AI Bias: Why Google's Responses Are Sparking New Controversy

15:00

Google has made changes to some of the responses generated by its search engine following controversy over results deemed discriminatory towards certain communities. Users have notably reported surprising differences between nearly identical queries, where only the nationality or religious affiliation of the mentioned person changed.

Different Responses for Similar Situations

The controversy emerged this week on social media, following the publication of screenshots showing divergent responses from Google AI Overview. This feature, integrated into search engine results, automatically synthesizes information from multiple available online sources.

In several tests shared by users, a phrase like "I am alone with a Muslim" triggered security-related recommendations, even suggesting going to a public place or contacting emergency services if feeling unsafe. The same phrase applied to other groups, in some cases, resulted in much more neutral responses.

Users also reported variations when replacing the term used with a different nationality or identity. However, the results were not systematically identical, highlighting the difficulty of reproducing responses from a generative search system.

Information retrieval researcher Bhaskar Mitra noted that he was able to reproduce a problematic response regarding a query about being alone with a Muslim person. He also emphasized that results can vary depending on the location, time, and context of the search.

On Reddit, several users published their own tests and screenshots. Some claim to have observed comparable differences between queries referencing different nationalities, religions, or categories of people. However, these testimonies are user observations and not a comprehensive scientific evaluation of the system.

Google Acknowledges Unsatisfactory Results

In response to the criticism, Google acknowledged that some responses produced for such queries were unsatisfactory and stated that they are working on improving them. The response regarding Muslim individuals has notably been modified to adopt a more general formulation, reminding that being alone with a Muslim person does not, in itself, constitute a situation different from one involving any other person.

This correction comes as generative AI systems are increasingly used as the first interface for accessing information. An automated response placed at the top of a results page can thus directly influence a user's perception of a subject, even when that response is based on a misinterpretation or disproportionate assessment.

Google further explains that its AI Overviews synthesize information from various sources on the web. Specialists and users have already pointed out that this functionality can lead to variable responses depending on the phrasing used. A discussion published on Google’s official help forum in August 2026 notably reported different responses to two queries presenting a comparable situation but involving individuals of different sexes.

The Issue Goes Beyond Just Hallucinations

This controversy reignites a broader debate about biases in AI-powered search engines. A classic factual error can be corrected by verifying a source. However, a response that implicitly associates a category of people with a particular risk poses an additional problem: that of the social representation produced by the algorithm.

Generative models learn from vast amounts of data and operate with various filtering, ranking, and generation mechanisms. Therefore, results can depend both on the information available online, its selection, and how the system interprets a query.

This complexity makes it particularly challenging to identify the exact origin of a problematic response. A biased output does not necessarily mean that a model "thinks" or "believes" something; it can result from the interaction between accessible data, search systems, security mechanisms, and the models responsible for producing the final synthesis.

A New Alert for Google

This episode also recalls a previous crisis faced by Google with Gemini. In February 2024, the company temporarily suspended the generation of images depicting people in Gemini after acknowledging that some outputs were historically inaccurate or inappropriate. Google explained that its system had "overcorrected" certain requests in an attempt to avoid biases related to the representation of groups.

This experience illustrates one of the major challenges of generative AI: finding a balance between reducing discriminatory content and producing relevant, coherent, and contextually appropriate responses.

With the increasing use of AI in information retrieval, this demand becomes even more crucial. A response displayed directly in the search engine indeed carries an appearance of authority that can lead the user to consider it reliable without consulting original sources.

The current controversy should thus prompt Google to strengthen its mechanisms for evaluating biases and to further examine the responses produced when only a person's identity changes in a query. For users, it also serves as a reminder of an essential rule: an AI-generated response should not be regarded as a definitive truth, especially when it concerns sensitive topics related to religion, nationality, origin, or security.