- 01Explaining a marker is easy; reading your number is hard
A chatbot can tell you what potassium or ferritin measures almost perfectly, because that knowledge is in every textbook. Reading your own result needs things the chatbot often lacks: the unit, the range printed by your lab, a note about the sample, the markers beside it and your last result. Most chatbot errors on lab reports come from that missing context.
Paste a blood test into ChatGPT and you get a fluent, line-by-line explanation in seconds, and what it says about each marker is usually right. The trouble starts with your own numbers: a glucose read in the wrong unit, a potassium raised because red cells broke open in the tube (a hemolyzed sample), or a flag that 1 in 20 healthy people get anyway.
Free, no accountRead your blood testHolding a report right now? Read every line of it against the range printed on it, free.Read my results
What chatbots get right
A general chatbot is a good translator for a lab report. ChatGPT, Claude and Gemini can expand abbreviations (MCV, eGFR, ALT), say what each marker measures, pick out the values your lab flagged as high or low, and turn a page of numbers into questions for your doctor.
The best direct test comes from the European Federation of Clinical Chemistry and Laboratory Medicine (EFLM). Its working group gave a 2023 version of ChatGPT a set of lab reports and had specialists grade the answers. ChatGPT recognized all the tests and detected the values outside the reference ranges (the normal range printed beside each result), but the group judged its interpretations superficial, not always correct, and only partly coherent when several tests had to be read together 1.
Chatbots rarely fail on medical knowledge. They fail on missing context: the unit, your lab's range, the note about the sample and your last result. Those blind spots are shared by every general chatbot, so the advice here applies to Claude and Gemini as well.
| Good at | Not built for |
|---|---|
| Saying what a marker measures | Knowing which unit your lab used if you leave it out |
| Translating jargon and abbreviations | Using your lab's own range instead of a textbook one |
| Pointing at flagged values | Telling a sample problem from a real result |
| Drafting questions for your doctor | Judging whether a change between two tests is real |
| Summarizing what a panel covers | Spotting a value it has misread from a photo |
Where chatbots go wrong with lab results
Chatbot errors on blood tests cluster in six places, and each can turn a harmless result into a scare or a real problem into reassurance. How to read any lab result walks through reading a report line by line yourself.
- Ranges. A chatbot tends to quote a textbook range, while your result should be read against the range your lab printed beside it. Ranges differ between labs, methods and age groups.
- Units. Glucose of 5.5 means about 99 mg/dL if the lab reports in mmol/L, and a dangerously low value if a chatbot assumes mg/dL. Units often drop off when you copy a table.
- One marker at a time. Low ferritin with low hemoglobin reads differently from low ferritin with normal hemoglobin. Chatbots often explain each line in isolation.
- Change over time. A chatbot compares two numbers. It does not know whether the gap is bigger than normal day-to-day variation.
- Sample problems. A hemolyzed sample releases potassium from the broken red cells and can read falsely high 10. High-dose biotin can push some antibody-based lab tests (immunoassays), including thyroid tests 12 and troponin, a heart-damage test, falsely high or low. A falsely low troponin can hide a heart attack, so if you take biotin and have chest pain, tell the emergency team 4. Biotin explains how long to stop it before a draw.
- Misreading the report. From a blurry photo or a multi-page PDF, a chatbot can pick up the wrong row or invent a value, then explain it with full confidence.
The last failure is measurable. When researchers planted a fabricated detail, such as an invented lab test, in clinical cases, large language models (the technology behind chatbots) expanded on the false detail in about 50% to 82% of cases, depending on the model. A prompt telling the model to be careful cut the average rate from about 66% to 44%, but did not remove it 2.
What the report says vs what a chatbot may say
| Line on the report | What a chatbot may say | The better read |
|---|---|---|
| Glucose 5.5 (unit lost when pasted) | Reads it as 5.5 mg/dL, which would be dangerously low | 5.5 mmol/L is about 99 mg/dL, under the American Diabetes Association's 100 mg/dL fasting cutoff 9 |
| Potassium 6.1 mmol/L H, comment: slight hemolysis | High potassium, a heart risk | Hemolysis can release potassium into the sample 10, and a clinician decides on a repeat draw. Until a repeat confirms it, treat a potassium above 6.0 as real: if the lab or your doctor calls, or you have palpitations, weakness or feel faint, get same-day care. |
| Calcium 10.3 mg/dL H, lab range 8.6 to 10.2 | Hypercalcemia, possible parathyroid problem | A value just above the range is often seen in healthy people. It is read with your last result and albumin (the main blood protein calcium rides on), and your doctor decides whether to repeat it. See high calcium. |
| HbA1c 5.6% last year, 5.9% now | Up 0.3: you now have prediabetes | 5.9% sits in the ADA prediabetes band of 5.7% to 6.4% 9. A diagnosis is your doctor's call, usually after a repeat test, and whether 0.3 is a real change depends on test variation. Prediabetes explains the band. |
| TSH 0.2 mIU/L L, taking a hair and nails supplement | Possible overactive thyroid | High-dose biotin can distort thyroid immunoassays; mention it to your doctor before reading the TSH 12. If it stays low off biotin, see hyperthyroidism. |
CalculatorGlucose converterConvert a glucose result between mmol/L and mg/dL before you paste it anywhere.Work it out Markers from the examples above
Why healthy people get flagged
A reference range is usually the middle 95% of results from healthy people, so about 1 in 20 healthy people land outside it on any single test.
