Benchmarking ChatGPT Against Human Expertise in Health Communication Assessment: Organizational Implications of Feasibility-Oriented Evaluation

Prominent narratives of generative artificial intelligence (AI) highlight these tools' cost-saving potential for health organizations via more efficient communication. However, even if a particular tool were to perform adequately at one point in time, evaluating models across updates and upgrades is complex. There are few examples from industry that are feasible enough and few examples from academia robust enough for health organizations. To help fill this gap, we examined the performance of ChatGPT to identify the presence or absence of 6 evidence-based practices from the Centers for Disease Control and Prevention (CDC) Clear Communication Index (CCI) in short-form texts. First, we carried out a prompt engineering process to identify optimal approaches for prompting ChatGPT to correctly label the clear communication practices in a test dataset of 27 tweets, from simulated front-end user and back-end developer perspectives. Using expert labels as comparison points, we found front-end users and back-end developers would reach different conclusions about how to best prompt ChatGPT. Then, we investigated how accurately different versions of ChatGPT's models could label these clear communication practices in a data set of 260 tweets. Using expert labels as comparison points, we found stable performance across two dated GPT-3.5 model versions, albeit with wide confidence intervals. Even so, we found that GPT-4 models could actually perform significantly worse on some CDC CCI labeling tasks. Our results imply the need for increased investment in health communication experts to evaluate generative AI performance over time, which may be less practical than investing in such experts to communicate with patients and engage communities directly.