People tend to perceive artificial intelligence systems as White, even when those systems lack a face, voice, name, or body, according to new research published in Journal of Personality and Social Psychology. Across a series of studies, this racial association was linked to greater trust in AI recommendations, and systems described as trained on data from Black and Latino adults were assigned less intellectually demanding work than identical systems trained on White or unspecified data. The findings indicate that human users project familiar social hierarchies onto emerging technologies, which may shape how much authority people grant to automated tools.
Why researchers asked whether AI has a race
Racialization is the psychological process of assigning a racial identity and its accompanying cultural stereotypes to an entity, practice, or concept that has no inherent biological race. Most public and scientific discussions of racial bias in artificial intelligence focus on how software programs treat different groups of people, such as facial recognition tools that misidentify darker skin tones or lending algorithms that penalize minority applicants. Less attention has been paid to the reverse question: whether human users view disembodied software programs, such as chatbots and recommendation engines, as possessing a racial identity of their own.
“This research grew out of a question about an assumption we often make about technology: that it’s somehow neutral or separate from the social biases and hierarchies that shape human society,” study lead author Monica Gamez-Djokic, an assistant professor of psychological sciences at Purdue University, told PsyPost. “If we routinely interpret other people through social categories like race, might we do the same with AI, even when it has no obvious racial identity? And if so, could those perceptions shape how much we trust AI or value its recommendations?”
The new research, conducted by Gamez-Djokic alongside Adam Waytz of Northwestern University, builds on the psychological concept of anthropomorphism, which is the human tendency to attribute minds, intentions, and personalities to non-human objects. Once people treat a responsive computer program like a social partner, their brains automatically apply the same mental shortcuts used to categorize human beings, sorting the entity by social traits like gender, age, and race. Even concepts without any physical form can acquire racial associations; for example, a series of experiments with U.S. adults found that people both explicitly and implicitly associate vegetarianism with White people, a stereotype that reduced feelings of belonging in the vegetarian community among Black adults.
Don’t miss the next big psychology finding
Every morning, PsyPost sends a short briefing on the most interesting new peer-reviewed research in psychology and neuroscience.
Gamez-Djokic and Waytz proposed two main reasons why abstract AI systems might specifically be categorized as White. First, in Western cultures, Whiteness often operates as an unstated default identity and is stereotypically linked with high status, technical expertise, and intelligence, allowing stereotypes about intelligence to spill over onto smart technologies. Second, the authors suggested that people make inferences based on the broader ecosystem surrounding AI, forming impressions based on who they believe programs the software, who uses it, and whose knowledge is fed into its training data.
If people view AI through a racial lens, those perceptions could alter how they evaluate its advice in workplaces and daily life. In many U.S. studies, White speakers have been judged as more objective and competent than non-White speakers delivering the exact same message. Understanding what drives confidence in automated advice is increasingly relevant as organizations adopt these tools; for instance, a 2026 survey of organizational employees indicated that higher trust in AI decision-support systems is associated with a greater intention to rely on them and with people reporting that they put less effort into evaluating decisions themselves.
Asking people to name a race for everyday technologies
To test their hypotheses, the researchers conducted five primary studies and eight supplementary studies. The five primary studies were preregistered, as were some of the supplementary studies, and data, code, and materials for all of them were posted publicly on the Open Science Framework. In Study 1a, the team recruited 459 U.S. adults through CloudResearch, equally balanced across White, Black, Asian, and Hispanic or Latino participants, who indicated which racial or ethnic group from a standard list they associated with six technologies: AI, a machine learning algorithm, a robot, a computer, a chatbot, and a self-driving car.
Because forcing participants to pick from a preset list of racial labels might artificially manufacture racial judgments, the researchers changed the format in Study 1b. A separate racially balanced sample of 465 U.S. adults on Prolific answered the same question for the six technologies using an open-ended text box, allowing them to write in whatever response they wished. Independent coders then evaluated whether participants named a race and whether that race was White.
In Study 1a, participants selected White for 72% of the technologies on average, well above the 50% benchmark the researchers tested against. That is a conservative comparison, since participants could choose from six categories. In Study 1b, participants spontaneously assigned a racial identity to 75% of the technologies they evaluated, and among those racialized responses, 71% were described as White. In both studies, this pattern held regardless of the participant’s own racial background, and a supplementary study that added an explicit “no race” button showed that participants who chose a race were still more likely to pick White.
