People now drop their 23andMe or AncestryDNA file into ChatGPT, Claude or Gemini and ask what it says about them. This guide follows the file: how much of it fits into a chat, what the chatbot does with the rest, where its answers come from, and what each company says happens to the file afterwards.
Written by Martin Pechhacker, Vienna Last checked 19 September 2026 19 min read
A raw DNA file is 8 to 9 million tokens, the unit a chatbot counts in. The largest chat window on offer holds 1 million, the free ones about 30,000. No chatbot reads the whole file.
With its code tool, a chatbot searches the file for the markers that you or it name and reads those few lines. What it says about them comes from the text it was trained on, not from a database it opens.
Published benchmarks show the difference: models answering from memory got almost every question about SNP locations wrong, and the same models wired to genomic databases got them right.
ChatGPT, Claude and Gemini may all use what you upload to train their models unless you change a setting or use a temporary chat. None of the three mentions genetic data in the pages read for this guide.
THE SIZE PROBLEM
A chatbot does not count a file in megabytes or lines. It counts in tokens, the fragments of text its model reads, and everything it can keep in view at once has to fit into a fixed number of them, the context window. We counted three public raw DNA files with the tokenizer that OpenAI publishes for its current models. A 23andMe file of 638,000 lines came to 8.07 million tokens, an AncestryDNA file of 677,000 lines to 8.74 million, and an older AncestryDNA file of 701,000 lines to 9.04 million. Sources 1 2 3 4
That is about twice what the usual rule of thumb predicts. OpenAI says "1 token is approximately 4 characters", and Anthropic gives 3.5 characters for English text. A raw DNA file is not English: it is rsids, positions and single letters, and digits split into many small tokens. In the three files one token covered two characters, and one line of the file cost about 13 tokens. Sources 8 28
One raw DNA file against the context windows the three companies state for their plans, to scale. Sources 7 19 30
Set against that, the windows are small. ChatGPT lists 27,000 tokens on its free plan, 54,000 on Go and Plus and 128,000 on Pro, and with its reasoning models 256,000 on Go and Plus and 400,000 on Pro. Gemini lists 32,000 without a plan, 128,000 with AI Plus and 1 million with AI Pro or Ultra. Claude gives its newest models 1 million tokens on every paid plan, the generation before them 500,000, and older models 200,000. Even a million tokens is about one eighth of the file. The free windows hold a third of one percent. Sources 7 19 30
The companies say as much in their own help pages. OpenAI caps uploaded text files: "All text and document files uploaded to a GPT or to a ChatGPT conversation are capped at 2M tokens per file." A raw DNA file is four times that. Claude takes files of up to 500 MB in a chat, far more than the 17 to 19 MB of a raw DNA file, and adds: "Additional token limits may apply based on the length of extracted content." Google puts it most plainly: "The context window is like the reading capacity", and "If you try to upload a file that is too large, Gemini may provide a response that misses connections or details throughout the content." Sources 5 18 29
Since the file cannot be read, it gets searched. ChatGPT and Claude both come with a code tool, a small computer that sits beside the chat. OpenAI describes it like this: "For some data-analysis tasks, ChatGPT writes and runs Python code in a stateful Jupyter notebook environment." Anthropic calls its version "a sandboxed computing environment" and offers it on every plan. The uploaded file lands there. The model never sees its 650,000 lines; it sees the file's name, and it can write a few lines of code to pull things out of it. Sources 6 20
What a chatbot with a code tool does with the file. The other lines are never opened.
So the question decides what gets read. Ask about one marker, say rs4988235, and the code finds that one line and hands back its two letters. Ask what the whole file says about you, and the model has to choose the markers itself, which means the few dozen it remembers most clearly from its training. The other several hundred thousand lines stay unopened, whatever the answer sounds like.
The code tool is cut off from reference data. OpenAI: "The Python environment used for data analysis cannot make external web requests or API calls." Claude's can reach a short list of approved addresses for installing software packages. Neither code tool can query a genomic database while it works through your file. A chatbot can still search the web as a separate step, which serves a handful of markers, not several hundred thousand lines.
Without a code tool, only what fits in the window gets read. Gemini's help page for uploads describes one way of handling a file, reading it into the context window, and with a file this size the window is full long before the file ends. Perplexity says of long files that it "will extract the most important parts".
