Millions of people now turn to AI chatbots for quick answers, and many researchers are doing the same. A new study suggests that when it comes to judging whether psychological tests treat everyone fairly, those confident answers can be seriously misleading.

The research examined whether popular large language models can stand in for the statistical checks behind questionnaires used in schools and psychology. These checks matter because a flawed test could make certain groups of children look less confident when the real problem lies in the measure itself.

Onur Ramazan of the University of Hong Kong and Danielle Ka Lai Lee of Hong Kong Shue Yan University published their findings in the journal Psychology International. They focused on science self-efficacy, a young person’s belief in their ability to tackle scientific tasks and solve problems.

Science self-efficacy is closely linked with success in science, technology, engineering, and maths, yet girls and pupils without English as a first language tend to report lower levels. Before comparing groups, however, researchers must show that a questionnaire works in the same way for everyone, a property psychologists call measurement invariance.

The team analysed responses from 11,323 American teenagers aged 15 and 16 who took part in the international PISA assessments in 2006 and 2015. Pupils rated how well they could handle eight science tasks, such as describing how antibiotics treat disease.

Using a standard technique called multigroup confirmatory factor analysis, the researchers found that the scale worked equally well for boys and girls. It also measured confidence fairly for English speakers and those who speak another language at home, in both years.

The researchers then put the same problem to six widely used chatbots: ChatGPT, DeepSeek, Grok, Microsoft Copilot, Meta AI, and Gemini. Crucially, the free versions of these tools were given none of the underlying data and were asked for realistic estimates from their existing knowledge.

The chatbots produced figures that looked plausible and often resembled one another, but they matched the real results poorly. They tended to paint an overly rosy picture of how well the questionnaire fitted the data, and some wrongly concluded that the test was not fair across groups.

On one common measure of statistical fit, agreement between individual chatbots and the real analysis was limited, and on another it was close to zero. The authors describe this as an illusion of consensus, where machines appear to agree while drifting from the evidence.

The findings arrive as universities grapple with how far generative AI should be trusted in academic work. Earlier studies suggested chatbots can help review test questions, but they appear to struggle with judgements that depend on real data.

The authors stress that traditional statistical analysis should remain the foundation for deciding whether a psychological test is fair. AI tools may still help with jobs such as debugging code, where mistakes are easily checked.

The study has limits, including its focus on one questionnaire and a single prompt for each chatbot. Even so, it offers a timely reminder that a confident answer from a chatbot is not the same thing as evidence.