Is it safe to upload your medical records or lab results to ChatGPT?

Your doctor's office and your lab are bound by HIPAA. A chatbot you open on your phone is not. That one fact decides most of the risk. When you upload medical records or a lab report to ChatGPT, Claude or Gemini, the protection on that copy comes from the company's privacy policy, the settings you chose, and how much identifying detail was on the page.

Why HIPAA stops at the upload

HIPAA's Privacy Rule applies to health plans, healthcare providers and the billing middlemen between them (the law calls these covered entities), plus the contractors who handle records for them (business associates). A chatbot you choose yourself is neither. HIPAA's right of access, strengthened by the 21st Century Cures Act's rules on standard apps and data connections, lets you send your records to any app you like, and once they arrive there HIPAA no longer applies. The U.S. Department of Health and Human Services (HHS) says a provider is not liable for what your chosen app does with the data, and may not refuse to send it only because it worries about the app ¹.

Once health information is received from a covered entity, at the individual's direction, by an app that is neither a covered entity nor a business associate under HIPAA, the information is no longer subject to the protections of the HIPAA Rules.HHS Office for Civil Rights, The Access Right, Health Apps, & APIs ¹

Some federal and state protections still apply outside HIPAA. The Federal Trade Commission's amended Health Breach Notification Rule, in force since 29 July 2024, requires health apps outside HIPAA that keep personal health records to tell users, the FTC and sometimes the media after a breach, and it counts sharing data without permission as a breach ³. The FTC can also act under its general consumer protection powers against companies that share health data with advertisers in ways their privacy promises did not allow. Some states, Washington among them, have consumer health data laws that require consent before a company collects or shares that data.

What ChatGPT Health promises, and what it leaves out

OpenAI launched ChatGPT Health in the US in January 2026 ⁴. Health is a separate space inside ChatGPT where you can connect medical records (through a records service called b.well) and Apple Health data. OpenAI's main promises for that space, and what each one means in practice:

• Connected records and Health conversations are not used to train OpenAI's AI models or to target ads ⁴. The promise covers the Health space, so a lab value pasted into an ordinary chat follows your normal training setting.
• Health has its own encryption, isolation and memory, and Health memories stay out of your regular chats ⁴.
• OpenAI says you can disconnect a records source in Settings, which stops Health's access to it ⁴. Deleting your Health chats is a separate step.
• The consumer Health feature is not a HIPAA product. OpenAI's HIPAA-eligible offerings, with a signed business associate agreement, are business products such as ChatGPT for Healthcare, sold to organisations.

Claude, Gemini and other chatbots: the settings that matter

Each consumer chatbot has a training switch, a temporary chat mode and a business tier with stronger terms. Anthropic changed Claude's consumer terms on 28 August 2025: Free, Pro and Max chats can be used for training only if you allow it, and Anthropic keeps that data for up to five years; if you opt out, chats are kept for 30 days ⁵. Google's Gemini app saves chats to your account through its activity setting unless you turn it off, and Google's own privacy page for Gemini explains how long it keeps chats and who can review them. The table reflects each company's settings as of November 2026. Policies change, so check the date on the setting page before you rely on it.

In short: on any of these apps, switch the training setting off and use a temporary chat. That keeps your lab values out of training and shortens how long the company holds them.

The risks that remain even with the right settings

Good settings reduce what a company keeps and learns from. They do not remove every way a copy can leak or be read later.

• Breach: any stored chat can be stolen in a hack.
• Account access: a shared laptop or a reused password lets someone else scroll your history.
• Legal process: a court order in a lawsuit can require a company to keep chats, including ones you deleted.
• Policy change: a company can rewrite its terms, and the new terms can cover chats you send after the change.
• Sale or bankruptcy: user data can pass to a buyer.
• Other people's data: family history, a partner's name or a child's result on your report belongs to someone who never agreed to share it.

What to remove before you share a lab report

HIPAA's Safe Harbor method, the standard for making a record de-identified, lists 18 identifiers to strip, including names, any address detail smaller than a state, every date element except the year, ages over 89, phone numbers, email, medical record numbers, health plan numbers and account numbers ². Applied to a typical lab report, that list becomes a short checklist. Removing these fields still leaves everything a chatbot needs to explain the results.

Units and printed ranges matter most on the keep list, because the same marker can be reported in different units and each lab sets its own range. The guide to lab report flags shows why the H and L marks only make sense next to the lab's own range.

Before you start: if a result is flagged critical or your doctor's office has called, contact them first. Reading it in a chatbot can wait.

1. 1Copy the results into plain textRetype each marker, value, unit and range, or export the PDF to plain text and delete the header. A black box drawn over a PDF often leaves the text underneath, and anyone can copy it out.
2. 2Crop photos and screenshotsCrop to the results table only. Phone photos can carry hidden details, including the GPS location where the photo was taken, so a screenshot of the cropped area is safer than the original photo.
3. 3Turn off training and use a temporary chatSwitch off the training setting for your chatbot, then start a temporary chat before you paste anything. Include earlier results too, so the chatbot can see the trend.

Is a chatbot accurate enough to be worth it?

For general explanations, often yes. In a 2024 study of 53 real lab-result questions, GPT-4's answers were rated above those of four other language models and above answers from other patients on relevance, correctness, helpfulness and safety ⁶. The same study found GPT-4 still gave some inaccurate answers and some that ignored the person's situation. Chatbots most often slip on four things with labs: assuming a textbook range instead of the lab's printed one, confusing units, reading one result without the earlier ones, and missing the effect of a medication or supplement such as biotin (see supplements that affect blood tests).

Two habits close most of those gaps. Always paste the range and unit with each value, and paste previous results too, since a trend says more than a single reading. The guide on reference ranges explains why your lab's range can differ from the one a chatbot quotes, and how to read any lab result walks through each column of a report so you can check the chatbot's answer yourself. Treat a chatbot's reading as a starting point for the conversation with your doctor, and check with them before you start, stop or change any medication or supplement because of it.

Free, no accountRead your blood testHolding a lab PDF? Read every line in plain words.Read my results

If you do not have a copy of your results yet, the guide to getting your lab results covers patient portals and direct requests, and the guides for Quest and Labcorp reports show where each field sits on the page.

Free, no accountDrop in your lab report and see every marker explained in plain words, placed on the body, against your lab's range. It's read on your device: nothing is uploaded and nothing is kept.

Read my results Download the app