Can ChatGPT Analyze Your 23andMe Raw Data? What Actually Happens

You downloaded your raw data file from 23andMe or AncestryDNA. It is a plain text file of rsIDs, chromosome positions, and letter pairs — hundreds of thousands of rows, completely unreadable to a non-specialist. So you did the obvious thing: you pasted it, or uploaded it, into ChatGPT and asked it what it means.

That is an extremely common thing to do, and this article is not here to tell you it is a foolish one. It is here to explain, as plainly as we can, what is actually happening on the other end when you do it — what the model is and is not doing with your file, why asking the same question twice can produce two different answers, and what the published research on language models and genetics questions actually shows.

One disclosure before any of that: DecodeMyBio sells a paid product that answers the same underlying question this article is about — what your raw data means for your medications. We have an obvious commercial interest in you believing that a general-purpose chatbot is not the right tool for that job. That doesn't make the argument below wrong, but you should read it the way you'd read any comparison written by someone with a stake in the answer — skeptically, and with the sources checked. That is exactly why every technical and research claim in this article links to its original source, rather than asking you to take our word for it.

What Happens When You Upload a Raw DNA File to an AI Chatbot

A 23andMe raw data export typically contains 600,000 to over 700,000 rows, each one a genetic marker (rsID), its chromosome and position, and the two letters you carry at that position. AncestryDNA files are similar in size and structure. It is a large, flat text file — not a curated report.

When you hand that file to a chat interface, two things determine what happens next. First, how much of it the model actually sees: language models process a limited context window per request, and a file this size may be partially truncated or only partly represented in what the model is actually reasoning over, depending on how the upload is handled. Second, and more fundamentally: what the model does with whatever portion of the file it does see.

Here is the part that is easy to miss. A general-purpose language model does not, at the moment it answers you, run a lookup against a curated table of genotype-to-guideline relationships — it is not querying CPIC, PharmVar, or ClinVar the way a purpose-built pipeline would. Unless a specific product has wired it up to do exactly that — the technique is called retrieval-augmented generation, and a plain chat session is not automatically doing it — the model is generating a written answer, one word at a time, based on patterns learned during training, conditioned on whatever text you gave it. That is a fundamentally different operation from looking a specific row up in a specific table.

Why the Same File Can Give You Two Different Answers

This is the part people find genuinely surprising: run the exact same file, with the exact same question, through the same model twice, and you are not guaranteed to get the same answer back. That is not a bug in one particular product. It is how this class of model works. As Anthropic's own technical documentation puts it, describing behavior that is not unique to any one company's models: “Users may encounter non-determinism in APIs. Even with temperature set to 0, the results will not be fully deterministic and identical inputs may produce different outputs across API calls.”

“Temperature” is the setting that controls how much randomness gets injected into which word the model picks next. Text generation is a sampling process — at each step, the model computes a probability distribution over what the next word could be and draws from it, rather than retrieving one correct, pre-existing answer. Turn that randomness down and answers get more repetitive; turn it up and they get more varied. But even at its most conservative setting, the underlying mechanism is still sampling, not lookup, and identical input is not guaranteed to produce identical output.

This has been measured directly on genetics questions specifically, not just as a general property of language models. Researchers at the National Human Genome Research Institute compared ChatGPT against 13,642 human responses on 85 multiple-choice genetics questions. On accuracy alone, the result was genuinely impressive: ChatGPT answered 68.2% correctly, against 66.6% for the human respondents — a gap the researchers found was not statistically significant (p = 0.8327). If accuracy were the only thing that mattered here, that result would argue for trusting the tool about as much as you'd trust a knowledgeable person.

Accuracy is not the only thing that matters. The same study asked ChatGPT the identical questions more than once, and on repeated questioning it gave a different answer 16% of the time — for questions it had previously answered correctly and questions it had previously answered incorrectly — and it “gave plausible explanations for both correct and incorrect answers”. Put the two findings together and the honest conclusion is not “ChatGPT is bad at genetics.” It matched human experts on accuracy, in a one-shot test, and still gave a different, equally confident answer when asked again. The lesson is that accuracy and reproducibility are different properties, and a model can score well on the first while failing the second — and a wrong answer delivered with the same fluent, structured confidence as a correct one gives you nothing to distrust in the moment. Genetics has a ground truth: a marker either is or is not in your file, and published evidence about it either exists or it doesn't. A model that generates plausible-sounding text does not have a built-in way to flag which of its sentences are which.

