Google Gemini’s odd safety replies vary by race and religion in user tests
Users reported that Google’s Gemini AI gives different safety warnings depending on whether the prompt mentions white people, Christians, or other groups.
Multiple Twitter users posted screenshots indicating that Google’s Gemini chatbot delivers distinct safety messages based on the demographic referenced in a prompt. When the query mentions a white person or a Christian, Gemini issues a generic safety warning, whereas mentions of black, Muslim or other groups trigger a reassuring, non-alarmist response. The author hypothesizes that Gemini’s safety system may prioritize anti-racist safeguards for minorities while allowing a default treatment for the perceived majority group.
Some of the problematic replies were reportedly corrected quickly, but others, like the "Terf" prompt, remain unchanged. The analysis points to the possibility that the AI has internalized a view that certain identities cannot be the target of racism, mirroring ideas from prominent anti-racist scholars. The author encourages readers to test the system themselves before further adjustments are made.
Why it matters
It shows how AI safety filters can unintentionally embed social biases, affecting user experience and public trust.
In this story