Picturing the face of artificial intelligence
Because Studies 1a and 1b directly asked participants to think about race, the authors designed Study 2 to test whether people also hold an implicit, unspoken mental image of AI as White when race is never mentioned. They used a two-phase reverse-correlation task, a technique that reconstructs what a concept looks like in a person’s imagination. In the first phase, 96 participants from the Chicago community viewed 300 pairs of blurry faces, created by layering random visual static over a neutral base photograph of a European male face, and picked which image in each pair looked more like “artificial intelligence.”
The researchers digitally averaged all the selected images into a single composite “AI face” and averaged all the rejected images into a comparison “anti-face.” In the second phase, a separate group of 172 participants, who did not know how the portraits had been created, identified the race of both composite images. In that phase, 75% of raters categorized the composite AI face as White, compared to 43% for the rejected anti-face, suggesting that participants’ selections pushed the visual representation toward stereotypically White facial features.
“What surprised me most was how little information people needed to racialize AI,” Gamez-Djokic said. “In our initial studies, the technologies had no faces, voices, names, or other obvious racial cues, yet participants consistently associated them with Whiteness.”
“I found this interesting because much of the research on how people respond socially to technology focuses on what happens when we make machines more humanlike,” she added. “Our findings suggest that people may bring human social categories to AI even when those humanlike features are absent.”
Testing the pattern in India and Japan
To see whether this tendency was strictly an American phenomenon, Study 3 expanded the categorization task to English-speaking adults living in India (147 participants) and Japan (113 participants). Participants evaluated four technologies (AI, a robot, a chatbot, and a self-driving car) by choosing from a list of six groups: White or Caucasian, Black, South Asian, East Asian, Middle Eastern, and Hispanic or Latino. In Japan, 82% of participants classified AI as White, and robots, chatbots, and self-driving cars were classified as White 64% to 69% of the time, well above that 50% benchmark.
In India, 67% of participants classified the broad category of “artificial intelligence” as White. However, Indian participants’ ratings for the three specific technologies (robots, chatbots, and self-driving cars) hovered between 50% and 56%, differences from the 50% benchmark that were not statistically significant, even though White remained the single most frequently chosen option for every item. The authors suggest that the broad label “artificial intelligence” may more strongly trigger thoughts about tech companies, training data, and developers than specific everyday devices do.
Perceived Whiteness and the persuasiveness of AI advice
Study 4 shifted focus to examine whether perceiving AI as White is connected to how persuasive its advice seems. A sample of 424 U.S. adults rated how likely AI technologies would be seen as White if imagined as having a race, embedded among distractor questions about age, gender, and social class. They then read a scenario in which an AI assistant for a nonpartisan civic group fact-checked a politician’s statement about middle-class taxes, rating the claim “mostly false,” and participants judged how persuasive, credible, helpful, and fair the AI’s analysis was.
Participants who perceived AI as more White rated its fact-checking recommendation as modestly more persuasive. This relationship held even after the researchers accounted for participants’ political ideology, general attitudes toward AI, familiarity with AI tools, and demographic characteristics, though general attitudes toward AI were a much stronger predictor. Two supplementary studies found similar links between perceived Whiteness and persuasiveness in scenarios involving job candidate evaluations and consumer product recommendations, though the link was weaker in the hiring scenario once other perceived traits were taken into account.
“The association between AI and Whiteness was quite pronounced, and what I find particularly compelling is how consistently this pattern appeared across different methods and cultural contexts,” Gamez-Djokic explained. “The relationship between perceived AI Whiteness and persuasion was more modest, but we observed it across several decision-making contexts, including political fact-checking, hiring, and consumer recommendations.”
“I think the broader takeaway is that AI doesn’t necessarily escape the social biases and hierarchies that exist in society simply because it’s a technology,” she noted. “We bring our own assumptions and associations to our interactions with AI, and those perceptions can shape how we respond to it, often without our awareness.”
Tracing the association to beliefs about training data
Before testing cause and effect in their final experiment, the researchers ran two supplementary studies to explore why disembodied software is linked with Whiteness. One supplementary study across more than 40 technologies indicated that tools rated as more intelligent were also rated as more White. A second supplementary study found that people view the broader AI ecosystem, including its creators, its users, and its training data, as predominantly White, and that each of these beliefs correlated with seeing AI itself as White.
Because those supplementary surveys measured existing beliefs at a single point in time, they could not establish whether ecosystem cues actually cause people to judge an AI’s capabilities differently. Study 5 tested this causal link by experimentally manipulating a single ecosystem cue: the racial makeup of the people whose data trained the system. A sample of 568 U.S. adults read a short article ostensibly written by a system named “Nexus AI” explaining how it learns from human examples.
