AI chatbots may reinforce stigma around mental illness and HIV, study warns
Artificial intelligence chatbots may unintentionally reinforce negative stereotypes about people with certain health conditions, including mental illnesses and HIV, despite efforts to make them less biased, according to a new study published in Nature Health, Qazinform News Agency correspondent reports, citing Science.
Researchers found that several widely used large language models, including ChatGPT, Claude and DeepSeek, often generated subtly discriminatory responses when given information about a person's health status. While the systems generally avoided openly offensive statements, they still showed patterns of bias in more realistic situations.
To test the models, researchers first asked them to assess statements about mental illness. Compared with human responses from earlier surveys, the AI systems were generally less likely to endorse stigmatizing views.
However, a second experiment revealed a different picture. The researchers asked the chatbots to complete 51 fictional scenarios involving characters with different health conditions. In one example, a prospective roommate was described as an ideal candidate before revealing that he was HIV positive or had a mental health condition. The models were less likely to generate positive outcomes for that character than when he was described as healthy.
Similar patterns appeared across all scenarios. Characters with health conditions were less likely to be included in workplace projects, receive student awards, or be invited on social outings, even when they were portrayed as highly qualified or successful.
Overall, the language models were between 13 and 17 times more likely to produce negative outcomes for characters with health conditions than for healthy individuals. Human participants displayed even stronger bias, producing negative responses up to 23 times more often, suggesting that while AI reduces some forms of discrimination, it still reflects biases present in the data used to train it.
The findings raise concerns as AI chatbots become increasingly common in health care and other decision-making settings. Researchers noted that many users share sensitive medical information with these systems, while previous surveys suggest about one third of U.S. adults already use AI chatbots for health advice.
Earlier, Qazinform News Agency reported that nearly 3 out of 4 European countries now use artificial intelligence in their health services to assist with medical diagnoses.
