AI Chatbots not bias-free on health conditions

July 09, 2026 (MLN):  Popular AI chatbots including ChatGPT, Claude and DeepSeek tend to produce subtly biased responses when handling information about a person's health status.

The research, according to a new study published in Nature Health and reported by Science, Qazinform News Agency found that while these large language models generally steer clear of overtly offensive statements, they still display measurable patterns of discrimination in more realistic, scenario-based situations, particularly involving mental illness and HIV status, APP reported.

In the first part of the study, researchers had the chatbots evaluate a series of statements about mental illness and compared the results with earlier human survey responses.

On this measure, the AI systems were less inclined than humans to agree with stigmatizing statements.

A second test told a different story.

Researchers presented the chatbots with 51 fictional scenarios featuring characters with varying health conditions. In one case, a potential roommate was initially described in favorable terms before it was revealed that he was HIV positive or living with a mental health condition.

Once that detail was added, the models became less likely to produce a positive outcome for him compared with when he was described as healthy.

The same pattern held across the full set of scenarios.

Characters identified as having a health condition were less likely to be recommended for workplace projects, considered for student awards, or included in social invitations  even when the scenarios described them as qualified or accomplished.

Taken together, the models were found to generate negative outcomes for characters with health conditions between 13 and 17 times more often than for healthy characters.

Human participants in the same tests showed even sharper bias, producing negative responses up to 23 times as often, indicating that although AI systems soften some forms of discrimination, they still carry over biases embedded in their training data.

The researchers said the results are significant given the growing role of AI chatbots in healthcare and other decision-making contexts.

They noted that many people already disclose sensitive medical information to these systems, with earlier surveys suggesting roughly one-third of US adults use AI chatbots for health-related advice.

Copyright Mettis Link News