Source: French to English Tester Published on: 2026-04-08
Source: The Conversation – France in French (3)– By Rebecca Payne, Clinical Senior Lecturer, Bangor University; University of Oxford
More and more people are consulting artificial intelligence-powered conversational agents to answer everyday questions. These new tools are also used to provide health-related information. However, they are still far from being able to replace doctors, as revealed by a new study.
From now on, to get advice on topics as varied as cooking or how to fill out their tax returns, millions of people turn to conversational agents (chatbots, in English) based on artificial intelligence (AI). An increasing number of individuals also ask them questions about their health. However, just asrecently recalled the chief physician of the United Kingdom, when it comes to making medical decisions, such an approach can prove risky.
In arecent study, my colleagues and I evaluated to what extent chatbots based onlarge language models(LLM) are truly capable of helping the general public deal with certain common health problems. Our striking results demonstrate that the chatbots we tested are not yet able to take on the role of a doctor.
Using a chatbot does not allow for better decision-making in health matters
First of all, let us emphasize that a common objection to research such as that which we have conducted is the assertion that AI evolves faster than the academic publication cycle: by the time an article is published, the models studied have often already been updated, rendering its conclusions obsolete. However, somestudies conducted in the context of medical triage(process aimed at determining at the initial phase of patient care, theprocess tailored to his condition in terms of timing and type of care, editor’s note) and concerning more recent versions of these systems suggest that the same problems persist.
Our work consisted of presenting participants with brief descriptions of common medical situations. The volunteers were randomly divided into two groups. Members of the first group were to interact with one of the three chatbots we had selected (common chatbots, easily accessible to the general public), while those in the other group could use the sources they were accustomed to consulting at home. At the end of the interaction, we asked them two questions: which condition was most compatible with the described symptoms? And, consequently, which healthcare facility was best to turn to?
Chatbot users were found to be less capable of identifying the correct condition than those who had not used one. They were also not better than the control group at determining the appropriate way to seek care. In other words, interacting with a chatbot did not enable participants to make better health decisions.
Solid knowledge, but disappointing results
These results do not mean that the medical knowledge of the tested chatbot models is lacking: LLMs are, indeed, capable ofeasily pass medical certification exams. Moreover, once we set aside the human element by directly submitting the same scenarios to the chatbots, their performance improved noticeably.
Without human intermediary, the models identified the relevant conditions in the vast majority of cases and most often suggested appropriate care options.
Why, then, do the results deteriorate as soon as real users come into play? The analysis of the exchanges highlighted several pitfalls. It often happened that the chatbots mentioned the correct diagnostic hypothesis during the conversation, without the participants retaining it or reproducing it in their final response.
In other cases, users transmitted partial information, or the chatbot misinterpreted essential details. The failure was therefore not due to a simple lack of medical knowledge. It stemmed from a communication problem between the human being and the machine.
Do not confuse theory and practice
This study shows that before deploying new technologies in environments where the stakes are high, policymakers must imperatively have data collected “under real conditions” to assess the true performance of the tools concerned. This is obviously the case in the health sector.
Our results highlight that many of the evaluations currently conducted to determine the interest of AI in medicine have significant shortcomings. Indeed, language models often achieve excellent results when it comes to answering structured exam questions or during simulated interactions between different models.
But the actual use of these tools is much more complex than that. In reality, patients describe their symptoms in a vague or incomplete way. They may also misunderstand the explanations given to them, or ask their questions in an unpredictable order. A system whose results during evaluation tests prove impressive can behave very differently once faced with real users.
This study also highlights a fundamental point about the very nature of clinical practice. As a general practitioner, my work is not limited to mobilizing previously memorized facts. Medicine is often described as an art as much as a science. A consultation’s sole purpose is not just to establish the correct diagnosis: it involves interpreting the patient’s narrative, probing uncertainty, and making decisions that result from negotiation.
This complexity of the unique dialogue between the doctor and his patient has long been recognized by medical educators. For decades, future doctors have been trained according to theCalgary-Cambridge model. This involves establishing arelationship of trust with the patient, to collect information by questioning him and listening to him with the greatest attention, to understand his concerns and expectations, to clearly explain the conclusions to him, and to agree with him on a care plan.
These processes rely on the establishment of a human connection, made possible through tailored communication, involving careful exploration, in order to reach a judgment shaped by context and trust. All these qualities cannot be easily apprehended by pattern recognition techniques (techniques at the base of AI models, which allow the computer to detect, from raw or preprocessed data, the presence of shapes or regularities,editor’s note)
AI chatbots, assistants more than doctors
Our work does not demonstrate that AI has no place in the healthcare sector – far from it. The lesson to be learned is that it is crucial to understand what these systems are currently capable of, and where their limits lie.
Current chatbots should be considered more as assistants than as doctors. They excel in organizing information, summarizing texts, and structuring complex documents. Tasks that are precisely those for which language models are designed.have already proven useful within health systems, whether it is about drafting clinical reports, synthesizing medical records, or generating referral letters, for example.
The promises of AI in medicine remain real, but in the short term, its role will likely be more supportive than a true revolution. We cannot expect chatbots to be the entry point into the healthcare system. They are not yet capable of making diagnoses or directing patients to the appropriate care.
Certainly, AI is already capable of passing medical exams. But just as passing a driving theory test does not make you a competent driver, practicing medicine is not limited to answering questions correctly.
Finding one’s way in the heart of the complexity that hides behind every clinical encounter requires the ability to show empathy and discernment. This remains, for now at least, the prerogative of humans.
Rebecca Payne is working on the REMEDY project, funded by Health and Care Research Wales, and also holds a Clarendon-Reuben scholarship from the University of Oxford. She is a member of the Royal College of General Practitioners and a senior member of the Faculty of Medical Leadership and Management.
–ref. Using AI when you are not a doctor does not help in making a better diagnosis –https://theconversation.com/using-ai-when-you-are-not-a-doctor-does-not-help-in-making-a-better-diagnosis-280196