The bias, and what mitigating it actually does

What has been measured

This is one of the better-documented properties of image models, because it is easy to test at scale: generate a few hundred images for an occupation prompt and classify what comes back.

The consistent findings across several published audits:

• Occupation prompts return images skewed well beyond real-world workforce statistics, in the direction of the stereotype. Prompts for high-status professions return predominantly light-skinned men; prompts for care and service work return predominantly women.
• Skin-tone distributions are compressed toward the lighter end of standard scales, including for prompts that specify a country where that is not representative.
• Prompts naming a country return the most touristically photographed version of it. An Indian street returns a crowded market with saturated colour; a German street returns clean cobbles. Neither is false; both are a magazine's idea of the place.
• Adding an attribute like poor or criminal shifts the demographic distribution markedly, which is the most troubling finding of the group because it shows the association is available and reachable.

None of this requires malice anywhere in the pipeline. It is the corpus, and the mode-seeking behaviour of the last lesson, working exactly as described.

What the mitigations do

Three families are in use, and each has a cost worth knowing.

Prompt injection. The service silently adds or substitutes demographic terms. It is cheap, it works on the immediate complaint, and it is unaware of context. This is the mechanism behind well-publicised incidents where diversity terms were applied to historical prompts and produced images that were absurd or offensive. The failure was not the goal; it was applying a blunt textual rule to every request without knowing which requests it made sense for.

Retraining and rebalancing. Filter or reweight the training data. This is the principled route and it is expensive, slow, and only as good as the labels used to rebalance — which means somebody has decided which categories exist and what the target proportions are. Those are contested choices, and they are usually made without publication.

User-level specification. Push the decision back to the person prompting. Honest, effective and unevenly distributed: it works if you know to do it, which puts the burden on the people most likely to be misrepresented.

What this means for your work

Two practical positions, and they are not in tension.

If you are producing images of people for anything public, specify. Leaving it to the default means shipping the corpus's assumptions under your name, and "the model chose it" is not a defence anybody accepts. This is the same discipline as casting: someone always decides, and pretending otherwise just hides the decision.

If you are evaluating a tool or a vendor's claim, ask what was measured and how. "We have addressed bias" is not a claim you can check. "Here is the distribution over skin tone for these fifty occupation prompts, before and after" is. Insist on the second form. The methods for measuring this are public and cheap; a supplier who has not run them has not looked.

The part that stays unresolved

There is no agreed answer to what the right output distribution is, and it is worth being clear that this is a genuine disagreement rather than a gap waiting to be filled.

Should a doctor reflect the real demographics of doctors in the user's country, which vary enormously? The global figures? An equal split, on the grounds that a generated image is not a statistic? The demographics of the people likely to see it? Each has been argued seriously, each implies different outputs, and no principle chooses between them.

What can be said without controversy is narrower and still useful: the unmitigated default is more skewed than any of the candidate answers, the skew is measurable, mitigation by silent prompt injection has produced visible failures, and the choice is currently being made by vendors without publication. Anyone who tells you this is a solved problem is describing their preferred answer as though it were the only one.

One further note for anyone working in a country the corpus under-represents. You will find that the model's idea of your own surroundings is a foreigner's, and that specifying harder produces a caricature rather than a correction, because the specific vocabulary has thin coverage. The route that works is the one from the previous module: photograph your own references and condition on them, or train a small fine-tune on a few dozen of your own images. Twenty photographs taken on a phone are enough to move a model's idea of a particular kind of street, shop or garment further than any prompt will. It is more work than it should be, and it is work you can actually do.