A white rabbit runs by
ChatGPT, an artificial intelligence (AI) chatbot released in November 2022, has caused controversy within the world of academia1. The number of articles on the subject continues to increase almost exponentially. A general theme is: To what extent can an AI platform write a reliable scientific text?
In our department – Clinical Pharmacology at Odense University Hospital – we decided to approach the question in a practical manner by testing ChatGPT in a routine task. In our Drug Information Centre (DIC) we answer clinical drug-related questions from healthcare professionals.
The foundation of our answers is evidence-based medicine2. We follow a standardized procedure to ensure scientific quality and reproducibility: From a focused clinical question, we perform a thorough literature search and write a mini-review based on the available evidence with an accompanying interpretation of the clinical relevance. The work is reviewed by the department academic staff on a bi-weekly conference before the report is sent to the inquiring health care professional.
We dedicate a significant portion of our work-day to writing and reviewing evidence. The emergence of ChatGPT made us wonder if it could serve as a valid review- or writing assistant – or even replace us. Since our product (the mini-review) is purely text, we can directly compare our answers to ChatGPTs text-based output.
When we asked ChatGPT to write a scientific review on a pharmacological topic, our initial discovery was, that although the answer on surface-level looks convincing, some of the provided references were utter nonsense. In other words: when prompted to write a scientific review, ChatGPT will provide a scientific-looking article including a list of references with an obscure mix of relevant and real articles, real but irrelevant articles, and articles that are complete bogus.
This finding is echoed in the rapidly exploding literature about ChatGPTs capability to produce reliable scientific texts3. Likewise, it is openly stated by the creators of ChatGPT that ChatGPT “sometimes writes plausible-sounding but incorrect answers4.
Intrigued by this, we decided follow the white rabbit…