How to use AI on your lab results without uploading them
AI chatbots are good at some things people want from their lab results: explaining what a test measures, putting a flagged value in plain language, or suggesting questions for the next appointment. The catch is where the results go. Paste a report into a chatbot and it now sits on that company’s servers, under that company’s rules.
You do not have to choose between using AI and keeping your results private. This post covers what happens to results you upload, how to run a model entirely on your own computer, and what to remove if you do use an online service.
What happens when you upload results
HIPAA protects your records while a doctor, hospital, lab, or health plan holds them. It does not follow a copy you hand to someone else. Once you paste a lab report into a consumer app, the copy there is governed by the app’s privacy policy and terms, plus whatever state privacy law and the FTC’s general consumer protection rules apply. That is true of chatbots and of most health apps.
The big AI companies now offer health features built for exactly this. OpenAI rolled out Health in ChatGPT across the US in July 2026, letting users connect medical records through a partner called b.well, and says records shared with it will not train its models or target ads. Coverage at the launch pointed out that HIPAA no longer protects those records once they leave the provider. (Paubox, July 2026) Anthropic added health record and lab connections for Claude Pro and Max subscribers in the US in January 2026, and says it does not use health data to train models. (Anthropic, January 2026)
Those commitments may be sincere and well kept. They are still company policies, which can change, rather than law. Whether that is acceptable is your call. If it is not, there are two other routes.
Option 1: Run the model on your own computer
Open models now run comfortably on an ordinary laptop, and nothing you type leaves the machine. Two free apps make this simple on a Mac or Windows PC:
- Ollama downloads and runs open models from a command line or a simple chat window.
- LM Studio has a chat interface much like ChatGPT’s and lets you browse and download models from inside the app.
Install one, download a mid-sized model (the apps suggest ones that fit your memory), and once the download finishes you can turn off Wi-Fi to prove to yourself nothing is being sent. Then give it your results.
Getting your results into it. A model reads text, so the easiest input is a table: test, value, unit, reference range, date. If your results are PDFs, our lab PDF to spreadsheet tool turns a report into a CSV in your browser without uploading it. Paste the rows into the chat.
What to expect. A model small enough to run on a laptop knows less than the large online ones and makes more mistakes, especially with arithmetic, unit conversions, and reference ranges it half remembers. Use it to explain and to draft questions, not to calculate. Check every number it repeats back against your report.
Option 2: Strip what identifies you, then use an online model
If you would rather use a large online model, send only what it needs. A lab value with no name attached reveals far less than a full report. Before pasting, remove:
- Your name, date of birth, and address
- Medical record, accession, and account numbers
- The ordering doctor’s and the lab’s names and addresses
- Exact dates, if they are not essential. “Six months later” works as well as a date.
What remains, a list of tests and numbers, is still personal. Combined with other details you mention in the conversation, it can point back to you. But it is a much smaller thing to hand over than a scanned report with your name on every page.
API terms differ from app terms. If you use a model through a developer API key, check that service’s API terms, not its consumer app’s. Google’s Gemini API, for example, says content sent on its unpaid tier may be used to improve Google’s products and read by human reviewers, while the paid tier’s content is not used to improve its products. (Gemini API terms)
What to ask, and what not to
AI is useful for:
- Explaining a test. What does ALT measure, and what tends to raise it?
- Translating a report. What does “hemolyzed” or “see note” mean?
- Preparing for an appointment. Turn your results and worries into a short list of questions.
It is weak at:
- Judging whether a change is real. Two results a few percent apart can be ordinary variation from draw to draw, or a change of lab. Our per-test guides, such as how to track your TSH over time, explain how much each test varies.
- Reference ranges. A model may quote a “normal range” that is not your lab’s. Use the range printed on your report.
- Diagnosis and treatment decisions. These belong with your doctor, who knows the rest of your history.
The most useful input you can give any model, local or online, is the trend rather than one report: the same test over several years, with the date each medication started. How to track your lab results over time covers building that history.
Where HealthViewer fits
HealthViewer is a lab results tracker for Mac and Windows that keeps your results in an encrypted file on your computer. Its AI features are optional and off until you add an API key, and we want to be plain about how they work: when you use one, such as the AI summary in the Appointment Prep report, the relevant results are sent from your computer directly to the AI provider you chose. HealthViewer does not run models locally yet.
If you want AI with nothing leaving your computer, HealthViewer still helps. Export your results as CSV or JSON, narrowed to the dates or category of tests you care about, and paste them into a local model in Ollama or LM Studio. Everything else in HealthViewer, from PDF import to charts, works without AI at all.