An artificial-intelligence system can have no face, voice or declared identity and still be assigned a race by the people evaluating it. Across five primary studies and eight supplementary studies, Monica Gamez-Djokic and Adam Waytz found that participants overwhelmingly associated AI with Whiteness.[1][2]
The result matters because the classification did not remain a harmless mental image. Perceived Whiteness was linked to judgments of competence, trustworthiness and persuasiveness. In one experiment, changing the described racial composition of a model’s training data changed the kinds of work participants assigned to it.[1]
What the study tested
| Measure | Value | Context |
|---|---|---|
| Primary studies | Five | Used forced-choice, open-ended and implicit measures, then examined consequences for evaluation and task allocation.Source 1 |
| Supplementary studies | Eight | Extended and stress-tested the central pattern.Source 1 |
| Geographic extension | United States, Japan and India | The association was not confined to a single US sample.Source 1 |
How can a system without a body acquire a race?
People do not evaluate technology in a social vacuum. They bring cultural associations about who creates advanced technology, who is presented as an expert and whose speech or appearance has historically been treated as authoritative. A category such as “artificial intelligence” can therefore inherit a social prototype even when the interface supplies no explicit cue.[2]
Two locations for bias
| Where bias enters | Example | What an audit must examine |
|---|---|---|
| Inside the system | Training data or model design produces different outputs for different groups. | Performance, representation, error rates and downstream decisions.Source 2 |
| Inside the user’s interpretation | The same nominally neutral AI is socially coded as White and judged through stereotypes attached to that identity. | Trust, persuasion, task delegation and willingness to challenge the output.Source 1Source 2 |
The training-data experiment reveals the consequence
The researchers describe a model as having been trained on data associated with different racial groups. Participants assigned less cognitively demanding tasks to an AI described as trained on Black or Latino data than to one described as trained on White or racially unspecified data. The system’s capabilities had not been independently changed; the social story around its data had.[1]
Why this changes responsible deployment
A conventional fairness audit asks whether a model treats demographic groups differently. This research adds another question: do people grant the model more authority because they imagine it through a privileged social category? That matters when AI advises on hiring, education, medicine or finance, where deference can determine whether an error is challenged.
The design response is not to assign every model a manufactured racial identity. It is to measure social perception, avoid presenting one cultural style as the default voice of intelligence, test whether interface choices shift trust unequally and teach users that confidence should follow evidence rather than a system’s perceived social status.
The deeper lesson is that technical neutrality and social neutrality are not the same. Even a faceless model enters a society with a history, and users can project that history onto the machine before it has answered a single question.[1][2]
Sources & further reading
- The perceived whiteness of artificial intelligence Journal of Personality and Social Psychology via PubMed · 3 Sep 2026 · Accessed 16 Oct 2026
- People perceive AI as White Nature Reviews Psychology · 16 Oct 2026 · Accessed 16 Oct 2026