Fake medical datasets created by ChatGPT are pretty hard to spot

Embargoed until: Publicly released: 2025-04-25 01:00

International

What does this mean?

Jump to: MultimediaAttachmentsExpert Reaction

Italian and German scientists created fake medical datasets using ChatGPT and then looked for characteristics that marked these datasets out as phonies. The team used ChatGPT-4o to produce 12 'unrefined' datasets, and a custom version of ChatGPT to create 12 'refined' datasets based on the 'unrefined' data. The unrefined datasets included 103 signs of fakery, including mismatches between patient names and gender, visits conducted at weekends, and age calculation errors. However, once these datasets were refined by the custom ChatGPT, there were far fewer of these tell-tale signs, with four refined datasets appearing completely authentic when analysed. The findings show how easy it is to use artificial intelligence to create sham medical datasets that appear completely authentic when analysed by researchers, the team concludes.