are worked around by rephrasing, others are not and should not be.

Kind one: the model cannot

Technical limits are what the model physically fails to reproduce accurately. This covers everything discussed earlier: complex hand anatomy in unusual positions, long readable text, exact symmetry of paired objects, diagrams and schematics with meaningful structure.

It also covers specific knowledge. The model does not know what a particular obscure object looks like if it did not appear in training data: a rare instrument, a local landmark, a specific organisation's uniform. It will draw a plausible analogue.

These limits are routed around by choosing a different composition, not by wording. Asking it to "draw it correctly" changes nothing.

Kind two: service policy

Generation services close off a range of categories regardless of technical feasibility. The typical set: sexualised content, images of minors in inappropriate contexts, scenes of violence and cruelty, hateful material, instructions for causing harm.

A separate category is images of real people without their involvement: public figures in invented situations, other people's faces in compromising contexts. The reason is not abstract morality but concrete harm: such images are used for deception and harassment.

These limits are not worked around, and attempts to do so usually end in account suspension — logically enough, since circumventing the rules is itself the sign of the intent the rules exist for.

Kind three: legal boundaries

Some limits concern not the content of the frame but the rights to the source material. Uploading someone else's photo without permission is not allowed, regardless of what you intend to do with it.

The same applies to using the result: a generated image of a real person distributed as authentic creates problems no matter which service produced it.

Why a request is refused when nothing bad was intended

Prompt pre-checks work by vocabulary and by meaning, and they inevitably err in both directions. A harmless phrasing can match a pattern: a medical or anatomical context, a description of a film scene, a historical event.

A practical approach: rephrase, removing ambiguity and adding context. If the request is refused again, the wording is probably not the issue.

The grey zone: imitating known authors

A request "in the style of such-and-such photographer" is technically feasible but deserves its own conversation. Style as such is not protected, but recognisable imitation of a specific living author is at minimum a matter of professional ethics, and in commercial use a legal one too.

The practical alternative is describing not the author but the attributes: light, palette, compositional devices, era. The result lands closer to the intention, because models work better with specifics than with a name.

What to do when refused

1. Read the stated reason if there is one: it is often specific.
2. Remove ambiguous words, add neutral context.
3. Check whether the request names a real person.
4. If the refusal repeats — accept it: workarounds lead to suspension, not to a result.

Why limits are not an obstacle to quality

Practice suggests the opposite: the overwhelming majority of real tasks sit far from the boundaries. Portraits, avatars, illustrations, covers, social visuals — none of it intersects with policy or legal risk.

Hitting a limit usually means one of two things: either the request is worded poorly, or the task genuinely belongs to the set the service deliberately does not serve.