AI for medical records: frequently asked questions
Answer in brief: 12 questions physicians and clinic teams ask before using AI on records, each answered in 2 to 4 sentences with a link to the page that goes deeper.
Limits: Vendor facts come from each vendor's own website, read 2026-11-12, and we haven't tested every tool; prices and features change, so check the vendor before you buy.
Safety and checking
What is the most common way AI gets a record wrong?
Leaving things out. In the 2025 Asgari study, omissions outnumbered hallucinations about 9 to 1 by sentence count. Flipped negatives are the most dangerous hallucination, because they look plausible. See the risks page.
How much of the AI output do I need to check?
Every fact your opinion, plan or signature rests on, by opening its cited page, plus a few random lines. You don't need to re-read the whole file if the tool cites its sources; you do if it doesn't.
Can AI read handwritten notes and faxes?
Many tools run OCR and handwriting recognition, with quality that varies by tool and by page. Put your worst handwritten page in the test file and check what came out; an unreadable page should be flagged, not skipped.
Can AI tell me a record is missing?
It can flag likely gaps by matching bills, orders and referrals against the records present, and by flagging long breaks between visits. It can't detect care that left no trace in the file. See missing records.
Privacy and contracts
Is it a HIPAA violation to paste a record into ChatGPT?
Pasting identifiable patient records into a consumer account with no BAA is the risk HHS guidance addresses: a service handling PHI for a covered entity is a business associate and needs a BAA. Use a tool or an organizational AI service that signs one.
Do AI vendors train on my patients' records?
Some say they don't and put it in the contract; others leave it to a setting. Ask for a written term, ideally in the BAA, that customer data never trains or tunes a model.
What should a BAA from an AI vendor cover?
Permitted uses of PHI, safeguards, breach notice, return or destruction of data at the end, and the vendor's subprocessors, including the company that runs the AI model. Ask before the trial, not after.
Cost and fit
How much does AI record review cost per file?
At published rates of $0.05 to $0.15 a page, a 1,000-page file costs $50 to $150. Allowance plans and seat pricing change the math with volume; use the calculator.
How large a file can these tools handle?
Medical-legal tools are built for files of thousands of pages; general AI chat tools may cap uploads or the length of text they read at once. Ask the vendor for the largest file a customer has run, and test with yours.
Should a small practice buy AI for records at all?
Only for a task that costs real hours today. For most clinics that is note writing or pre-visit chart review, often already in the EHR. For physicians who review outside files, per-page tools with no subscription fit occasional work.
Patients and AI
What can patients' AI apps read from my notes?
With the patient's permission, ChatGPT Health can connect records from supported US provider portals, and Claude can read athenahealth records through the athenaPatient MCP connector: labs, medications, conditions, visit history and more. See patient apps.
How should I respond when a patient brings an AI summary?
Treat it like any outside summary: ask what tool they used and what records it read, then check the points that matter against the chart. OpenAI states its Health product is not intended for diagnosis or treatment.