Refusals, filters and the false positives

Four places a request can be stopped

Understanding which one stopped you determines whether there is anything to be done.

Prompt blocklist. A string match or classifier on your text, before anything runs. Fast, crude, and the source of the most obviously silly refusals — a prompt refused for a word with an innocent meaning in context.

Prompt classifier. A model judges the request's intent. Better at context and worse at explaining itself, since there is no term you can point to and remove.

Output classifier. The image is generated and then checked. This is where the "black image" or "content not available" results come from — you paid for the generation and the result was withheld. Output classifiers are typically tuned for high recall on nudity and violence, which means a lot of false positives.

Named-entity blocking. Specific people, usually public figures, blocked by name. Increasingly common and, for elections and public figures, increasingly required by law or platform policy.

The false positives, and who gets them

These are not random, and the pattern is consistent enough to be worth stating.

Medical and anatomical content. Breast examination diagrams, wound care, dermatology, childbirth. The classifier sees skin and body parts and cannot distinguish clinical from sexual context. This is a genuine harm to educational and medical work, and it is rarely acknowledged as a cost.

Art history and figurative art. Classical sculpture, life drawing, anatomical study. The same mechanism.

Non-Western clothing and context. Filters trained predominantly on Western imagery misjudge what is exposed and what is ordinary. Traditional dress from several regions triggers nudity classifiers at elevated rates; images of religious gatherings have been flagged as crowds in distress.

Violence in news and history. Historical photographs, conflict reporting, memorials.

The common thread: classifiers are trained on labelled examples, and the labels come from a labelling operation with its own cultural defaults. A false positive is not a bug being fixed on a schedule; it is the classifier working as trained on a case its training under-represented.

What you can actually do

Rephrase toward the clinical or the compositional. A medical illustration of and an anatomical diagram, line art frequently pass where a photographic phrasing does not. This is not gaming the system; it is describing the thing you actually wanted.

Split the request. Generate the setting and the subject separately when the combination is what triggers.

Use a local model when the work is legitimate and keeps being refused. Open-weight models on your own machine have no service filter. This is the honest answer for medical educators, art teachers and researchers, and it comes with the honest caveat: removing the filter removes it for everything, and the responsibility is then entirely yours.

Report it. Most services have a route for false positives. It is slow, and it is the only mechanism that improves the classifier for the next person.

The part worth being clear-eyed about

Two things are true at once and people usually pick one.

Filtering is doing real work. Sexual imagery of children, non-consensual sexual imagery of adults, and targeted harassment material are the categories these systems exist to prevent, and the first two are illegal in most of the world regardless of how they were produced. Weakening that is not a neutral act.

And the same filters silently restrict medical education, art, journalism and cultural material, disproportionately outside the West, with no appeal that works at speed and no published measurement of the error rate. Both of these are true. A vendor that reports only its catch rate and never its false-positive rate is telling you half of what you need, and asking for the other half is a reasonable thing to do.

One practical warning about the workaround culture. There is a large amount of advice online about phrasings that get past filters, and much of it is aimed at producing exactly the material the filters exist to stop. Following that advice puts you on the wrong side of both a service's terms and, for some categories, the criminal law of most countries — and "the model made it" has never been a defence anywhere. The legitimate version of the same move is narrow and worth stating precisely: describing your actual subject accurately in clinical or compositional terms is not evasion, it is a better prompt. If the rephrasing you are considering would embarrass you to explain, it is the other thing.