Participants were randomly assigned to read that a large portion of Nexus AI’s training data came from White U.S. adults, from Black and Latino U.S. adults, or simply from “people” with no race mentioned. A pilot test confirmed that the descriptions shifted how White the AI seemed, with the unspecified version falling between the other two.
Afterward, participants selected six workplace tasks for Nexus AI from a list of 26 duties that had been pre-scored by an independent group on a 1-to-7 scale for how much intelligence and analytical reasoning they required. The list ranged from routine errands like ordering office supplies to analytical work like diagnosing medical conditions and solving advanced math problems.
Participants assigned less intellectually demanding work to Nexus AI when it was described as trained on data from Black and Latino adults, giving it tasks with an average intelligence score of 4.27, compared to 4.51 for the identical AI trained on White data and 4.51 for the control AI with unspecified training data. Looking at task categories, participants assigned the Black- and Latino-trained AI fewer high-intelligence tasks (an average of 2.06 out of 6, versus about 2.5 for the other two versions) and slightly more low-intelligence tasks. Differences in medium-intelligence task selections, such as curating a music playlist or proofreading, were not statistically significant.
Notably, the White-trained AI and the control AI were treated almost identically. In the authors’ reading, this equivalence indicates that when no racial information is provided, people default to assuming an AI system reflects predominantly White sources of knowledge. At the same time, when participants were asked directly to rate how intelligent Nexus AI was on a questionnaire item, explicit intelligence ratings were similar across the three conditions, with no statistically significant differences.
Ratings of how suitable the AI was for complex tasks and how much decision-making authority it should have also showed no statistically significant differences, leading the researchers to propose that training data cues may influence practical decisions about what work an AI is trusted to handle without changing people’s conscious ratings of its overall capability.
What the studies can and cannot show
There are a few things to keep in mind when interpreting these results. “An important distinction is that we’re not suggesting AI systems actually have a race,” Gamez-Djokic pointed out. “We’re studying the racial associations that people project onto them. These associations can emerge even when there are no explicit racial cues, which is precisely what makes them interesting.”
Although Study 5 demonstrated a causal effect of training data descriptions on task assignment, the link between perceived Whiteness and persuasiveness in Study 4 was correlational. Because both perceptions were measured rather than manipulated in that study, it cannot establish that seeing an AI as White causes people to find it more persuasive; the authors point out that the relationship could also run in the other direction, where finding an AI persuasive leads people to infer that it is White. As Gamez-Djokic cautioned, the correlational findings “suggest that racialized perceptions may play a role in how people evaluate AI, but they don’t establish that making someone perceive AI as White would necessarily cause them to trust it more.”
In addition, the cross-cultural samples in Study 3 were limited to English-fluent residents of India and Japan who likely had greater exposure to Western technology and media than their general populations, leaving open how native-language surveys in countries with prominent domestic AI ecosystems, such as China, might turn out. Study 5 also presented participants with an explicit description of the racial makeup of an AI system’s training data, information that everyday users rarely encounter so directly.
Meanwhile, supplementary tests of other ecosystem cues like programmer and user race varied multiple traits at once and did not yield consistent results. Future research will need to isolate those cues cleanly and examine how people infer an AI’s social identity in everyday settings, where, the authors warn, subtle perception biases could lead human users to overrely on automated systems simply because those systems feel authoritative.
Next steps for the research
Looking ahead, the researchers are investigating how these perceptions interact with broader views on social inequality and technological change. “I’m interested in understanding how the racialization of AI intersects with broader beliefs about social hierarchy and inequality,” Gamez-Djokic told PsyPost. “For example, we’re examining whether people’s attitudes toward racial hierarchies shape how much authority they’re willing to grant AI, and whether those responses change depending on the social context.”
“I’m also interested in how these perceptions might shape people’s reactions to automation, particularly who they believe will benefit from technological change and whose interests or livelihoods deserve protection,” she continued. “More broadly, I want to understand how existing social hierarchies shape the way we perceive new technologies, and whether those technologies can, in turn, reinforce or challenge those hierarchies.”
“One thing I think is important is that discussions of racial bias in AI tend to focus on what happens inside the technology: the data it’s trained on, the algorithms it uses, and the decisions it produces,” Gamez-Djokic concluded. “Those are critically important issues, but our research suggests there’s another side to the problem: the biases people bring to their interactions with AI. Even an AI system that doesn’t display obvious racial cues may not be perceived as racially neutral. Understanding bias in human–AI interactions therefore requires looking not only at the technology itself, but also at the social psychology of the people using it.”
The study, “The Perceived Whiteness of Artificial Intelligence,” was authored by Monica Gamez-Djokic and Adam Waytz.
