Are AI chatbots concordant with evidence-based cancer screening recommendations?

Abstract

OBJECTIVE: This study aimed to assess whether information from AI chatbots on benefits and harms of breast and prostate cancer screening were concordant with evidence-based cancer screening recommendations. METHODS: Seven unique prompts (four breast cancer; three prostate cancer) were presented to ChatGPT in March 2024. A total of 60 criteria (30 breast; 30 prostate) were used to assess the concordance of information. Concordance was scored between 0 and 2 against the United States Preventive Services Task Force (USPSTF) breast and prostate cancer screening recommendations independently by international cancer screening experts. RESULTS: 43 of 60 (71.7 %) criteria were completely concordant, 3 (5 %) were moderately concordant and 14 (23.3 %) were not concordant or not present, with most of the non-concordant criteria (9 of 14, 64.3 %) being from prompts for the oldest age groups. ChatGPT hallucinations (i.e., completely made up, non-sensical or irrelevant information) were found in 9 of 60 criteria (15 %). CONCLUSIONS: ChatGPT provided information mostly concordant with USPSTF breast and prostate cancer screening recommendations, however, important gaps exist. These findings provide insights into the role of AI to communicate cancer screening benefits and harms and hold increased relevance for periods of guideline change. PRACTICE IMPLICATIONS: AI generated information on cancer screening should be taken in conjunction with official screening recommendations and/or information from clinicians.

Ähnliche Arbeiten