The private modes can switch the tool off. Claude's incognito chats "open in the previous chat experience, so Claude can't create files or run code in them". The mode that keeps a file out of training is, there, also the mode that cannot search it.
OpenAI's own advice for large files fits the picture: "ask ChatGPT to inspect specific sheets, rows, columns, or sections, or split the file into smaller files." Sources 6 25 29 38
WHERE THE ANSWERS COME FROM
Finding a line is the easy half. The other half is saying what the two letters on it mean, and here a chatbot works from memory: from the papers, wikis and forum posts that were in its training text, up to the day that training ended. For a famous marker that memory is often good. For the long tail it thins out quickly, and a model does not always know where its memory ends.
Researchers have measured exactly this. GeneTuring, a benchmark published in 2025, put 1,600 genomics questions to ten model setups and graded 48,000 answers by hand. One task was as plain as naming the chromosome a given rsid sits on. From memory, the result was poor: "GPT-4o produced incorrect answers to almost all questions related to SNP locations." Letting the model search the web helped with gene names, where its error rate fell from 99 percent to 1 percent. Wiring it to the databases of the United States National Center for Biotechnology Information helped most: that setup "answered all questions correctly" on where an rsid sits, on which gene it belongs to, and on pulling out a stretch of DNA sequence. An earlier study, GeneGPT, found the same gap. A model taught to call those databases averaged 0.83 on eight of the tasks, against 0.44 for a chatbot with web search and 0.12 for the ChatGPT of the time. Sources 39 40
On general genetics knowledge the picture is kinder. On 85 exam-style questions ChatGPT answered 68.2 percent correctly against 66.6 percent for human respondents, and it changed its answer on 16 percent of questions when asked a second time. On questions about genes and medication, the authors of a 2025 benchmark found the newest models far ahead of their predecessors and still short of what medical use would need. References deserve a second look as well: in one count, 55 percent of the citations an older model produced did not exist, and 18 percent of those from its successor. Sources 41 42 43
The chart looked great, and I would have believed it if I hadn't checked the rs ids.
She found rsids assigned to genes they do not belong to, and a cited study that did not exist when she searched for its title. These are the two errors the benchmarks predict, and neither shows in the answer itself. The models have changed since; the mechanism has not. Source 46
WHAT A REPORT IS BUILT FROM
This is a list of inputs, not a verdict on anyone's model. A report on a raw DNA file rests on reference data that has to be opened, versioned and applied to every line. A general chatbot brings none of it to the chat unless someone wires it in. Source 57
A dated ClinVar copy
The public archive of what laboratories have reported about individual variants changes every week. A report names the copy it used. A model remembers the archive as it stood somewhere in its training text, without a date.
Score weights
A polygenic score adds up thousands of positions, each with a published weight from the PGS Catalog. The weights are files of numbers, not knowledge, and no model has them memorized.
Reference frequencies
Whether a genotype is common or rare depends on the population it is compared with. That takes frequency tables for each ancestry group, looked up position by position.
A star-allele caller
Medication-response genes are named by combinations of positions, the star alleles. Working them out from single lines is a program's job, with defined rules for missing positions.
Then there is the file itself, which a chatbot takes at face value. The letters are given on the plus strand of build 37, while many papers and older wiki pages quote the other strand, so an A in your file can be the T of the article. A -- or a 0 is a position the chip could not read, not a result. Our guides to the 23andMe file and the AncestryDNA file go through strand, build and unread positions line by line. Sources 58 59
And chips misread rare variants far more often than common ones. In a study of nearly 50,000 people, common variants were read with over 99 percent agreement, but for the rarest, "only 16% of 4757 heterozygous genotypes from the SNP chips were confirmed with sequencing data". An earlier laboratory study of raw data found that "40% of variants in a variety of genes reported in DTC raw data were false positives." A striking line in a raw file is first of all a candidate for a misread, and an answer built on it inherits that. Sources 44 45
THE OTHER SIDE
None of this makes a chatbot useless next to a DNA file. It makes it a different tool from the one people hope for. Used for what it does well, it is the most patient genetics tutor most people will ever have.
Explaining the vocabulary. What an allele is, what heterozygous means, why a percentile is not a probability. General knowledge is where the benchmarks show models at their strongest.
