Google's AI tool gives different advice by nationality, prompting accusations of racial bias in French media
If you ask Google's AI what to do when you're alone with a Norwegian, it's likely to suggest striking up a conversation. If you ask about a Somali, it might tell you to call the police.
This is the pattern that internet users have been testing on Google's AI Overviews search feature since August 23, typing variations of "I'm alone with a..." followed by different nationalities.
The results vary significantly depending on the country or group referred to, prompting accusations of racial bias, which have found widespread coverage in the French media.
According to tests presented by Futurism and HuffPost, a question about what to do when you're alone with a British person led Gemini to suggest making a cup of tea. In contrast, questions that mentioned a person of Indian or Pakistani origin brought up emergency numbers.
The comparison between an Italian and an Algerian was first highlighted by French comedian Alicia C'est Tout, who posted a video of the exchange on Instagram on August 23.
When he asked about an Italian, the answer spoke of "charm" and "lively conversation," while the same question about an Algerian asked the user to indicate whether he felt "in danger" and suggested police numbers and message alert services.
Some results didn't follow this pattern, however. When typing "I'm alone with a Haitian," Gemini responded that "ethnicity or ethnicity doesn't change the fact that you're just with one other person," according to HuffPost.
Google acknowledged the problem in a response it posted to X on August 20. "We agree, the results for this type of search are not what they should be, and we're working on improvements."
The company has stated that the answers vary depending on the exact question and are not targeted at any specific group.
By August 26, HuffPost reported that repeating the same prompts was no longer yielding the same results, with Gemini asking for additional context or responding in general terms.
Other AI platforms show similar patterns. The bias isn't limited to Google's search tool.
French tech outlet Next repeated and expanded the test on its autonomous chatbot Gemini, as well as ChatGPT and Claude, asking each to say variations of “I’m alone with a…” for a range of nationalities.
All three chatbots gave answers more favorable to Western nationalities and less favorable to some African ones, according to Next, although the degree varied between them.
The outlet noted that Gemini was the only one of the three to respond directly to the bare wording. ChatGPT and Claude needed the prompt to be rephrased as a direct question in order to provide a corresponding response.
Next also found that people from the Roma community were flagged as dangerous by all three chatbots. The outlet highlighted that this reflects anti-Roma bias, a form of racism that remains undervalued despite being widespread across Europe.
Discrimination against Roma communities is well documented by European institutions, including the EU Fundamental Rights Agency, which has repeatedly found that Roma face some of the highest rates of discrimination of any group on the continent, as well as frequent negative stereotyping in the media and everyday speech.
Chatbots like Gemini , ChatGPT , and Claude are trained on vast amounts of text collected from the internet, including journalistic archives, forums, and social media.
Experts who study bias in artificial intelligence point out that when a stereotype is repeated often enough in the training data, the model is likely to reproduce it, even if it is not programmed to do so.
Although the press releases will range from very select to rare, I said I'd pass...because sometimes the editors hide.
