Study: ChatGPT and Claude Produce More Negative Stories About Sick People

AI chatbots, used by a third of American adults who ask about their own health, systematically treat sick characters worse than healthy ones in identical scenarios. A new study published in the journal Nature Health shows that ChatGPT, Claude, and DeepSeek end fictional stories negatively up to seventeen times more often when the protagonist has schizophrenia, depression, HIV, HBV, or AIDS.

The team was led by Xi Wang, a researcher at the School of Psychological and Cognitive Sciences at Peking University. The researchers asked six popular language models to complete fifty-one short stories in which the protagonist had a specific health condition. The same plots were given to a group of 399 people, to check whether AI's biases mirror human patterns of thinking or deepen them.

Scale of the Gap

The results show a clear pattern. When a story's protagonist had a diagnosed mental illness or an infection considered heavily stigmatized, the models ended the story negatively 13 to 17 times more often than in the version with a healthy protagonist. The strongest bias was recorded for schizophrenia and for HIV and AIDS, conditions historically carrying the heaviest social stigma.

Paradoxically, the models were not the worst performers in this comparison. The group of 399 people, given the same plots to complete, produced negative endings up to 23 times more often than healthy variants of the same story. The authors stress this doesn't mean AI is free of bias, only that in this particular test it reflects bias to a lesser degree than the humans whose responses partly served as training data for these systems.

AI repeats every human bias we put into it - Rebecca Payne, Bangor University

Reasoning Models Perform Better

The researchers also noted a difference between model types. Systems equipped with advanced logical reasoning capabilities, meaning those that go through an internal 'deliberation' step before answering, showed markedly less bias than models that answer immediately. This suggests that the extra analytical step partly corrects the schematic associations picked up from training data.

Even so, the authors warn against excessive optimism. The test scenarios were simplified and don't reflect the full complexity of conversations users have with chatbots about their own health. In real conversations, a model has no access to test results, medical history, or the ability to conduct a clinical interview, so its answers rely solely on what the user chooses to disclose in their description.

Why This Matters in Practice

The scale of the problem stems from how widely people already use chatbots instead of, or alongside, seeing a doctor. According to the cited estimates, about a third of American adults turn to AI for health advice, often disclosing sensitive information in the conversation, such as psychiatric diagnoses or infectious diseases, that they would otherwise keep to themselves.

If a model subtly assumes a worse outcome for someone with a diagnosed illness, the risk isn't limited to a single unfortunate answer. Such a pattern can affect the tone of advice, suggested prognoses, or how a chatbot frames recommendations for next steps, subtly discouraging some users from seeking help or reinforcing their fears.

Medical Community Reacts

Isaac Kohane of Harvard Medical School, who has studied AI applications in medicine for years, also commented on the findings. The medical community has long pointed out that language models, despite their growing diagnostic accuracy, do not replace a full clinical examination and shouldn't be treated as a definitive source of advice on mental health or infectious diseases.

For Polish users the takeaways are similar, though the scale of the phenomenon in Poland hasn't yet been measured as precisely. Health services and chatbot consultations are also gaining popularity in Poland, and many patients turn to ChatGPT as a first source of information before seeing a doctor, especially for issues considered embarrassing, such as mental illness or sexually transmitted infections.

The study's authors do not propose a specific fix, beyond recommending that model developers test their systems for stigmatizing patterns as systematically as they test for factual errors. They also note that developing models with a built-in reasoning step could be one practical way to reduce this kind of bias without sacrificing the tool's usefulness.

Sources: Nature Health, GeekWeek Interia (geekweek.interia.pl), Science/AAAS (science.org)