What These Tools Are Actually Good At

None of this makes general-purpose AI chatbots useless for learning about genetics — that would overstate the case just as much as claiming they are reliable for it. The honest picture is narrower and more useful than either extreme.

Studies that test these models on general genetics knowledge — explaining what a gene does, answering conceptual questions about inheritance and testing, walking through genetic counseling guidelines — tend to find respectable performance. One study evaluating ChatGPT 3.5 (“the version that is readily available to the public,” per the authors) on questions adapted from professional genetic-counseling guidelines for gynecologic cancers found it gave correct and comprehensive answers to 82.5% of questions overall, and 100% of the genetic-counseling-guideline questions specifically. A separate study directly comparing GPT-4 to GPT-3.5 on BRCA1, HFE, and MLH1 questions, reviewed by an expert panel of genetic counselors and clinical geneticists, found GPT-4's accuracy had improved significantly over GPT-3.5 — while also noting “notable errors in accuracy remained” and flagging outdated information as an ongoing problem, even in the newer model.

Worth naming plainly: both of those studies, and the accuracy half of the genetics-questions study above, were run on GPT-3.5 and GPT-4 — models from 2023 and 2024. Newer models may do better or worse on the same questions; we haven't tested the current one, and neither has anyone else in a published study we could find. An article about AI reliability that quietly implied its evidence was about whatever model you happen to use today would have the same problem it's describing, so we are naming the versions instead.

Put together, that is a coherent and fair picture: these models can be a genuinely useful way to get a plain-language explanation of a concept, summarize a topic you don't know the vocabulary for yet, or help you draft a sharper question to bring to a clinician. What none of the research above tested — and what is a materially different task — is whether a model can reliably tell you what a specific row in your specific raw data file means, on repeat, without drift. That is the retrieval-and-lookup problem described above, not the explain-a-concept problem these studies measured.

The Privacy Question Worth Asking Before You Upload

A raw DNA file is about as identifying as data gets, and it identifies more than just you — it carries information about blood relatives who never agreed to share it. Uploading it to any third-party chat service, free or paid, sends that file outside of whatever purpose-built, access-controlled environment a dedicated genetics tool would otherwise use.

How that data is retained, whether it is used to train future models, and how long it is kept differs by provider, and often by which plan or tier of that provider you are using — and these terms change over time. We are not going to assert what any specific company currently does with uploaded files, because that is exactly the kind of claim that goes stale and becomes wrong without anyone updating it. Check the provider's current privacy policy and terms of use yourself, specifically for language about training-data use and retention, before you upload a file this personal anywhere.

How DecodeMyBio Does This Differently — and What It Still Doesn't Do

Your Decode result is looked up, not generated. Your star alleles, your metabolizer phenotype, and the guideline that applies to them come from matching your file against curated evidence tables — CPIC and DPWG drug-gene guidelines, PharmVar allele definitions, ClinVar variant classifications. No language model decides what your result is. The assistant that answers your follow-up questions inside Decode+ is a language model, and it is deliberately fenced to the findings that lookup already produced — it can phrase them, it cannot invent them.

We are not going to tell you that makes DecodeMyBio infallible, because it doesn't. Consumer DNA arrays read a subset of the variants that define each star allele; where a defining variant isn't genotyped, the common one is assumed, which means rare alleles can be missed. That is a limitation of consumer raw data generally, not something specific to how we process it, and we have written it up in full — read exactly where the boundaries are on the methodology page and the limitations page before you decide how much weight to put on any tool's output, including ours.

What To Actually Do With the Raw Data File You Already Have

1. Use a general-purpose AI assistant for what it is good at. Ask it to explain a term, summarize a concept, or help you word a question for your doctor. That is squarely within what the research above shows these tools can do reasonably well.
2. Do not treat a specific claim about your file as verified until you have checked it. If a chatbot tells you your file shows a particular variant or that a particular guideline applies to you, that is a claim about a specific row in a specific dataset — exactly the kind of claim generation-without-retrieval is least equipped to get right consistently. Check it against a primary source, or a tool built to look it up rather than generate it.
3. Bring anything medication-related to the person who prescribes for you. That is true regardless of where the claim came from — a chatbot, a lookup tool, or a website. Genetics is one input among several a prescriber weighs, alongside your other medications, your history, and your kidney and liver function.
4. If you want your own file checked against curated guideline data rather than generated text, that is the specific problem Decode+ is built to solve. See how pharmacogenomic analysis from raw DNA data works or what else you can do with a 23andMe or AncestryDNA export.