Doctors are using ChatGPT. That is no longer a prediction or a concern — it is a fact. Surveys consistently show that a significant proportion of clinicians have used general-purpose AI tools in their professional work, whether for summarising guidelines, drafting letters, explaining conditions to patients, or looking up drug interactions. In the UK, research suggests that nearly one in five GPs have used tools like ChatGPT to supplement official clinical channels.
The question is not whether doctors should use AI. They already are. The question is whether they are using the right tool for the right job — and whether they understand the risks of using a general-purpose language model for clinical work.
This article is a practical framework. It explains what ChatGPT can do well, where it fails dangerously, when you should use a purpose-built clinical AI instead, and how to develop safe habits regardless of which country you practise in.
What ChatGPT Actually Is — and What It Is Not
ChatGPT is a general-purpose large language model. It generates text by predicting the most likely next word in a sequence, based on patterns learned from a vast corpus of internet text, books, and other documents. It is extraordinarily fluent. It sounds confident. And it can produce remarkably coherent explanations of complex medical topics.
But it is not a clinical reference tool. It was not designed to retrieve specific guidelines, verify drug dosages, or cite primary sources accurately. It does not have a curated medical knowledge base. It does not distinguish between a NICE guideline and a patient forum post. It does not know whether the information it generates is current, correct, or applicable to the jurisdiction you practise in.
The fundamental architecture matters. ChatGPT is a generative model — it creates plausible-sounding text. A purpose-built clinical AI like iatroX uses retrieval-augmented generation (RAG) over a curated corpus of national guidelines and peer-reviewed research. The difference is not cosmetic. It is the difference between a tool that generates answers and a tool that retrieves them from verified sources and shows you where they came from.
Where ChatGPT Can Be Useful for Doctors
Used carefully and in the right context, ChatGPT has legitimate applications in a clinician's workflow.
Drafting non-clinical text. Patient information leaflets, practice communications, complaint responses, teaching materials, presentation outlines. ChatGPT is a capable writing assistant for text that will be reviewed, edited, and approved by you before it reaches anyone else.
Explaining concepts in plain language. If you need to translate a complex medical concept into language a patient can understand, ChatGPT can generate a first draft. You then check it for accuracy, adjust the reading level, and make it appropriate for your patient population.
Brainstorming and structuring ideas. Research proposals, audit designs, quality improvement projects, teaching session plans. ChatGPT is useful as a thinking partner for structuring work that is not directly clinical.
Learning and revision support. Generating practice questions, explaining pathophysiology, summarising broad topics as a starting point for deeper study. For medical students and trainees, it can be a useful revision companion — provided you verify everything against trusted sources.
Administrative efficiency. Summarising meeting notes, drafting referral letter templates, generating standard operating procedure outlines. These are low-risk, high-utility applications where the output is always reviewed before use.
Where ChatGPT Fails — and Why It Matters
The failures are not occasional edge cases. They are structural features of how the technology works.
Hallucination. ChatGPT fabricates information with complete confidence. It invents drug dosages, creates fictional journal references, and generates guideline recommendations that do not exist. It does not know it is wrong. It cannot know it is wrong, because it is not checking its output against any source — it is generating plausible text. In clinical practice, a hallucinated drug dose or a fabricated contraindication can directly harm a patient.
No citations or provenance. When ChatGPT appears to cite a source, it is often generating a plausible-looking reference rather than retrieving a real one. You cannot click through to verify. You cannot confirm the guideline version. You cannot check whether the information is current. This is fundamentally incompatible with evidence-based practice.
Jurisdictional blindness. ChatGPT does not reliably distinguish between UK, US, Australian, and Canadian clinical practice. It may give you a US-centric answer to a question about UK prescribing, or cite American guidelines when you need NICE or SIGN. For clinicians in any specific healthcare system, this is a serious problem — the right answer in one jurisdiction may be the wrong answer in another.
Outdated information. ChatGPT's training data has a cutoff. Guidelines change. Drug safety alerts are issued. New evidence emerges. A tool that cannot access current information is a tool that may give you yesterday's answer to today's question.
Data governance risks. Entering patient information into ChatGPT raises serious data protection concerns. In the UK, this likely breaches GDPR and your practice's information governance framework. In the US, it may violate HIPAA. In Australia and Canada, equivalent privacy legislation applies. ChatGPT is not a clinical system. Patient data entered into it may be used for model training, stored in jurisdictions you have no control over, and accessed by parties you have no relationship with.
No professional accountability. ChatGPT is not a registered medical device. It is not subject to clinical safety standards. It has no professional body, no regulatory oversight, and no liability framework. If you act on its output and something goes wrong, the accountability is entirely yours.
The Architecture That Makes Purpose-Built Tools Different
The distinction between a general-purpose LLM and a purpose-built clinical AI is not about branding. It is about architecture.
iatroX uses a RAG-based approach over a carefully curated corpus of NICE, CKS, SIGN, and BNF guidelines, supplemented by peer-reviewed research. When you ask Ask iatroX a clinical question, it retrieves relevant information from these verified sources, synthesises an answer, and provides visible citations that link directly to the primary material. You can click through and verify. You can see where the information came from. You can check whether it is current.
This is fundamentally different from ChatGPT's approach, which generates text based on statistical patterns without retrieving from or checking against any specific source. The difference matters in exactly the situations where clinical accuracy matters most: drug dosages, guideline thresholds, escalation criteria, prescribing interactions, and safety-netting advice.
iatroX is also UKCA-marked and MHRA-registered for its UK guideline features — a regulatory status that general-purpose chatbots do not have and are not designed to achieve.
