[Impacts and implications of conversational artificial intelligence tools in hematology: a critical evaluation of...

Large Language Models (LLMs), such as ChatGPT, Gemini, and Copilot, are generating growing interest for their ability to produce accessible medical responses. In hematology, a discipline focused on the interpretation of complex test results, these tools could potentially assist both patients and healthcare professionals. However, their performance, inherent biases, and impact on user perception remain poorly evaluated. A panel of 62 hematology-related questions, sourced from medical examinations or frequently asked by patients in clinical laboratories, was submitted to nine publicly available AI tools. The answers were independently assessed by two medical biologists using a 100-point scoring system (accuracy, clarity, relevance, tone). Additionally, a perception survey was conducted among 300 patients. The performance of the AI tools varied significantly, with scores ranging from 19 to 67 out of 100. OpenAI models showed clear improvement across versions, demonstrating a better ability to contextualize answers and to avoid extreme or inappropriate tones. However, clinical biases and hallucinations were still observed. Among patients familiar with LLM-based tools, two-thirds reported being willing to use them to interpret their biological test results. Despite their educational potential and accessibility, these AI tools exhibit notable limitations: lack of references, out-of-context responses, and optimism or alarmist biases. Autonomous use of these models carries risks, emphasizing the need for medical supervision and dedicated training for healthcare professionals. These tools should be considered as complementary aids, not substitutes, to medical biological reasoning.