AI can support clinical decision-making — but the word "support" is doing important work. Support means the AI provides information that the clinician evaluates, verifies, and integrates with patient-specific context before making a professional decision. It does not mean the AI makes the decision and the clinician rubber-stamps it.
Every AI-generated clinical answer should pass through a verification process before it influences patient care. This checklist applies to every clinical AI tool — whether it is Doximity Ask, ChatGPT, OpenEvidence, iatroX, or any other platform.
The Checklist
1. What source is the answer based on?
Can you see where the information came from? Is the source visible and specific — a named guideline, a specific SmPC section, a cited study? Or is the answer presented without attribution, requiring you to trust without verifying? If the tool cannot show its source, the answer is unverifiable.
2. Is the source appropriate for my country and clinical setting?
A guideline from the American College of Cardiology may not apply to UK practice. A US drug label may cite different licensed indications from the UK SmPC. A hospital-level recommendation may not be appropriate for primary care. The source must match the clinician's jurisdiction and clinical setting.
3. Is this a guideline, a drug-label source, expert opinion, or primary research?
Not all evidence is equal. A NICE guideline recommendation carries different weight from a single case report. An SmPC contraindication is a regulatory statement with legal implications. An expert opinion in a journal editorial is a professional perspective, not a definitive recommendation. The type of source matters for how the clinician weighs the answer.
4. Is there a visible review or update date?
Clinical evidence changes. A guideline that was current two years ago may have been superseded. A drug interaction list that was accurate at the AI's training date may be missing more recent additions. The clinician should be able to assess the currency of the information.
5. Does the answer show uncertainty where evidence is unclear?
Real clinical evidence has gaps, conflicts, and areas of uncertainty. An AI answer that presents everything with equal confidence — regardless of evidence quality — is hiding uncertainty behind fluency. The tool should distinguish between strong evidence ("NICE recommends...") and weaker evidence ("limited evidence suggests...") and frank uncertainty ("guidelines do not address this specific scenario").
6. Does the tool have a fail-safe mechanism?
When the available evidence is insufficient, conflicting, or poorly matched to the question, what does the tool do? Does it narrow the answer and show uncertainty? Does it surface the source trail so the clinician can evaluate the evidence themselves? Does it decline to provide a definitive conclusion when one is not supported? Or does it generate a confident-sounding answer from insufficient evidence — the most dangerous failure mode in clinical AI?
7. Can I report an error or unclear answer?
Feedback mechanisms are quality-improvement infrastructure. A tool that allows clinicians to flag inaccurate, unclear, or potentially harmful outputs creates a correction loop that improves the system over time. A tool without a feedback mechanism has no systematic way to improve from real-world clinical use.
8. Is the tool being used within the right governance framework?
Is the tool approved for use in your clinical setting? Does your organisation have a governance policy for clinical AI use? Is the tool registered with the MHRA (for UK) or FDA (for US) as a medical device, if its intended use warrants it? Is there a clinical safety case? Are data processing agreements in place?
UK Governance Context
The MHRA recognises that software, including AI, may be regulated as a medical device in the UK depending on intended use. NHS England's guidance on ambient scribing products emphasises implementation considerations including governance, safety, and user review of outputs. The regulatory expectation is proportionate governance: tools with greater clinical influence require greater governance oversight.
How iatroX Aligns with This Checklist
iatroX is built around this verification framework. Professional-facing (designed for clinicians and healthcare professionals), source-grounded (curated UK clinical sources including NICE, CKS, SmPC/eMC), citation-visible (provenance display showing where answers come from), fail-safe (narrowing, abstaining, or escalating when evidence is insufficient), and feedback-enabled (clinician reporting of unclear or inaccurate outputs). UKCA-marked, MHRA-registered.
The Standard
A trustworthy clinical AI answer is not just fluent. It is sourced, scoped, reviewable, cautious where evidence is insufficient, and correctable through professional feedback. That is the standard every clinical AI tool should be measured against.
Ask iatroX is designed for clinicians who need answers they can verify →