Putting ChatGPT’s Medical Advice to the (Turing) Test

Abstract

Importance: Chatbots could play a role in answering patient questions, but patients' ability to distinguish between provider and chatbot responses, and patients' trust in chatbots' functions are not well established. Objective: To assess the feasibility of using ChatGPT or a similar AI-based chatbot for patient-provider communication. Design: Survey in January 2023 Setting: Survey Participants: A US representative sample of 400 study participants aged 18 and above was recruited on Prolific, a crowdsourcing platform for academic studies. 384 participants filled out the full survey. After removing participants who spent less than 3 minutes on the survey, 360 respondents remained. 53.3% of respondents analyzed were women; their average age was 45.4. Exposure(s): Ten representative non-administrative patient-provider interactions were extracted from the EHR. Patients' questions were placed in ChatGPT with a request for the chatbot to respond using approximately the same word count as the human provider's response. In the survey, each patient's question was followed by a provider- or ChatGPT-generated response. Participants were informed that five responses were provider-generated and five were chatbot-generated. Participants were asked, and incentivized financially, to correctly identify the response source. Participants were also asked about their trust in chatbots' functions in patient-provider communication, using a Likert scale of 1-5. Main Outcome(s) and Measure(s): Main outcome: Proportion of responses correctly classified as provider- vs chatbot-generated. Secondary outcomes: Average and standard deviation of responses to trust questions. Results: The correct classification of responses ranged between 23.4% to 86.7% for different questions. On average, chatbot responses were identified correctly 60.0% of the time and provider responses were identified correctly 62.3% of the time. On average, responses toward patients' trust in chatbots' functions were weakly positive (mean Likert score: 3.4), with lower trust as the health-related complexity of the task in questions increased. Conclusions and Relevance: ChatGPT responses to patient questions were close to indistinguishable from provider responses. Laypeople appear to trust the use of chatbots to answer lower risk health questions. It is important to continue studying patient-chatbot interaction as chatbots move from administrative to more clinical roles in healthcare.

Trialling a Large Language Model (ChatGPT) in General Practice With the Applied Knowledge Test: Observational Study Demonstrating Opportunities and Limitations in Primary Care

“…However, the accuracy of computational models’ answers to medical questions is yet to exceed that of fully trained physicians, with findings in the present context of primary care being no exception [ 16 , 17 ]. When ChatGPT is used as a medical advice chatbot, advice seekers are only able to identify that the source of provided advice is computational 65% of the time [ 19 ]. It follows that health care providers must protect their patients from inaccurate information provided by this technology, as they are unable to differentiate between computational and human advice [ 19 ].…”

“…When ChatGPT is used as a medical advice chatbot, advice seekers are only able to identify that the source of provided advice is computational 65% of the time [ 19 ]. It follows that health care providers must protect their patients from inaccurate information provided by this technology, as they are unable to differentiate between computational and human advice [ 19 ]. This requirement for oversight limits the potential of LLMs to meaningfully change practice, as performance equivalent to that of experts is the minimum standard to justify autonomous deployment: there must be confidence in the accuracy and trustworthiness of answers from these applications [ 20 , 21 ].…”

ChatGPT- versus human-generated answers to frequently asked questions about diabetes: a Turing test-inspired survey among employees of a Danish diabetes center

“…We identified only one study with a similar design to ours. 18 The authors extracted patient-provider communication from electronic health records and presented five cases with provider-written answers and another five with ChatGPT-generated answers. Participants, who were recruited online, could identify 65% of AI-generated answers correctly, which is of similar magnitude to our findings, although the authors neither have formally tested any hypotheses, nor presented precision estimates for the results.…”

Does ChatGPT Provide Appropriate and Equitable Medical Advice?: A Vignette-Based, Clinical Evaluation Across Care Contexts

“…2 Some preliminary work in the medical domain highlighted ChatGPT's ability to write realistic scientific abstracts, 3 pass medical licensing exams, 4 and accurately determine appropriate radiology studies. 5 Although ChatGPT can triage medical cases, 6 answer clinical questions consistent with the judgment of practicing physicians, 7 and provide medical advice that is perceived as human-like by non-clinicians, 8 its ability to provide appropriate and equitable advice to patients across a range of clinical contexts remain unknown. These knowledge gaps are important because the underlying training data and approach for ChatGPT have not been released, 9 and there are substantive concerns about the safety, fairness, and regulation of LLMs and clinical AI systems.…”