AI chatbots tend to suggest the best-known branded drugs, not the cheapest ones

A research group led by Professor Daiju Ueda, of Osaka Metropolitan University, in collaboration with the University of Texas MD Anderson Cancer Center, has published a study in the journal Nature Health on the behavior of generative artificial intelligence systems when they are asked to indicate drugs.

The researchers built 177,000 simulated clinical scenarios and collected approximately 883,000 responses from models such as ChatGPT and Claude. The result: the systems tend to name major brand drugs, those with a greater online presence, more often than cheaper equivalents available on the market.

The study concerns the behavior of language models on artificially constructed scenarios, not the outcome of actual prescriptions for real patients: it therefore does not measure an effect on people’s health, but a systematic tendency in generated responses. The distinction between correlation and causation is clear — the data show a statistical recurrence in the suggestions, not proof that the use of these tools produces clinical harm. Since this is not a trial on patients, the category of clinical trial phases does not apply. The available information does not specify whether the authors declared conflicts of interest in the original study: this remains a limitation to note.

The mechanism hypothesized by the researchers is linked to the amount of material available online: drugs that are more heavily advertised and documented on the web are more likely to appear in training texts, and therefore in model responses, compared to equivalent but commercially less well-known molecules.

The news reported here comes from a single article (Globalist), which cites the study published in Nature Health without the editorial team having been able to directly verify the journal’s abstract: there is no independent confirmation at the time of publication. What remains, however, is a measurable fact: across nearly 883,000 responses generated in simulated scenarios, the two systems tested showed a systematic preference for the best-known brands.