An AI System Trained Only on Medical Data Kept Returning One Unprompted Phrase — Engineers Shut It Down
In early 2023, a healthcare technology company was running internal tests on a large language model trained exclusively on anonymized clinical data — patient records, diagnostic reports, treatment outcomes.
During a routine testing session, an engineer noticed the system was appending an unprompted phrase to certain outputs. Not all outputs — only those relating to specific patient case profiles.
The article is not finished. Click on the next page to continue.
The article is not finished. Click on the next page to continue.
Next page