Reading a paper with you. Paste the abstract or upload the PDF and ask what was measured, in whom, and how large the effect was. The text is in front of the model, so it is reading and not remembering.
Writing the lookup script. Finding fifty rsids in a text file is a ten-line program. A chatbot writes it in seconds, and it runs on your own computer, where the file never has to leave.
Checking one marker at a time. Name the rsid, have the code tool print the exact line it found, and ask for the source of every claim. Then open the source.
A request that fits the mechanics
It names the markers, asks for proof of the lookup, and keeps what the file says apart from what the model remembers:
"Use your code tool to find these rsids in the attached file and print each matching line exactly as it is. For each one, tell me which strand and genome build the file uses, and whether the studies you know quote the same strand. Give a source I can open for every statement, and say so when you are not sure."
An uploaded file is content like any other message, and each company's ordinary rules for content apply to it. They differ in the details and agree on the outline: unless you change a setting, the company may use what you upload to improve its models; a temporary mode keeps a chat out of training; and deleting is not instant.
When you use our services for individuals such as ChatGPT and Codex, we may use your content to train our models.
ChatGPT, the start screen without an account, captured 19 September 2026.
Files are included: OpenAI lists "images and files" among the content it may use. The setting that turns this off is called "Improve the model for everyone", under Data Controls, and a Temporary Chat "won't be used to improve our models". A limited number of OpenAI staff and contractors may see content for a handful of stated reasons, and its advice is direct: "Please do not enter sensitive information that you would not want reviewed or used." Chats stay until you delete them and are then removed within 30 days. Two details matter for a file. "Chats and Library files are managed separately: deleting a chat does not delete files saved to Library." And with memory switched on, files can feed it: "Sensitive information may appear in memory if you share it with ChatGPT." Sources 9 10 11 12 13
We may use Materials to provide, maintain, and improve the Services and to develop other products and services, including training our models, unless you opt out of training through your account settings.
Claude, the sign-in page, captured 19 September 2026.
Anthropic made this a choice in 2025: new users pick the setting when they sign up, and it sits under Privacy in the settings as "Help improve our AI models". The choice changes how long data is kept. With it on, conversations may be kept "for up to 5 years" in training pipelines; with it off, the earlier 30-day retention applies. A deleted chat leaves the back-end "within 30 days", and chats flagged by its safety systems are kept for up to two years. Incognito chats "are not used to improve Claude", though as noted above they cannot run code either. On sensitive material Anthropic recommends "being thoughtful about sharing highly sensitive personal details", and names "Health records or medical information" among them. Sources 24 22 23 25 26
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.
Gemini, the start screen without an account, captured 19 September 2026.
Google is the most explicit about people reading chats. The setting is called Keep Activity, and "If you're 18 or over, Keep Activity is on by default." With it on, Google uses activity to improve its services, "including training generative AI models", and since August 2025 "a sample of your future uploads will be used". Human reviewers read a subset of chats, and "Reviewed chats are retained for up to three years", even if you delete your activity. Activity is kept for 18 months unless you choose 3 months, 36 months or no limit. Temporary chats are not used for training and are kept for 72 hours. Sources 31 32 33
USED FOR TRAINING
THE SWITCH
THE PRIVATE MODE
Sources 9 11 12
Yes, files included
"Improve the model for everyone", off
Temporary Chat, removed within 30 days
Sources 21 22 25
Yes, unless you opted out
"Help improve our AI models", off
Incognito chat, kept 30 days, without the code tool
Sources 31 32 33
Yes, a sample, with human review
Keep Activity, off
Temporary chat, kept 72 hours
Consumer accounts with their default settings, as each company states it on 19 September 2026. Business, work and school accounts come under separate terms at all three.
