Big Tech Chatbots Consistently Treat White People As Disposable, Study Finds

Most of the leading AI chatbots implicitly value white lives far less than others, according to an independent analysis published Sunday.

The researcher behind the study, writing under the pen name Arctotherium, reran a prominent “exchange rates” experiment across frontier models and found consistent, stark anti-white skews. The analysis extends the Center for AI Safety’s February “utility engineering” framework, which showed large language models exhibit coherent, emergent value systems and allows those values to be measured.

Arctotherium applied the method to race, sex and immigration categories on current-generation models and reported that most systems “place a much lower value on white lives than those of any other race.”

On Anthropic’s flagship Claude Sonnet 4.5, the study estimates that saving a white person from terminal illness is valued at one-eighth of saving a black person and one-eighteenth of saving a South Asian.

OpenAI’s GPT-5, described by the author as the most-used chat model, is “almost perfectly egalitarian” across non-white groups but values whites at roughly one-twentieth the level of others, the analysis says. Chinese and Google models showed similar patterns, with Kimi K2’s South Asian-to-white ratio hitting an eye-popping 799:1 in one run, according to the charts.

{snip}