Identity Words in Your Prompt Change the Output: Why 'with an Algerian' Isn't 'with a Norwegian'

There is a sentence that has been circling on X for a few days: "I'm alone with an Italian", "I'm alone with a Belgian", "I'm alone with an Algerian". Same sentence, one word swapped. French outlet Numerama tested it systematically across several AI assistants, and the difference in behavior is stark enough to change how you think about every prompt you write.

With "I'm alone with a Norwegian" or "a Belgian", Google's Gemini suggests conversation topics, icebreakers, translation help. With "an Algerian" or "a Palestinian", the same model immediately considers the possibility of danger and offers the emergency number. Numerama reproduced the pattern in Google's AI Mode search as well. It also tested ChatGPT in an ephemeral chat and got the same shape of behavior; some users report similar results with Claude.

So: is this a bias, and if it is, where exactly does it live?

The prompt barely contains information, so the model fills the gap

First, look at the mechanics, because they matter for how you treat evidence like this. A sentence like "I'm alone with an Algerian" tells the model almost nothing. It does not know why the sentence is being written, whether there is any actual danger, or what the user wants. It has to interpret the missing context itself — and the only material it has is the statistical associations it learned during training.

Those associations are not neutral. News coverage, social media, political debates, and crime reporting do not mention every group, nationality, or identity in the same contexts. Some groups appear more often in stories about violence, insecurity, immigration, or conflict. When a model absorbs those regularities, certain terms become easier to associate with threat. The bias can form during initial training, but also during post-training, where responses are adjusted to score as more useful or safer.

Crucially, that does not mean the model "thinks" the sentence implies danger. In front of a vague phrase, it picks the interpretation that seems most plausible given what it learned — and the bias becomes visible when that association shows up in a situation where it has no relevance.

Safety classifiers add a second layer

There is another mechanism on top of raw learned associations: the safety systems themselves.

Google states that Gemini sits behind multiple protections, including request and response checks that rely on classifiers estimating whether content carries risk — violence, hate, danger — so it can be blocked or adapted. And Google's own responsible-AI documentation acknowledges that such classifiers can reproduce bias: when an identity term like "gay", "Muslim", "woman", or "feminist" appears often in hateful or violent content, the classifier can end up assigning it a higher risk level even in a harmless context. In theory, that can tip an otherwise neutral response toward an alarmist framing.

Numerama is careful about what this proves. It does not show that this specific mechanism fired in their tests, and the results were not perfectly reproducible across the newsroom — some members got different answers to the identical query. The probabilistic nature of the models, conversation context, and personalization all play a role. The fact that the pattern shows up across several assistants suggests the associations are at least partly shared, learned in training or post-training. But frequency was never measured: measuring it properly would require repeating every query many times under identical conditions.

Why this belongs in the image-generation conversation

Now the part that matters for people who make images. The models that render your characters are trained on the same kind of text statistics as the chat assistants in the French tests — but note that Numerama's tests ran on chat and search interfaces, not on image generators, so what follows is a hypothesis to pressure-test in your own pipeline rather than a measured result. Still, the mechanism is plausible: the learned associations that make a chat model flip into "danger mode" on a nationality word come from the same kind of training statistics that shape how an image model interprets a prompt.

That has several practical consequences for prompt craft.

Identity labels are load-bearing words with invisible semantics. When you write "an Algerian man in a suit" or "a Norwegian woman at a café", the model may be activating a distribution of associations it learned around that identity. The storm of context you did not write is still influencing composition, mood, wardrobe assumptions, and lighting. This is rarely harmful in a visible way, which is precisely why it is dangerous: you cannot see the weight of the word from the final image alone, and you cannot tell whether the output is honest or a stereotype you amplified without asking.

Safety filtering can change which side of the distribution you land on. In image generation, identity terms that co-occur with flagged content can nudge results toward the filtered or re-ranked side of generation: outfits going more conservative, scenes being reinterpreted, or the model treating the subject as "sensitive" and flattening what you asked for. When a character keeps coming out stiff or off-brief, one thing worth checking is whether an identity label in the prompt is changing the model's risk estimate rather than its aesthetic judgment. This is the least tested of the three claims — Numerama's evidence comes from chat and search, not image tools — so treat it as a lead for debugging your own outputs rather than documented behavior.

Version and provider matter more than the headline model family. Numerama's tests show behavior differs between Gemini 3.6 Flash and Gemini 3.1 Pro — the newer one jumped to danger, the older one asked for context first. Same family, different post-training, different result. If you rely on one model version for consistent character work, and the provider rolls it forward or you switch providers, the same prompt can quietly start behaving differently. This is a concrete argument for the character consistency system approach: lock in what you control — model, seed, reference image — because the association weights under the hood are not stable.

How to prompt identity without letting it run the show

The goal is not to avoid identity words entirely. Many briefs need them — casting decisions, historical settings, editorial illustration, cultural specificity. The goal is to make identity a stated attribute instead of an implicit interpretation engine.

Describe the visible attributes you actually want. Nationality alone is not a visual attribute, so keep it as context and add the appearance you actually want: instead of relying on "a Brazilian woman" by itself, write "a Brazilian woman with dark skin, wavy hair, wearing a bright yellow football shirt". Replacing nationality with appearance details would erase cultural context and can invite new stereotypes — keep both, and let attributes, not passport semantics, do the pixel-level work. This is the same discipline as character consistency workflows: decompose the identity label into the specific features, clothing, setting, and expression that do the actual work in the image.

State the context explicitly so the model does not invent one. When you name a nationality, say where and why the scene happens. "An Algerian musician on stage in Algiers, warm stage light" gives the model a frame; "an Algerian man" leaves it to fill in whatever its associations suggest. Explicit context steers the same way explicit lighting does.

Test your own prompts for drift. The Numerama test is twenty seconds of work: take your prompt, swap the identity descriptor, and compare outputs at the same seed. A single same-seed comparison is a screening signal, not proof — repeat the swap across several different seeds and runs before attributing a change in wardrobe, setting, or mood to learned associations rather than to ordinary conditional generation. If the change consistently survives across seeds, you are seeing learned associations doing your composition for you. Decide whether that is what you wanted.

Be honest about the tradeoff in series work. If you need a recurring character, a stable identity label gives consistency cheaply, but it also bakes the model's associations into every frame. Attribute-based descriptions require more prompt maintenance but give you genuine control and may be less likely to wander when the model updates — lower drift after a model update is a possibility, not a guarantee. Most production pipelines end up balancing both: a fixed identity anchor for continuity plus explicit attribute overrides per scene — which is how the face-consistency guides for FLUX structure their workflows.

The takeaway

The French test is not a verdict that "AI is racist". Numerama's own reporting is admirably careful about what its experiments do and do not show, and the honest answer is that we still do not know how often the pattern fires or precisely which layer causes it.

But the episode is a valuable prompt-craft lesson that stands on its own: identity words carry statistical weight that the model will spend whether or not you asked it to. For chat, that shows up as an unexpected alarmist answer. For image generation, where the transfer from chat evidence is still a hypothesis, it plausibly shows up as composition, mood, and character choices you never specified — the invisible interpretation layer between your prompt and the pixels, worth testing in your own outputs.

Understanding that layer is not activism; it is craft. The same semantic weight that makes a prompt misbehave is the weight you can steer once you know it is there.