AI for doctors, nurses and health workers

Where things stand (as of 2026)

AI tools are genuinely helpful for administrative and communication tasks: drafting clinical notes from dictation or shorthand for a clinician to review, turning technical information into patient-friendly language, and helping search medical literature more efficiently. Some health systems have integrated approved, governed AI scribe or documentation tools directly into clinical workflows, with proper information governance sign-off.

What remains genuinely dangerous: using general-purpose AI chatbots for diagnosis, or for dosing and drug interaction information, without independent verification against an authoritative source. These tools have been documented to produce confident, fluent, and wrong medical information, including invented dosing figures and plausible-sounding but incorrect drug interaction claims. No general-purpose AI chatbot available to the public is currently approved as a medical device or diagnostic tool in most jurisdictions, and using one as if it were is both a patient safety risk and, in many settings, a professional conduct issue. Regulatory frameworks for AI in healthcare are actively developing and vary significantly by country and by whether a tool is marketed as a medical device - check current guidance rather than assuming last year's position still holds.

The mental model

Picture the AI as a well-read but unlicensed assistant who has absorbed enormous amounts of medical text but has never examined a patient, doesn't know this patient's history beyond what you type, and cannot be held accountable the way you can. It's genuinely useful for turning your own clinical judgment into clear notes or plain language for a patient. It must never be the thing that supplies the clinical judgment itself.

Scenarios that work

Where it fails, and what it costs you

Diagnosis-by-chatbot is dangerous. A model producing a fluent, confident differential diagnosis from a patient's typed symptoms, or a clinician's brief prompt, is not the same as clinical assessment. It has no access to physical examination, vital signs, imaging, or the countless small observational cues a trained clinician picks up in person. Patients using consumer chatbots for self-diagnosis, or clinicians leaning on them as a shortcut for genuine clinical reasoning, both risk missing something a proper assessment would catch.

Hallucinated dosing information. This is one of the most serious documented failure modes of general-purpose AI in a healthcare context. A model can state a drug dose, frequency, or maximum daily limit with total confidence while being simply wrong. Unlike a typo in a report, a wrong dose can directly and immediately harm a patient. This must always be checked against your jurisdiction's authoritative formulary, never accepted from an AI tool alone.

Patient data and consent. Typing identifiable patient information into a consumer AI tool without explicit organisational approval, a proper data processing agreement, and - in many jurisdictions - patient consent, is very likely to breach data protection law and your professional confidentiality obligations, regardless of good intent. Some tools retain and use input for further training unless specifically excluded by contract.

Device and regulatory boundaries. In many jurisdictions, a tool that makes or materially informs a diagnostic or treatment decision may cross into being regulated as a medical device, requiring formal approval. A general-purpose AI chatbot has almost never gone through that process. Using it in a way that functions as a diagnostic aid, even informally, can put you outside your organisation's approved practice and outside the regulatory framework meant to protect patients.

Literature summaries can misrepresent findings. A one-line AI summary of a study can flatten an important caveat, misstate a sample size, or blend two different papers' findings together. Practice-changing decisions should never rest on a summary alone.

Human-in-the-loop is not optional. Every credible framework for safe AI use in healthcare rests on a human clinician retaining final judgment and accountability. If a workflow is structured so that AI output goes straight to a patient-facing decision without a qualified human checking it, that workflow is unsafe regardless of how good the tool's average accuracy looks in testing.

How to check the answer

1. Confirm any dose, frequency, or interaction claim against your jurisdiction's current authoritative drug reference before acting on it.
2. Check any AI-drafted clinical note line by line against the original observation or dictation - confirm nothing was added or changed.
3. Verify any literature claim by reading the actual paper's abstract or full text, not a one-line AI summary.
4. Confirm no unapproved use of identifiable patient data occurred, per your organisation's information governance policy.
5. Have a second qualified clinician review any AI-assisted content before it affects a patient care decision.
6. Ask yourself directly: would I be comfortable explaining this AI use to my regulator or professional body if asked?

How to read the docs and find the truth

For clinical facts, go to your jurisdiction's authoritative sources: your national formulary (such as the BNF), clinical guideline bodies (such as NICE in the UK, or your national/specialty equivalent), and peer-reviewed literature accessed directly, not summarised secondhand. These sources publish review and update dates - check them, since guidance changes.

For the AI tool itself, look for a genuine model card or system card from the vendor describing intended use and known limitations - a credible one will explicitly state it is not intended for diagnostic or treatment decisions if that's true. Check whether your organisation's medicines and healthcare products regulator (such as the MHRA in the UK, or the FDA in the US) has published any classification or warning about the specific tool or type of tool you're considering. Be wary of vendor marketing that implies clinical reliability without a corresponding regulatory approval or peer-reviewed validation study behind it - a real safety claim will point to a specific, checkable study or approval, not just a general assertion of accuracy.

Your one-page playbook

ROLE: You are helping with [documentation / plain-language explanation /
literature search strategy] only - not diagnosis, dosing, or treatment decisions.
INPUT: [dictation, shorthand, or anonymised case details - no identifiable
patient data unless this tool is specifically approved for it].
BOUNDARY: Do not add clinical detail not explicitly given. Do not suggest a
diagnosis, dose, or treatment. Flag gaps or ambiguity for me to resolve.
OUTPUT: [format: note structure / patient reading level / search strategy].
FINAL CHECK: I will verify all clinical content against [authoritative source]
and my own clinical judgment before this affects patient care.

Check yourself

1. What specific patient data can and cannot go into the AI tool you're considering, according to your organisation's policy?
2. If an AI tool gives you a drug dose or interaction claim, what is the one authoritative source you must check before acting on it?
3. Why is a fluent, confident-sounding AI response not the same as a safe one, in a clinical context?
4. Who remains accountable for a clinical decision that was informed by AI output - and does your current practice reflect that clearly?