How to read a lab report PDF with AI (and what to check yourself)

You can ask an AI chatbot or a lab-results app to explain a lab report PDF, and it can be a starting point, but treat the answer as a draft to check, not as a conclusion. Before you upload, read the service’s data policy; afterwards, verify the values, the units and the reference range against the printed report, and take anything that worries you to a doctor.

This article is about choosing the route, what to consider before the upload, and how to check the result. For how a lab report is laid out and what reference ranges mean, see how to read your lab results. For the step-by-step import in an app, see how to import lab results from a PDF or photo.

Which AI route should you use for a lab report?

None of these replaces the others. A chatbot can help you prepare questions; your doctor can interpret the result for you. The NHS advice is simple: “Ask to talk to a healthcare professional if you have questions about your results or do not understand them.”

Is it safe to upload lab results to ChatGPT, Gemini or Claude?

A lab report usually carries your name, date of birth and health data. Under GDPR Art. 9, data concerning health is a special category of personal data, and one of the grounds for processing it is explicit consent. Check what each service says before you hand it a file. Policies change, so these notes are dated.

Google Gemini. Google’s Privacy Hub says: “Human reviewers (including trained reviewers from our service providers) review some of the data we collect.” and “Please don’t enter confidential information that you wouldn’t want a reviewer to see or Google to use to improve our services, including machine-learning technologies.” Chats you have with Keep Activity off “are retained with your account for 72 hours”; with Keep Activity on, activity “is auto-deleted after 18 months” by default. Chats selected for human review are not deleted with your activity: Google says “Reviewed chats are retained for up to three years.” Google’s file-upload page says Gemini “may provide a response that misses connections or details throughout the content” for files that are too large.

Claude (consumer plans). Anthropic’s consumer terms update (28 Aug 2025) applies to Free, Pro and Max plans. It says retention extends to five years “if you allow us to use your data for model training”, and that you can change the choice in Privacy Settings at any time. Without that permission, the retention period is 30 days. These are the 2025 terms; check the current ones.

ChatGPT. Reporting by The Register and The Decoder says ChatGPT Health was announced on 8 January 2026; at launch it was not available in the EEA, Switzerland or the UK, and medical-record integrations were US-only. The Register quotes OpenAI: “Conversations in Health are not used to train our foundation models.” Check OpenAI’s current data-controls settings before uploading.

Practical rules that follow from this:

• Crop or redact your name, date of birth and ID numbers before uploading when the question does not need them.
• Prefer a setting that keeps chats out of training, and delete the conversation when you are done, though Google says reviewed chats are kept separately.
• If you would not want a human reviewer to read it, do not paste it into a general chatbot.

How accurate is AI at reading lab values?

Studies differ in design, and several tested older models on simulated or narrow tasks.

• Cadamuro et al. (2023), a working group of the European Federation of Clinical Chemistry and Laboratory Medicine, gave ChatGPT 10 simulated lab reports and concluded that it “may only be considered a tool capable of interpreting a laboratory report on a test-by-test basis at best”.
• In a pilot study, Kaftan et al. (2024) tested ChatGPT-3.5, Copilot and Gemini on 10 simulated patient datasets and found that “Copilot is more accurate in interpreting biochemical data than Gemini and ChatGPT-3.5”.
• He et al. (2024) compared five language models with human answers to 53 lay-patient questions from Yahoo! Answers. “GPT-4’s responses achieved better scores than all the other LLM responses and human responses”, although the authors also noted incorrect statements, missing context and lack of references.
• Ye et al. (2026) tested GPT-5, Grok 4 and DeepSeek R1 on 100 complete blood count reports from haematology patients. GPT-5 matched the gold-standard diagnosis (top-1 diagnostic concordance) in 93% of cases when the prompt included the clinical suspicion and in 79% without it (Grok 4: 89% to 72%; DeepSeek R1: 92% to 77%). The authors describe it as a single-centre, Chinese-language, exploratory study that provides only preliminary evidence, so the figures may not carry over to other settings or languages. They conclude that “current clinical deployment requires human oversight and quality control.”

Track your lab results in HealthLab

First try free

Download

The same answer can be right for one report and wrong for another, which is why the checklist below matters.

A checklist to verify any AI answer

1. Check that the values were copied correctly. Compare each number in the AI’s answer with the PDF. A misread digit or a dropped decimal point changes the meaning completely.

2. Check the units. MedlinePlus notes that “test results are measured with different units of measurement”. The same analyte can be reported in different units, so a number from one report can look wrong against a reference found elsewhere. These factors come from the Labcorp table; the vitamin D factor is from Mayo Clinic Laboratories.

These factors are rounded, so a converted value can differ slightly from the one a lab prints; the Labcorp figures are used except vitamin D (Mayo Clinic Laboratories).

3. Use the reference range printed on your report. Testing.com says “A normal result in one lab may be abnormal in another.” and “You must use the range supplied by the laboratory that performed your test”. If an AI answer quotes a “normal range” that differs from your report, trust the report.

4. Remember what a range means. A review in the J Am Board Fam Med (2025) says “The central 95% of the values is usually adopted as the normal range.” and “This practice, by definition, labels 5% of healthy individuals as having abnormal laboratory results.” That is about 1 in 20 healthy people on a given test. Labs mark results as low, high or abnormal, and MedlinePlus says “It’s common for healthy people to sometimes have results outside the reference range.”

5. Out of range is not a diagnosis. Testing.com says “A test result outside the reference range may or may not indicate a problem”. A chatbot does not know your history, symptoms or medicines unless you tell it, and cannot examine you.

6. Know that critical values go to the clinician. Labcorp states that “Critical results are considered life threatening and require immediate notification of the physician”. If you have symptoms that worry you, do not wait for a chatbot’s opinion.

7. Look at the trend, not one point. One value is a snapshot; several results over time show direction. See how to read a trend chart.

When to talk to a doctor

MedlinePlus advises that you “talk with your provider about your results, even if they’re normal.” Do it sooner when an AI answer and your report disagree, when several values are outside the range, or when you have symptoms. Bring the original report, not just the AI summary.

HealthLab: recognition with a human check

HealthLab is an iPhone app for storing lab results. You can import a PDF, a photo, a clipboard image or a file from the Share sheet; Gemini recognizes the values, units and reference ranges, and you review and edit the rows before saving. It works with reports from any lab.

• Import tiers: one import is free, unlimited import is Pro. Pro is a subscription with a 7-day free trial.
• Manual entry: manual entry is free and unlimited.
• Family: separate profiles per person; additional profiles for family members come with Pro.
• Charts per biomarker, comparison of up to 3 biomarkers, and a PDF report for your doctor are free. An AI summary of one analysis has one free try, then Pro.

Privacy: there is no account. Data is stored locally on your iPhone via SwiftData; iCloud sync is optional and off by default. Recognition and AI go through our server to Google Gemini and are not stored. Everything in the policy notes above concerns the chat apps’ own settings and does not describe HealthLab.

The app is not a medical device and does not provide medical advice, diagnoses, or treatment recommendations. Try HealthLab on the App Store.

Frequently asked questions