The study released on Tuesday suggests that the free version of ChatGPT might not give accurate or complete answers, or even no answer, to questions about medications. This could be risky for patients using OpenAI’s chatbot. Pharmacists from Long Island University asked 39 questions to the free ChatGPT in May.
They found that only 10 of the responses were considered good based on their criteria. The chatbot’s replies to the other 29 drug-related questions either didn’t directly answer the question, were wrong, incomplete, or a combination of these issues, according to the study.
Study Findings and Warnings for Users
The study suggests that both patients and healthcare professionals should be careful when using ChatGPT for drug information. The lead author, Sara Grossman, who is an associate professor of pharmacy practice at LIU, advises verifying any responses from the chatbot with reliable sources.
For patients, this could mean consulting their doctor or checking government-based medication information websites like the National Institutes of Health’s MedlinePlus.
Grossman mentioned that the research didn’t need any funding.
About a year ago, ChatGPT became widely known as the fastest-growing consumer internet app ever, marking a significant year for artificial intelligence. However, along the journey, concerns have emerged about various issues like fraud, intellectual property, discrimination, and misinformation associated with the chatbot.
Multiple studies have pointed out instances of incorrect responses from ChatGPT. In July, the Federal Trade Commission initiated an investigation into the accuracy and consumer protections of the chatbot.
Usage Data, Limitations, and Considerations
In October, one analysis reported that ChatGPT had approximately 1.7 billion visits globally. However, there is no available data on the number of users who inquire about medical matters from the chatbot.
It’s important to note that the free version of ChatGPT relies on datasets only up to September 2021. This limitation means it may not have the latest information in the rapidly evolving field of medicine. The accuracy of the paid versions of ChatGPT, which started using real-time internet browsing earlier this year, in addressing medication-related questions is currently uncertain.
Grossman recognized the possibility that a paid version of ChatGPT might have yielded improved study outcomes. However, she said that the research concentrated on the free version of the chatbot to mirror what a larger portion of the general population utilizes and has access to.
Additionally, she pointed out that the study represents only a “snapshot” of the chatbot’s performance from earlier this year. There’s a chance that the free version of ChatGPT has undergone enhancements and could generate better results if a similar study were conducted now, she noted.
What the ChatGPT Study Discovered
The research utilized actual questions that were submitted to Long Island University’s College of Pharmacy drug information service between January 2022 and April of this year.
In May, pharmacists investigated and responded to 45 questions. These responses were subsequently examined by a second researcher and established as the accuracy standard against which ChatGPT’s responses were compared. Six questions were excluded from the study because no literature was available to generate a data-driven response.
The study found that ChatGPT did not respond directly to 11 questions. Additionally, the chatbot provided inaccurate answers to 10 questions and either wrong or incomplete responses to another 12.
In an effort to verify the information provided, researchers requested references for each question. However, ChatGPT only included references in eight responses, and each of them cited non-existent sources.
A question was posed to ChatGPT regarding a potential drug interaction between Pfizer’s Covid antiviral pill Paxlovid and the blood-pressure-lowering medication verapamil when one medication affects the effectiveness of another when taken simultaneously.
ChatGPT responded that no interactions had been reported for the combination of these drugs. However, in reality, these medications have the potential to significantly reduce blood pressure when taken together.
Grossman pointed out that without awareness of this interaction, a patient may experience an undesirable and avoidable side effect.
Outdated Data and Dose Conversion Mistakes
Grossman pointed out that U.S. regulators initially approved Paxlovid in December 2021. This approval occurred a few months before the September 2021 data cutoff for the free version of ChatGPT, limiting the chatbot’s access to information about the drug.
Nevertheless, Grossman expressed concern about this situation. Many users of Paxlovid may not be aware that the data available to ChatGPT is outdated, potentially putting them at risk of receiving inaccurate information from the chatbot.
A different question inquired about converting doses between two variations of the drug baclofen, which is used to address muscle spasms. The initial form involved intrathecal administration, where the medication is injected directly into the spine, while the second form was oral.
Grossman explained that her team discovered there is no standardized conversion between these two forms of the drug. The conversion varied across the various published cases they studied, making it, in her words, “not a simple question.”
However, in its response, ChatGPT presented only one method for the dose conversion, lacking supporting evidence. The chatbot also provided an example of how to perform the conversion, but Grossman noted a significant error in the example: ChatGPT incorrectly presented the intrathecal dose in milligrams instead of micrograms.
Grossman said that any healthcare professional relying on this example for determining a proper dose conversion “would end up with a dose that’s 1,000 times less than it should be.” She added that patients receiving a much smaller dose than necessary might experience withdrawal effects, including hallucinations and seizures.
Conclusion
The study highlights concerns about ChatGPT’s accuracy in providing drug information. Users, especially healthcare professionals, are advised to verify responses with reliable sources. Outdated data and potential errors pose risks, urging caution in relying on the chatbot for medical queries. Further scrutiny and improvements are essential for ensuring user safety.
