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

Abstract

Background Large language models have received enormous attention recently with some studies demonstrating their potential clinical value, despite not being trained specifically for this domain. We aimed to investigate whether ChatGPT, a language model optimized for dialogue, can answer frequently asked questions about diabetes. Methods We conducted a closed e-survey among employees of a large Danish diabetes center. The study design was inspired by the Turing test and non-inferiority trials. Our survey included ten questions with two answers each. One of these was written by a human expert, while the other was generated by ChatGPT. Participants had the task to identify the ChatGPT-generated answer. Data was analyzed at the question-level using logistic regression with robust variance estimation with clustering at participant level. In secondary analyses, we investigated the effect of participant characteristics on the outcome. A 55% non-inferiority margin was pre-defined based on precision simulations and had been published as part of the study protocol before data collection began. Findings Among 311 invited individuals, 183 participated in the survey (59% response rate). 64% had heard of ChatGPT before, and 19% had tried it. Overall, participants could identify ChatGPT-generated answers 59.5% (95% CI: 57.0, 62.0) of the time. Among participant characteristics, previous ChatGPT use had the strongest association with the outcome (odds ratio: 1.52 (1.16, 2.00), p=0.003). Previous users answered 67.4% (61.7, 72.7) of the questions correctly, versus non-users 57.6% (54.9, 60.3). Interpretation Participants could distinguish between ChatGPT-generated and human-written answers somewhat better than flipping a fair coin. However, our results suggest a stronger predictive value of linguistic features rather than the actual content. Rigorously planned studies are needed to elucidate the risks and benefits of integrating such technologies in routine clinical practice.

Academics' perceptions of ChatGPT-generated written outputs: A practical application of Turing’s Imitation Game

“…The outcome of the game rests on two interrelated factors: first, the capacity of the machine to produce communicative artefacts that imitate the attributes of those produced by humans. As has been alluded to already, and will be reviewed in more detail below, there is emerging evidence that ChatGPT currently possesses ample capacity to interpret text prompts and produce sophisticated human-like texts (Gao et al, 2023;Hulman et al, 2023;Nov et al, 2023); second, and the focus of the current study, is the degree to which the human interrogator is sensitive to the attributes of communicative artefacts that indicate whether they are produced by a human or a machine.…”

“…Another recent study inspired by Turing's Imitation Game was undertaken by Hulman et al (2023) among 183 employees of a large health provider service in Denmark. The objective of the study was to determine how adequately ChatGPT could answer 10 frequently asked questions that were of relevance to the healthcare service (i.e., questions about diabetes).…”

“…ChatGPT presents a chat interface into which users can enter text prompts and be immediately provided with bespoke text outputs. The application is surprisingly powerful; simple text prompts can be used to produce complex human-like written responses (Gao et al, 2023;Hulman et al, 2023;Nov et al, 2023).…”

Creating Trustworthy LLMs: Dealing with Hallucinations in Healthcare AI

“…FActScores [29] evaluates text generated by LLMs via an evaluation method that breaks a generation into atomic facts which are in turn evaluated by human evaluators. Majority voting for evaluating healthcare related answers generated by LLMs has been employed for myopia care [21], maternity [21], diabetes [26], cancer [27], infant care [21] etc.…”

A survey of consumer health question answering systems

“…A study conducted by Hulman et al. (2023) to evaluate the answers given by ChatGPT for diabetes‐related questions against answers given by humans, reports that ChatGPT's answers to two out of 10 diabetes‐related questions contain misinformation. With respect to this, the quality of the answers provided by more traditional CHeQA approaches as described in this survey, whose answers come from validated scientific content, may be higher than those provided by systems based on generic LLMs or PLMs fine‐tuned on smaller sets of domain‐specific data.…”