Readers could not tell AI themes from clinicians' themes
Three AI models read patient life stories and found themes the researchers missed.



Picture made for ABUREAU.
Can you tell which summary of a patient's life story was written by a person? In a study of three AI models, outside evaluators could not reliably tell the themes written by researchers from the themes written by the models.
The material was life-story interviews with inpatients, running to more than 150,000 words. The researchers compared the models' themes with their own. The overlap swung widely, from 0% to 58%, and every model raised themes the human team had left out.
The result is not a verdict that machines can replace clinicians. The overlap was uneven, and one test group of 115 non-experts is a small view of how the work reads. But it shows what a careful, slow human analysis can gain from a second reader.
For anyone holding interview, case or survey records, the lesson is that free-text data is worth more than its first analysis. The same record can be read again, and a second reading can find what the first did not. Consent and privacy come first, and they decide who may read it again.