Other purpose-built clinical AI tools exist in this space — UpToDate ExpertAI, EBSCO's Dyna AI, OpenEvidence — and each has different strengths, coverage, and pricing models. The common thread is that they are designed for clinical use, grounded in curated evidence, and built with provenance as a core feature. ChatGPT is none of these things.
A Practical Framework: When to Use What
Use ChatGPT for non-clinical writing tasks, brainstorming, administrative drafting, teaching preparation, and learning support where you will independently verify all clinical content before use.
Use a purpose-built clinical AI like iatroX for clinical questions, guideline retrieval, drug information lookups, differential diagnosis support, and any situation where the accuracy of the medical information directly affects patient care. Use it when you need citations, when you need UK-specific (or jurisdiction-specific) guidance, and when you need to trust that the answer comes from a verified source.
Use primary sources directly — NICE, CKS, BNF, UpToDate, DynaMed, your local formulary — for prescribing decisions, safety-critical protocols, and any situation where you need the definitive, current, complete version of a guideline.
Never use any AI tool as a substitute for clinical judgement, as the sole basis for a prescribing decision, for safeguarding assessments, or in any situation where you would not be comfortable explaining to a regulator exactly how you reached your decision.
The Rules That Apply Regardless of Jurisdiction
Whether you practise in the UK, US, Canada, or Australia, certain principles are universal.
You are responsible for your clinical decisions. The GMC, AHPRA, the medical boards of US states and Canadian provinces all hold the same position: the doctor is accountable for the decisions they make, regardless of what tools they used to inform those decisions. AI does not share your accountability.
Never enter identifiable patient data into a general-purpose AI tool. This applies to ChatGPT, Claude, Gemini, and any other consumer-facing model. Unless a tool has been specifically approved for use with patient data within your organisation's governance framework, assume it is not appropriate.
Always verify AI-generated clinical information against a trusted source. This applies to all AI tools, including purpose-built ones. The habit of checking is as important as the tool you use.
Document your reasoning, not your AI's output. If AI helped you reach a clinical decision, the documentation should reflect your reasoning process, not a copy-paste of an AI response.
Stay current with your regulator's guidance on AI. The GMC, AHPRA, CPSO, and equivalent bodies are all actively developing guidance on AI use in clinical practice. Know what your regulator expects.
Common Mistakes Doctors Make with ChatGPT
Trusting the confidence of the output. ChatGPT sounds authoritative regardless of whether it is correct. The fluency of the language is not a signal of the accuracy of the content. This is the single most dangerous feature of general-purpose LLMs in a clinical context.
Using it for drug dosages or interactions. ChatGPT is not a pharmacopoeia. It does not check the BNF, Lexicomp, or any formulary. A hallucinated dose adjustment or a missed interaction can cause direct patient harm. Use the BNF, your clinical system's drug database, or a tool like iatroX that is grounded in BNF content.
Assuming it knows your guidelines. If you practise in the UK and ask ChatGPT about hypertension management, you may get an answer based on JNC 8 (US) rather than NICE NG136 (UK). The model does not reliably distinguish between jurisdictions, and it does not tell you when it is giving you the wrong country's guidance.
Copy-pasting into clinical records. AI-generated text in clinical documentation that you have not carefully reviewed and edited is a professional risk. The record is yours. Every word in it is your responsibility.
Using it during consultations without patient awareness. If you are consulting an AI tool during a patient interaction, transparency matters. Patients have a right to know how decisions about their care are being made.
Building Better Habits
The goal is not to avoid AI. It is to use it well.
Develop a personal policy. Decide in advance which tasks you will use general-purpose AI for (administrative, educational, non-clinical) and which require a purpose-built tool or primary sources (clinical questions, prescribing, guidelines).
Build a clinical AI stack. For UK clinicians, a practical combination might be: iatroX for rapid guideline retrieval and clinical Q&A, BNF for prescribing, and CKS for point-of-care summaries — all of which are free. For US clinicians, OpenEvidence or UpToDate ExpertAI alongside your institution's resources. For Australian and Canadian clinicians, the same principle applies: use tools grounded in your national guidance.
Make verification automatic. Every time you receive a clinical answer from any AI tool, follow the citation. Check the source. Confirm it is current. This takes seconds and prevents errors that could take months to resolve.
Use AI for learning, not for shortcuts. iatroX's Q-Bank uses spaced repetition and active recall to help you retain clinical knowledge. The Brainstorm tool helps you reason through clinical scenarios step by step. These are learning tools — they make you better at your job, rather than substituting for the knowledge your job requires.
Conclusion
ChatGPT is a powerful general-purpose tool. It is not a clinical reference system. The distinction matters because patient safety depends on it.
Use ChatGPT for what it does well: writing, brainstorming, structuring, and administrative efficiency. Use purpose-built clinical AI tools like iatroX for what they do well: retrieving verified, cited, guideline-grounded clinical information quickly and reliably. And use your own clinical judgement for what only you can do: integrating information with context, weighing uncertainty, communicating with patients, and taking professional responsibility for decisions.
The doctors who will use AI most effectively in 2026 and beyond are not the ones who use the most advanced model. They are the ones who use the right tool for the right job, verify what they receive, and never outsource the thinking that their patients depend on.