AI for Doctors: A 60-Second Safety Checklist for Answers

AI for doctors saves time only if you check the answer. Run the five checks below in about a minute, and treat any answer that fails as unverified until you read the primary source.

What should I check on any AI answer before I use it?

Run five checks: cite, support, date, PHI, scope. The first two catch most errors.

1. Cite-check. Click every citation. Prefer answers that link the paper, FDA label, or society guideline directly. A bare author-and-year string needs a manual PubMed search.
2. Support-check. Find the sentence in the source that states the number or recommendation. Discard the claim if you can't find it.
3. Date-check. Find the society name and guideline year, then look for a newer version on the society's site. Compare any dose, cutoff, or risk score with the current guideline or FDA label. The PICO and PubMed search guide helps, and the PCE versus PREVENT comparison shows how much calculator version changes the output.
4. PHI-check. Leave identifiers out unless your organization has cleared the tool.
5. Scope-check. Ask what the tool is. Software that analyzes images or signals, or whose basis you can't review, stays under FDA oversight.

How often do AI citations fail?

Often enough to check every one. Published tests found fabricated references and real references that don't support the text.

Can I put patient information into an AI tool?

Not unless your organization has cleared the tool for PHI. Otherwise, strip identifiers and write the question about a generic clinical scenario.

HIPAA lets a covered entity disclose PHI to a business associate only with satisfactory assurance, documented in a written contract or other written arrangement. A business associate creates, receives, maintains, or transmits PHI on the covered entity's behalf. Whether a chatbot vendor qualifies depends on the product tier and contract, so ask your compliance office. I could not open HHS's own pages on AI tools, so check hhs.gov for current OCR statements.

Where does FDA oversight of clinical decision support software end?

FDA oversees CDS software functions that fail any of four statutory criteria. Non-device CDS must meet all four.

The CDS guidance was reissued January 29, 2026 and supersedes the January 6, 2026 version. The software must not analyze medical images or signals. It must display, analyze, or print medical information. It must support or recommend to a health care professional about prevention, diagnosis, or treatment. And it must let that professional independently review the basis, so they don't rely primarily on it. The guidance cites automation bias, over-reliance on automated suggestions, as the reason the fourth criterion matters most in time-critical settings.

Regulatory status leaves accountability with you. The ANA position statement says nurses are accountable for their practice even when technology fails, and AANP says the provider's clinical decision making should stay the defining factor. I found no standalone AAPA statement I could open, so PAs should check their employer's policy.

Where does this fall short?

The checklist catches visible errors. It misses a real source that agrees with a wrong claim, and a search that skipped a contradicting trial. The accuracy studies tested specific, older models, and no source here evaluates any tool's clinical accuracy. FDA guidance, as I read it, doesn't mention large language models, so it doesn't settle how general-purpose chatbots are classified. I confirmed the January 29, 2026 date on the document itself on 2026-10-10 and checked for no later revisions.

CliniWiz displays its reasoning and links its sources, which speeds the first two checks.