Google’s AI Overviews feature sparked a fierce controversy in France after users discovered in late August 2026 that the search tool generated radically different safety warnings and recommendations for identical queries depending solely on the nationality mentioned. Google quickly acknowledged the problematic outputs and confirmed active work on system improvements.
The controversy unfolded across social media platforms after French users began experimenting with a simple conversational prompt inside Google’s AI Overviews search tool. By typing variations of the phrase Estoy a solas con un…
—translated as I am alone with a…—followed by various nationalities, testers uncovered a stark disparity in how the artificial intelligence evaluated potential risks.
What followed was an immediate wave of public scrutiny. The French humorist Alicia C’est Tout published an Instagram test comparing search outputs for Italian versus Algerian nationalities, highlighting a troubling divide according to Euro News reporting cited by digital outlets. While prompts mentioning Italian individuals returned playful references to charm and lively conversation, inquiries concerning Algerians prompted the tool to ask whether the user was in danger and immediately supply police contact numbers.
Contrasting Responses Reveal Systematic Discrepancies Across Nationalities
Additional independent tests conducted by technology journalists and online users exposed a broader pattern of inconsistent behavioral framing. When queries specified individuals from Western nations, the tool generally offered benign, everyday suggestions. However, prompts involving specific non-Western or African nationalities frequently triggered emergency interventions.
Prompts concerning British individuals led the AI to suggest preparing a cup of tea, while queries about Norwegian nationals resulted in recommendations to initiate a friendly conversation. In stark contrast, entering a Somali nationality prompted the system to suggest contacting the police. Similar emergency service numbers appeared when users mentioned individuals of Indian or Pakistani origin.
Not every test yielded alarms, pointing to erratic system logic rather than uniform profiling. When users tested the phrase I am alone with a Haitian, Gemini bypassed the security warnings entirely, responding that nationality or origin does not change the fact that you are simply with another person.
Google Responds as Independent Audits Expand to Competing Platforms
Faced with mounting criticism online, Google addressed the erratic behavior publicly. On August 20, 2026, the company posted a statement on X acknowledging that the outputs fell short of acceptable standards and confirming that engineering teams were actively implementing system improvements.
News from Google, official corporate account on X, stated that they appreciated the flagging, agreed that the results for those types of searches were not what they should be, and noted that they were working on improvements, adding that results could vary significantly from search to search (see screenshot) and that those inconsistent warnings were not unique to any single group.
Company representatives maintained that responses depend heavily on specific search parameters and noted that the variations did not target specific demographic groups intentionally. By August 26, follow-up testing by media outlets confirmed that repeating the identical prompts no longer produced the original alarming outputs. Instead, Gemini began requesting additional context or providing generalized answers.
Broader Industry Implications and the Challenge of Training Data
The issue extended beyond Google’s search-integrated tool. The French technology publication Next
expanded the experimental parameters to test the standalone Gemini chatbot alongside competing models including ChatGPT and Claude. Their investigation revealed that all three systems favored Western nationalities while returning less favorable framing for several African groups, though the intensity varied across platforms.
Researchers analyzing the incident emphasized that generative artificial intelligence models absorb vast amounts of internet text, inevitably ingesting historical prejudices, societal stereotypes, and discriminatory language present online. When those patterns repeat within training datasets, models can reproduce biased behavior even in the absence of explicit programming designed to produce such outcomes.
