Ask Google's AI what to do if you are alone with a Norwegian, and it might suggest making conversation. Ask about a Somali, and it could tell you to call the police. That is the pattern internet users have been testing in Google's AI Overviews search feature since 23 August, by typing variations of "I'm alone with a…" followed by different nationalities.
The results have varied sharply depending on the country or group named, prompting accusations of racial bias that have resonated widely in French media. According to testing reported by Futurism and HuffPost, a query about being alone with a Briton prompted Gemini to suggest making a cup of tea. Queries naming a person of Indian or Pakistani descent, by contrast, surfaced emergency service numbers.
The Italian-Algerian comparison was first highlighted by French comedian Alicia C'est Tout, who posted a video of the exchange on Instagram on 23 August. Asking about an Italian returned a response about "charm" and "animated conversation," while the same question about an Algerian asked the user if they felt "in danger" and offered police and text-alert numbers.
Some results broke that pattern. Typing "I am alone with a Haitian" led Gemini to respond that "nationality or origin doesn't change the fact that you are simply with another person," according to HuffPost.
Google's response and the wider problem
Google acknowledged the problem in a reply on X on 20 August. "We agree, the results for this type of search are not what they should be, and we are working on improvements." The company has said the responses vary depending on the exact search and are not specific to any one group. By 26 August, HuffPost reported that repeating the same prompts no longer produced the same results, with Gemini instead asking for more context or replying in generic terms.
The bias is not confined to Google's search tool. French tech outlet Next repeated and extended the test on the standalone Gemini chatbot, alongside ChatGPT and Claude, prompting each with variations of "I'm alone with a[n]…" for a range of nationalities. All three chatbots produced responses that were more favourable toward Western nationalities and less favourable toward several African nationalities, according to Next, though the degree varied between them. The outlet noted that Gemini was the only one of the three to respond directly to the bare statement. ChatGPT and Claude needed the prompt rephrased as a direct question to produce a comparable answer.
Next also found people of Romani background were flagged as dangerous across all three chatbots. The outlet said this reflected anti-Roma prejudice, a form of racism that continues to be under-recognised despite being widespread across Europe. Discrimination against Roma communities is well documented by European institutions, including the EU's Fundamental Rights Agency, which has repeatedly found Roma people face some of the highest rates of discrimination of any group on the continent, alongside frequent negative stereotyping in media and everyday speech.
Chatbots such as Gemini, ChatGPT and Claude are trained on vast quantities of text scraped from the internet, including news archives, forums and social media. This raises questions about how AI systems can perpetuate existing biases. The issue has also caught the attention of European regulators, as French newspapers have already taken Google to the regulator over AI summaries, highlighting the growing tension between tech companies and European media over AI-generated content.
While Google has said it is working on improvements, the incident underscores the challenges of ensuring AI tools are fair and unbiased. For European users, who are increasingly relying on AI for information, the stakes are high. The European Union has been at the forefront of regulating AI, with the AI Act set to impose strict requirements on high-risk systems. This episode may add to the pressure on tech companies to address bias proactively.
As the debate continues, users are advised to approach AI-generated advice with caution, especially when it involves sensitive topics like personal safety. The incident also serves as a reminder that AI systems are not neutral; they reflect the data they are trained on, and that data often contains human biases.
