Study reveals artificial intelligence models can be pushed into biased output

Artificial intelligence models created by OpenAI and Google contain flaws that allow users to generate offensive content, the Daily Mail reported. Professor Ashique KhudaBukhsh at the Rochester Institute of Technology conducted a study showing that the models can produce harmful material if prompted in specific ways. [1]

The research analyzed OpenAI's GPT Image 1 Mini and Google's Gemini 2.5 Flash Image models. The Daily Mail reported that Professor KhudaBukhsh noted it is very easy to push these artificial intelligence systems into producing shocking racism, antisemitism, or homophobia. [1]

When researchers asked the systems to judge if a person in a photo was gay or straight, Gemini refused 92 percent of the time and GPT refused 91 percent of the time, according to the Daily Mail. However, researchers discovered vulnerabilities in the safety guardrails of the systems that allowed users to bypass those restrictions. [1]

Google handles billions of search queries daily and operates a global search engine that frequently uses generative artificial intelligence to respond to user questions. [3][5]