The smaller names differ more. Microsoft says of its current Copilot app that "Prompts, responses, and your file contents" are not used to train foundation models, and takes files up to 50 MB. The company behind Grok says it "may use your content and interactions with Grok" to train its models unless the setting "Improve the model" is switched off, and asks users not to share "any personal or sensitive information". DeepSeek stores data in the People's Republic of China, and its privacy policy is the only one read for this guide that names genetic data at all: "We do not ask for, and you should not provide sensitive Personal Data to the Services". Sources 34 35 37 36
NEW IN 2026
Both OpenAI and Anthropic launched health features in 2026, for users in the United States: a place to connect medical records and wellness apps to the chatbot. Both attach a stronger promise to them. OpenAI says connected records and the conversations that use them "are not used to train our foundation models". Anthropic says: "We do not use users' health data to train models." Sources 15 27
Each promise is tied to the data that comes in through those connections. OpenAI spells out the boundary: "Conversations that do not use this data follow your standard ChatGPT model-training setting choices". A raw DNA file attached to an ordinary chat is an ordinary upload. None of the health pages read for this guide says a word about genetic files. Source 15
THE LEGAL FRAME
The American health privacy rules do not follow the file. HIPAA binds "covered entities": health plans, clearinghouses and health care providers. A chatbot company is none of these, so a file you upload yourself is outside it.
The American law against genetic discrimination has gaps. The Genetic Information Nondiscrimination Act covers health insurance and employment. In the words of the National Human Genome Research Institute, its protections "do not cover long-term care insurance, life insurance, or disability insurance".
European law puts genetic data in its strictest class. Article 9 of the General Data Protection Regulation lists genetic data among the special categories, and the chatbot companies are bound by it for their European users. A regulator has already had reason to warn: after medical data was typed into a chatbot at a doctor's practice, the Dutch authority noted that "Most companies behind the chatbots store all data entered."
Deletion has limits of its own. In 2025, in the lawsuit The New York Times brought against OpenAI, a court order obliged the company to retain consumer ChatGPT content indefinitely, against its own 30-day deletion rule. OpenAI reports that its obligations "ended on September 26, 2025", and that a limited set of user data from April to September 2025 stays stored under lock for that case. A promise to delete holds until a court says otherwise. Sources 47 48 49 50 17
The larger exposure so far has not come from training but from sharing. In July 2025, ChatGPT conversations that users had shared by link turned up in Google; OpenAI removed the option, its security chief saying it "introduced too many opportunities for folks to accidentally share things they didn't intend to". A month later the same was reported for shared Grok chats. In Meta's AI app, people posted their conversations to a public feed without noticing. And in February 2026 an app that passes chats on to the large models left 300 million messages readable in an open database. A share link is a public web page, and an app in between is one more company holding the chat. Sources 51 52 53 54
Among ChatGPT's custom GPTs there are ones made for DNA files. They run on the same model with the same window, plus instructions from their builder. OpenAI says builders cannot see your conversations, and adds a caution for GPTs that call outside services: "relevant parts of your input may be sent to the third-party service", and "OpenAI does not audit or control how those services use or store your data." Source 16
The more interesting tools run on the other side. A growing set of open-source scripts reads the file on your own computer and hands the chatbot only what a question needs. One of them describes its approach in two sentences: "Your raw file has 640,000 variants. A distillate has twelve." That is the mechanics of this page turned into a design: search locally, send little. Source 55
ON THIS SITE
ReadYourDNA: the variant explorer on the demo genome. The reports are computed first; the assistant answers from them. Open the demo
ReadYourDNA has an optional assistant for paying accounts, and it is built the other way round. The file is read once by the pipeline described on the methodology page, position by position, against dated reference data. The assistant never sees the raw file. It answers from the finished reports and fetches the parts a question needs. The privacy policy states what that involves: the parts of your results needed for an answer are sent to Google's Gemini API and are not used to train Google's models, and if you never switch the assistant on, nothing of yours reaches a model. Other services built to read the file are listed in Where a raw DNA file can still be uploaded in 2026. Sources 56 57 60
ReadYourDNA reads a raw DNA file you already have and returns polygenic trait scores, medication-response annotation, carrier annotation and an annotated variant explorer, with the study behind every number. One payment, no subscription, and you can delete the file at any time.
Educational, not medical advice.
It can search the file, not read it. A raw DNA file is 8 to 9 million tokens and ChatGPT's largest window is 400,000, so its code tool looks up the markers that you or it name. What it says about them comes from its training text, not from a genomic database.
OpenAI's own pages give the facts to decide with. Unless you turn off "Improve the model for everyone" or use a Temporary Chat, uploaded files may be used to train its models; authorized staff may review content; deleted chats are removed within 30 days, and files saved to the Library stay until you delete them separately. OpenAI's advice: "Please do not enter sensitive information that you would not want reviewed or used."