Most labs set a range as the middle 95% of results from a healthy reference group 11, and that 1 in 20 adds up across a panel. Assuming independent tests, the chance that a healthy person gets at least one flag is about 64% on a 20-test panel. A chatbot that treats every flag as a finding will find something wrong with most healthy people. What the H and L flags mean covers how to read a single flag.
People lose the answer in the conversation
In a randomized study of 1,298 UK adults, published in Nature Medicine in 2026, chatbots given medical scenarios on their own named the relevant conditions about 95% of the time. People using the same chatbots to work through the scenarios named them in fewer than 35% of cases, no better than a control group using their usual sources 3. People left out key details, and the chatbots' correct suggestions got lost among weaker ones.
Does it remember your last results?
It can, if you turn memory on. ChatGPT memory and projects can keep earlier results, and in January 2026 OpenAI launched ChatGPT Health, which can connect medical records and wellness apps 6. Remembering a number does not tell it whether a change from 5.6% to 5.9% is bigger than the normal variation of the test and of your body. Why a trend beats a single reading explains how to judge a real change.
Is it safe to paste your results in?
In the US, HIPAA covers healthcare providers, health plans and their business partners. A consumer app you choose to give your results to is generally outside HIPAA 5, so consumer ChatGPT is not a HIPAA-covered service. OpenAI says consumer chats can be used to train its models unless you turn off "Improve the model for everyone" in Data Controls 7, and that ChatGPT Health conversations and connected records are not used to train its foundation models 6. How LifeFrom reads your data explains what happens to results you give us.
- Strip your name, date of birth, address and any patient or insurance number before you paste.
- Turn off model training in the app's data settings, or use a space that states it does not train on your chats.
- Paste the values and units as text instead of uploading the full PDF with its letterhead.
How to ask a chatbot about your blood test
Yes, you can copy and paste your blood work into ChatGPT, Claude or Gemini. What you paste decides how good the answer is, and these steps close most of the gaps above.
- 1Type each value with its unit and the lab's rangeWrite "Ferritin 18 ng/mL (lab range 30 to 400)" rather than "ferritin 18". Lab units explained covers the common unit traps.
- 2Add the context a doctor would ask forAge, sex, pregnancy, medicines, supplements (especially biotin), whether you fasted, and any comment the lab printed about the sample. How to prepare for a blood test covers fasting and which supplements to pause.
- 3Include your previous result if you have oneGive the date and value, then ask whether the change could be normal variation.
- 4Ask about patternsAsk which markers should be read together and what other result would change the interpretation.
- 5Ask what to raise with your doctorDo not ask it to diagnose you or to pick a supplement or dose.
That pattern (low hemoglobin, low ferritin, small red cells) usually points to low iron. Iron deficiency and anemia explain what to ask next.
Can AI suggest supplements or catch what your doctor missed?
A chatbot will make the obvious link: low vitamin D, consider vitamin D. It does not know why the value is low or what else you take, and some supplements distort the next test, as biotin does with thyroid results 12. Supplements explained covers how to judge the evidence before you buy anything.
Most values a chatbot calls out are already flagged on the report. A real miss is more often a slow trend across several tests or a pattern across markers, which is where chatbots are weakest. If you think something was missed, bring the earlier results to your doctor and ask about the trend. For a single borderline flag with no symptoms, ask your doctor whether repeating the test in a few weeks makes sense before acting on it.
Free, no accountRead your blood testGot a panel with a few H and L flags? See what each one measures and which belong together, free, and take the questions to your doctor.Read my results
Citations
- Cadamuro J et al. Potentials and pitfalls of ChatGPT and natural-language artificial intelligence models for the understanding of laboratory medicine test results. An assessment by the European Federation of Clinical Chemistry and Laboratory Medicine (EFLM) Working Group on Artificial Intelligence (WG-AI). Clin Chem Lab Med 2023;61(7):1158-1166.
- Omar M et al. Multi-model assurance analysis showing large language models are highly vulnerable to adversarial hallucination attacks during clinical decision support. Communications Medicine 2025;5:330.
- Bean AM et al. Reliability of LLMs as medical assistants for the general public: a randomized preregistered study. Nature Medicine 2026;32:609-615.
- US Food and Drug Administration. Biotin interference with troponin lab tests: assays subject to biotin interference.
- US Department of Health and Human Services, Office for Civil Rights. Health apps and HIPAA.
- OpenAI. Introducing ChatGPT Health. January 2026.
- OpenAI Help Center. Data Controls FAQ.
- Kaftan AN, Hussain MK, Naser FH. Response accuracy of ChatGPT 3.5 Copilot and Gemini in interpreting biochemical laboratory data a pilot study. Sci Rep 2024;14:8233.
- American Diabetes Association Professional Practice Committee. 2. Diagnosis and Classification of Diabetes: Standards of Care in Diabetes 2025. Diabetes Care 2025;48(Suppl 1):S27-S49.
- Lippi G et al. Haemolysis: an overview of the leading cause of unsuitable specimens in clinical laboratories. Clin Chem Lab Med 2008;46(6).
- Ozarda Y. Reference intervals: current status, recent developments and future considerations. Biochem Med 2016;26(1):5-16.
- Li D et al. Association of biotin ingestion with performance of hormone and nonhormone assays in healthy adults. JAMA 2017;318(12):1150-1160.
Educational context only, not medical advice or a diagnosis. Always discuss your results with a clinician.
