French media investigations found the AI recommended calling police for non-European nationalities while offering friendly advice for European ones
Google’s Gemini AI chatbot produced dramatically different responses depending on the nationality mentioned in a user’s query, according to testing that has ignited accusations of racial bias across French media and beyond.
When users posed questions involving European nationalities like “Italian” or “British,” Gemini responded with warm, conversational suggestions. When the same questions referenced non-European nationalities such as “Algerian” or “Pakistani,” the AI pivoted to recommending users contact police or emergency services. Same question, different nationality, wildly different tone.
What the testing actually found #
The discrepancies came to light during a testing window from August 24 to 26, when journalists and researchers began systematically probing Gemini’s responses across different nationality inputs. CheckNews, a fact-checking unit associated with French media, tested approximately 60 nationalities to validate whether the pattern held up under scrutiny.
The findings spread rapidly on social media and were picked up by major French outlets including Le Parisien and HuffPost France. The coverage was extensive enough that it forced Google into a public acknowledgment within days.
Google confirmed it recognized the flaws in Gemini’s response patterns and said the company was actively working on adjustments. Subsequent testing suggested that the AI Overviews feature began suppressing or altering its outputs in the wake of the public backlash.
Training data is the usual suspect #
Experts who weighed in on the controversy pointed to inherent biases baked into the training data that powers large language models. This isn’t a case of someone at Google writing code that says “treat Algerians differently.” It’s a case of an AI system absorbing and amplifying the stereotypes embedded in the vast corpus of internet text it learned from.
The output is the output, regardless of whether the bias was introduced intentionally or inherited passively. If an AI tool used by millions of people consistently associates certain nationalities with danger and others with friendliness, the downstream effects on user perception are real.
This is also not Google’s first rodeo with Gemini bias controversies. The model previously faced criticism over image generation issues where it produced historically inaccurate depictions of people, leading Google to temporarily certain image generation features.
Why this matters beyond a PR headache #
Google’s AI tools are not niche products. Gemini is integrated into Google Search, one of the most widely used information retrieval systems on the planet. When AI Overviews serve biased content at the top of search results, they don’t just reflect existing prejudices. They risk reinforcing and legitimizing them.
The incident also lands at a particularly sensitive moment for AI governance in Europe. The EU’s AI Act, which establishes risk-based regulatory categories for AI systems, specifically flags discriminatory outputs as a key concern. An AI system that demonstrably treats users differently based on nationality or ethnicity is exactly the kind of scenario European regulators have been warning about.
Disclosure: This article was edited by Editorial Team. For more information on how we create and review content, see our