← Back to Home

AI-written targeted therapy consent forms are more complete, but over a quarter include unsupported inferences

A paired study of 70 rheumatology patients shows that AI can explain treatment risks and alternatives more comprehensively, while also increasing the reading burden. Higher document scores still leave a further validation step before patients’ actual understanding can be established.

By SURL BioNews

Before receiving targeted therapy, patients need to understand not only how much benefit a drug may offer, but also its risks and what other options are available. AI appears capable of covering these points more comprehensively, but a new study found that complete texts may also contain personalized inferences unsupported by medical records, introducing another risk to informed consent.

Published on November 4 in Frontiers in Pharmacology, the study retrospectively examined 70 adult rheumatology patients who received targeted therapy at Beijing Tongren Hospital in 2025. Using de-identified case data and a standardized prompt, the team had DeepSeek-V4-pro generate explanatory texts, then compared them with physician-written versions in electronic medical records. This was a study of routine treatment documents, not a test of participant consent forms for a new drug clinical trial.

Three rheumatologists, blinded to the source of the texts, evaluated a total of 140 texts using a scale developed by the research team that covered eight dimensions and 13 items. The median overall quality score was 60 for the AI versions and 50 for the physician-written versions, with differences mainly in explanations of risks, benefits, and alternatives. A report on the same study republished by Bioengineer.org on November 13 presented the same results; these scores reflect content performance within a specific scoring framework.

However, 19 AI versions, or 27.1%, contained “hallucinations,” while none were found in the physician-written versions. Here, hallucinations mainly involved presenting unsupported inferences as facts about the patients themselves. The study separately examined information inaccuracies and omissions, and neither difference between the groups reached statistical significance. Higher content scores therefore cannot be directly interpreted as greater factual reliability, nor do they establish that AI actually reduced omissions.

The reading threshold likewise did not fall as content increased. The mean score from the Chinese readability tool CRIE 3.0 was 7.87 for the AI versions and 7.10 for the physician-written versions, with higher scores indicating a higher required reading level. This finding suggests that greater detail may bring a greater comprehension burden, but the tool’s scores alone still cannot reveal how much patients actually understood.

This is the study’s key boundary: the AI texts were generated retrospectively for research, were not given to patients to read, and were not used in clinical care. The team evaluated specific explanatory content within consent forms and did not measure patient understanding, decision quality, or treatment outcomes. The single hospital, custom scale, and single model also limit how broadly the findings can be generalized.

The authors therefore positioned AI’s potential role as assisting with drafting and checking content coverage, while requiring physicians to verify each item before clinical use. However, the study has not tested the safety of documents after review or quantified the burden of revision. The next step is to validate whether this process, after errors are corrected, can actually help patients understand their treatment options.

References

  1. Frontiers in Pharmacology
  2. Bioengineer.org (republished from Scienmag)