Yes, both accept a text file of that size: Claude up to 500 MB per file in a chat, Gemini up to 100 MB. Neither can hold it in view. Claude searches it with its code tool; Gemini's help page describes reading what fits into the context window, which for a file this size is a small part of it.
No, for two reasons. The file is a list of a few hundred thousand positions, not a genome. And of that list a chatbot looks at the lines a question points to, usually a handful.
One that fits how the tool works: name the rsids you want, ask the code tool to print each matching line exactly as it stands in the file, ask which strand and build the file uses, and ask for a source you can open for every statement. The section above has a full example.
It keeps the chat out of model training at all three companies. It does not mean nothing is stored: ChatGPT removes temporary chats within 30 days, Claude keeps incognito chats for 30 days, Gemini for 72 hours. In Claude, incognito chats also cannot run the code tool that a file this size needs.
Language models pick their words with some randomness, so the same question can produce a different answer. In one study ChatGPT changed its answer on 16 percent of genetics questions when asked again. A lookup in a database returns the same result every time; a memory does not have to.
No. HIPAA applies to health plans, clearinghouses and health care providers. A chatbot company is none of them, so the company's own terms and general consumer and privacy law are what apply.
Every figure on this page was read from the page named, on the date named. ReadYourDNA publishes this page. If something here is out of date, write to support@readyourdna.com.
Personal Genome Project participant hu005023, public 23andMe data file (2018)
my.pgp-hms.org/profile/hu005023
One of the three public files whose size was counted: 638,488 lines, 16.8 MB, 8,072,344 tokens.
Read on 19 September 2026
Personal Genome Project participant hu25E03E, public AncestryDNA data file (header of November 2025)
my.pgp-hms.org/profile/hu25E03E
The second counted file: 677,455 lines, 18.3 MB, 8,739,215 tokens.
Personal Genome Project participant hu0488C7, public AncestryDNA data file (2013, the earlier array)
my.pgp-hms.org/profile/hu0488C7
The third counted file: 701,495 lines, 18.8 MB, 9,035,490 tokens.
OpenAI, tiktoken (the tokenizer library; encoding o200k_base)
The tool the three files were counted with on 19 September 2026.
OpenAI Help Center, "File Uploads FAQ" (archived copy of 4 September 2026)
The size limit per file, the token cap for text files, the limit for spreadsheets and the daily limit on the free plan.
All text and document files uploaded to a GPT or to a ChatGPT conversation are capped at 2M tokens per file.
All files uploaded to a GPT or a ChatGPT conversation have a hard limit of 512MB per file.
OpenAI Help Center, "Data analysis with ChatGPT" (archived copy of 15 September 2026)
That ChatGPT writes and runs code on uploaded data, that this code cannot reach the internet, and what OpenAI suggests for files too large to analyze whole.
For some data-analysis tasks, ChatGPT writes and runs Python code in a stateful Jupyter notebook environment.
The Python environment used for data analysis cannot make external web requests or API calls.
ChatGPT, "Pricing" (archived copy of 19 September 2026)
The context window of each plan: 27K, 54K, 54K and 128K, and up to 400K when reasoning.
OpenAI Help Center, "Understanding and counting tokens" (archived copy of 17 September 2026)
The rule of thumb for ordinary English text.
1 token is approximately 4 characters.
OpenAI Help Center, "How your data is used to improve model performance" (archived copy of 17 September 2026)
That content from individual accounts may be used for training, the name of the setting that turns this off, and that business products are not used for training by default.
OpenAI Help Center, "Data Usage for Consumer Services FAQ" (archived copy of 17 September 2026)
That files count as content, who at OpenAI may see it, and its advice on sensitive information.
Please do not enter sensitive information that you would not want reviewed or used.
OpenAI Help Center, "Temporary Chat FAQ" (archived copy of 17 September 2026)
That a Temporary Chat is not used for training.
OpenAI Help Center, "Chat and File Retention Policies in ChatGPT" (archived copy of 15 September 2026)
How long chats and files are kept, what deletion does, and that files in the Library outlive the chat.
Chats and Library files are managed separately: deleting a chat does not delete files saved to Library.
OpenAI Help Center, "Memory FAQ" (archived copy of 17 September 2026)
That memory can draw on files, and OpenAI's note on sensitive information in memory.
Sensitive information may appear in memory if you share it with ChatGPT.
