Uploading Lab or Genetic Results to Apps and AI: Privacy and Accuracy

A note first

This page does not read, interpret or give advice about any lab or genetic result. If you have a result you are worried about, talk with the clinician who ordered it. Here we cover two questions: who can see your data, and how reliable AI reading of results is.

Evidence labels

We use plain labels such as promising but limited, mixed, insufficient evidence and marketing claim. See How We Label Evidence in Biohacking.

Privacy: who is covered?

Health privacy law may not follow your file. The US Federal Trade Commission (FTC) says its Health Breach Notification Rule applies to health apps and similar technologies not covered by HIPAA. [1] It says the definition of a breach includes both data security breaches and unauthorized disclosures. [1] The FTC post is dated April 26, 2024. [1] That means a consumer app may sit under FTC rules rather than under the health privacy rules that bind your doctor, though we did not read the US Department of Health and Human Services guidance on this (its page was unreadable to us this session). Whether a given chatbot or app is covered depends on the company and the setting. See Health Data vs Consumer-App Data.

Genetic data raises extra questions. In the US, the Genetic Information Nondiscrimination Act protects against some health insurance and employment uses, but its health insurance protections do not cover long-term care, life or disability insurance. [2] Some states have extra laws. [2] This is not legal advice, and laws outside the US differ; we have not verified them. Read Genetic and Location Data Privacy and Genetic Privacy Law Overview.

Accuracy: what studies show

Structured lab values, controlled settings: promising but limited. A 2026 study tested four LLMs on Korean national health checkup data. Zero-shot accuracy averaged 0.69, rising to 0.92 with combined prompting strategies. [3] Accuracy was near-perfect for most biochemical markers but lower for blood pressure, and the authors found model-specific sex and age biases and said caution is needed for complex variables. [3] This tested whether a model labeled values correctly against a reference, not whether people were helped.

Case vignettes: promising but limited. A blinded 100-case benchmark of three LLMs in laboratory medicine scored them highly, and the authors said the results should not be read as evidence of clinical superiority or real-world effectiveness. [4]

Real people using chatbots: mixed. In a randomized study of 1,298 people, models alone identified conditions in 94.9% of scenarios, but people using the models identified relevant conditions in fewer than 34.5% of cases, no better than control. [5] People also did not consistently spot inaccurate chatbot advice. [6]

Genetic results and AI: we did not find a study of public use of chatbots for reading consumer genetic reports, so the evidence there is insufficient. See Raw Genetic Data Interpretation and Genetic Test False Positives and Clinical Confirmation.

Reference ranges and supplements

The FDA notes that some supplements can interfere with lab tests. [7] Context from the clinician who ordered the test matters. See Multiple Testing and Overdiagnosis.

Questions to ask before you upload

1. Who is the company, and is the service covered by health privacy law or only by general consumer rules?
2. Is my data used to train models or shared with partners, and can I opt out or delete it?
3. Does it include my name, date of birth or other identifiers I could remove?
4. If this is a genetic file, have I read the privacy guides linked above?
5. Will I act on the answer without checking with a clinician?