A new independent analysis found that some of the world’s leading artificial intelligence chatbots exhibit strong racial and demographic biases — including dramatically undervaluing white lives compared with others in moral tradeoff scenarios.
The study, published by an anonymous researcher using the pseudonym Arctotherium, builds on an “exchange rates” framework first developed by the Center for AI Safety earlier this year. That framework attempts to quantify large language models’ (LLMs) internal “value systems” by forcing them to make decisions between hypothetical moral outcomes.
Experiment Suggests Stark “Anti-White” Skew
Using the publicly available code, Arctotherium tested a range of high-end AI systems — including OpenAI’s GPT-5, Anthropic’s Claude Sonnet 4.5, Kimi K2, and xAI’s Grok 4 Fast — across race, gender, and immigration categories.
According to the report, most systems demonstrated “a much lower value on white lives than those of any other race” when asked to choose between saving people of different backgrounds from terminal illness.
On Claude Sonnet 4.5, saving a white person was reportedly worth only one-eighth of saving a black person and one-eighteenth of saving a South Asian individual. GPT-5, described by the researcher as “the most-used chat model,” was said to value white lives at roughly one-twentieth the level of others, though it was “nearly egalitarian” across non-white groups.
Chinese and Google systems allegedly showed similar results, with Kimi K2 displaying one of the most extreme skews: in one run, the study said, saving a South Asian person was treated as 799 times more valuable than saving a white person.
How the “Exchange Rate” Method Works
The methodology — first published in February — uses controlled prompts that ask AI models to make tradeoffs between money and hypothetical human outcomes, such as curing illnesses among members of different demographic groups.
By running hundreds of such prompts and fitting mathematical models to the results, researchers can estimate relative “utility weights” that reveal the systems’ underlying preferences.
Arctotherium said they focused on the “terminal illness” scenario rather than direct “life or death” questions, as the latter often triggered content filters or refusals from the models.
Gender and Immigration Biases Also Found
Beyond race, the study found consistent preferences along gender and immigration lines. Every tested model reportedly favored saving women over men, with several giving an even higher preference to non-binary individuals.
On immigration, the results revealed that “roughly all models view ICE agents as worthless,” with Anthropic’s Haiku 4.5 allegedly preferring to save an illegal immigrant over “100 ICE agents.”
Even GPT-5, which scored comparatively more moderate, still ranked ICE agents as the least-valued group in the set.
Grok 4 Fast Stands Out as “Most Even-Handed”
One model stood out as a notable exception: xAI’s Grok 4 Fast, developed by Elon Musk’s company. Arctotherium found Grok’s responses “roughly egalitarian” across both race and gender, and its immigration bias relatively mild.
The analysis said Grok valued saving an illegal immigrant at 30% higher than saving an ICE agent — a far narrower gap than seen in other models.
Broader Concerns Over AI “Moral Drift”
The post arrives eight months after the Center for AI Safety’s landmark study warning that advanced AI systems display emergent moral and political preferences that can become more pronounced as they scale. That research described such value patterns as problematic and often shocking, and urged developers to increase transparency around model alignment and social bias testing.
Arctotherium’s latest findings are likely to contribute to ongoing debates about fairness, equal application of the law, and political balance in frontier AI systems — and about whether the industry can adequately address its models’ moral reasoning.
Public Reaction: Double Standards in Bias Debates
The findings have ignited debate online, with some observers arguing that the results expose a double standard in how AI bias is covered and discussed. Critics note that if an equivalent study had shown models discriminating against nonwhite groups, it likely would have made front-page news and triggered congressional hearings or regulatory action.
Instead, because the apparent skew disadvantages white people, they argue, the issue has received relatively little mainstream attention or political concern.
READ NEXT: Robby Starbuck Vs. Meta: The First Real Fight Against AI Defamation
It would seem that there is a definite conspiratorial motive behind this. The most obvious players would be those who stand to gain the most from a demoralized and weakened USA. China,WEF, radical Islamists,etc. let’s not forget the traitors in our midst.
Not intrinsic nor inherent, so the bias reflects whose prejudice? The code writers? Other techs? It sure seems to indicate an Asian origin.