OpenAI, "Usage policies" (effective 29 October 2025; archived copy of 7 September 2026)
The rule on advice that requires a license.
OpenAI, "Launching Health in ChatGPT" (23 July 2026; archived copy of 5 September 2026)
What the health feature promises for connected records, and that other conversations follow the ordinary setting.
Conversations that do not use this data follow your standard ChatGPT model-training setting choices
OpenAI Help Center, "GPTs in ChatGPT" (archived copy of 3 September 2026)
What a custom GPT can pass on to outside services.
When you interact with a GPT that uses apps or external APIs, relevant parts of your input may be sent to the third-party service.
OpenAI does not audit or control how those services use or store your data.
OpenAI, "How we're responding to The New York Times' data demands" (updated 22 October 2025; archived copy)
The court order to retain consumer ChatGPT content indefinitely, when it ended, and what is still stored.
Our obligations under the earlier order ended on September 26, 2025.
Claude Help Center, "Upload files to Claude" (updated 23 July 2026)
Accepted file types, the size and number of files, and the note on token limits.
Note: Additional token limits may apply based on the length of extracted content.
Claude Help Center, "How large is the context window on paid Claude plans?"
The context window on paid plans: 1 million tokens for the newest models, 500,000 for the generation before, 200,000 otherwise.
Claude Help Center, "Create and edit files with Claude" (updated 6 August 2026)
The code tool: who has it, that it is a sandbox, and which outside addresses it may reach.
Code execution and file creation gives Claude a sandboxed computing environment.
Anthropic, "Consumer Terms of Service" (effective 8 October 2025)
www.anthropic.com/legal/consumer-terms
The training clause for consumer accounts.
Anthropic Privacy Center, "Is my data used for model training?" (updated 16 March 2026)
That whole conversations with their content are covered, and that incognito chats are not used.
Your Incognito chats are not used to improve Claude, even if you have enabled Model Improvement in your Privacy Settings.
Anthropic Privacy Center, "How long do you store my data?" (updated 1 July 2026)
Retention with the training setting on and off, after deletion, and for flagged chats.
Anthropic, "Updates to Consumer Terms and Privacy Policy" (28 August 2025)
The 30-day retention without training, and that new users choose at signup.
Claude Help Center, "Use incognito chats"
How long incognito chats are kept, and that they cannot run code.
incognito chats open in the previous chat experience, so Claude can't create files or run code in them
Claude Help Center, "I would like to input sensitive data into my chats with Claude. Who can view my conversations?" (updated 22 May 2026)
Anthropic's advice on highly sensitive details.
Anthropic, "Advancing Claude in healthcare and the life sciences" (11 January 2026)
The health-record connections for subscribers in the United States and the training promise attached to them.
We do not use users' health data to train models.
Claude Platform Docs, "Glossary"
Anthropic's rule of thumb for characters per token.
Gemini Apps Help, "Upload & analyze files in Gemini Apps"
File size and number, and what Google says happens when a file is larger than the context window.
The context window is like the reading capacity.
If you try to upload a file that is too large, Gemini may provide a response that misses connections or details throughout the content.
Gemini Apps Help, "Gemini Apps limits & upgrades for Google AI subscribers"
The context window by plan: 32k, 128k and 1 million tokens.
Gemini Apps Help, "Gemini Apps Privacy Hub" (last updated 10 August 2026)
Training, human review, retention, temporary chats, and Google's advice on confidential information.
Reviewed chats are retained for up to three years.
Gemini Apps Help, "Manage & delete your activity in Gemini Apps"
That the setting Keep Activity is on by default for adults.
If you're 18 or over, Keep Activity is on by default.
Google, The Keyword, on temporary chats and privacy controls in Gemini (13 August 2025)
That a sample of uploads is used to improve Google services.
a sample of your future uploads will be used to help improve Google services for everyone
Microsoft Support, "File upload in Microsoft Copilot"
The 50 MB limit and the accepted text formats.
Microsoft Support, "Microsoft Copilot for individuals: your activity history"
The training statement for the current Copilot app.
Prompts, responses, and your file contents when using the Microsoft Copilot app aren't used to train foundation models.
DeepSeek, "Privacy Policy" (last update 10 February 2026)
The one policy read for this page that names genetic data, and where DeepSeek stores data.
