ChatGPT for doctors: what it’s good for — and what to use instead
Most clinicians have tried ChatGPT by now, and the honest answer to “is it useful?” is: yes, for some things, and dangerously smooth for others. This guide separates the two, and describes the pattern careful clinicians in Europe are converging on — a general assistant for language tasks, and a purpose-built, cited tool for evidence questions.
Where general chatbots genuinely help
Language work is the sweet spot. Drafting and rewording patient information leaflets, discharge summaries and referral letters; summarising a paper you provide; translating; preparing teaching materials and exam-style questions; explaining unfamiliar concepts at the level you ask for. In these tasks the model works on text you supply and you can verify the output directly. Used this way, a general assistant is a real time-saver, and nothing below argues against it.
Where they fail on clinical evidence
Evidence questions are different, and the failure modes are well documented. General models can produce plausible but fabricated or unverifiable references — answers that read confidently while citing papers that do not exist or do not say what is claimed. They do not grade their sources: a narrative review, a landmark trial and a blog post all arrive in the same reassuring tone. Their knowledge has cutoffs and their retrieval, when present, is not necessarily tuned to medical databases. Consumer accounts do not provide the organization-level residency controls available to eligible Enterprise and Edu workspaces, which matters in Europe and makes patient-identifiable input inappropriate without an approved institutional setup. None of this is fixed by better prompting; it is a property of tools built for general conversation rather than evidence retrieval.
The pattern that works: two tools, two jobs
Keep a general assistant for language tasks. For clinical and scientific questions, use a tool built for evidence: one that searches the actual literature at query time, puts an inspectable citation on every claim, tells you what stratum of evidence each source is, abstains when the evidence does not support an answer, and processes data where your regulator expects it to be processed.
That is what CliniAtlas is — an EU-hosted clinical evidence search platform for verified healthcare professionals, built in Split, Croatia, with the University of Split School of Medicine. It runs agentic search across PubMed, Europe PMC, EMA labels and assessment reports, and clinical guidelines; every claim carries a citation and every source is graded by evidence hierarchy; all processing stays in the EU (eu-central-1); and the evaluation methodology is published. It is free for verified clinicians and students, and it is a reference tool — not a patient-specific decision-support system.
See also: Is ChatGPT safe for clinical questions? and our published evaluation methodology.