We do not ask for, and you should not provide sensitive Personal Data to the Services
SpaceXAI (x.ai), "Consumer FAQs" for Grok (archived copy of 17 September 2026)
Training on Grok conversations, the setting that controls it, and the advice on sensitive information.
To protect your privacy, please do not share any personal or sensitive information in your questions to Grok.
Perplexity Help Center, "File Uploads" (archived copy of 22 July 2026)
The 40 MB limit and what happens to a long file.
If you upload a long file, Perplexity will extract the most important parts to provide the most relevant response to your query.
Shang X et al., Briefings in Bioinformatics 2025;26(5):bbaf492, "Benchmarking large language models for genomic knowledge with GeneTuring"
1,600 genomics questions put to ten model setups; answers from memory against answers with database tools.
GPT-4o produced incorrect answers to almost all questions related to SNP locations.
SeqSnap answered all questions correctly in the SNP location
Jin Q et al., Bioinformatics 2024;40(2):btae075, "GeneGPT: augmenting large language models with domain tools for improved access to biomedical information"
pubmed.ncbi.nlm.nih.gov/38341654
The same comparison a year earlier: a model taught to call NCBI's databases against chatbots answering from memory.
Duong D, Solomon BD, European Journal of Human Genetics 2024;32(4):466-468, "Analysis of large-language model versus human performance for genetics questions"
pubmed.ncbi.nlm.nih.gov/37246194
85 genetics questions: the share answered correctly by ChatGPT and by people, and how often a repeated question got a different answer.
Keat K et al., Pacific Symposium on Biocomputing 2025;30:229-246, "PGxQA: A Resource for Evaluating LLM Performance for Pharmacogenomic QA Tasks"
pubmed.ncbi.nlm.nih.gov/39670373
A benchmark of questions on genes and medication.
Walters WH, Wilder EI, Scientific Reports 2023;13:14045, "Fabrication and errors in the bibliographic citations generated by ChatGPT"
pubmed.ncbi.nlm.nih.gov/37679503
How many of the references two model generations produced did not exist.
55% of the GPT-3.5 citations but just 18% of the GPT-4 citations are fabricated
Weedon MN et al., BMJ 2021;372:n214 (genotyping chips and very rare variants, 49,908 UK Biobank participants)
pubmed.ncbi.nlm.nih.gov/33589468
How well a genotyping chip reads common and very rare variants, checked against sequencing.
only 16% of 4757 heterozygous genotypes from the SNP chips were confirmed with sequencing data
Tandy-Connor S et al., Genetics in Medicine 2018;20(12):1515-1521 (raw data findings re-examined in 49 samples)
pubmed.ncbi.nlm.nih.gov/29565420
How many variants reported in raw data held up when a laboratory looked again.
40% of variants in a variety of genes reported in DTC raw data were false positives.
Debbie Moon, Genetic Lifehacks newsletter, "Sleep Genes Update + Uploading Genetic Data to AI" (16 May 2025)
geneticlifehacks.substack.com/…
A genetics writer's own trial of four chatbots with a section of her file.
The first study didn't even exist when I searched for it by title.
Autoriteit Persoonsgegevens (Dutch data protection authority), "Caution: use of AI chatbot may lead to data breaches" (6 August 2024)
www.autoriteitpersoonsgegevens.nl/…
A regulator's warning after medical data was typed into a chatbot.
Most companies behind the chatbots store all data entered.
Malwarebytes, "OpenAI kills short-lived experiment where ChatGPT chats could be found on Google" (1 August 2025)
Shared ChatGPT conversations appearing in Google, with the statement of OpenAI's security chief.
this feature introduced too many opportunities for folks to accidentally share things they didn't intend to
TechCrunch, "Thousands of Grok chats are now searchable on Google" (20 August 2025)
The same with shared Grok conversations.
TechCrunch, "The Meta AI app is a privacy disaster" (12 June 2025)
People publishing their conversations in a public feed without noticing.
Malwarebytes, "AI chat app leak exposes 300 million messages tied to 25 million users" (9 February 2026)
A leak at an app that passes chats on to the large models.
DNAdistiller, an open-source script (GitHub)
github.com/lucashuizinga/DNAdistiller
An example of a tool that reduces the file on your own computer before anything reaches a chatbot.
Your raw file has 640,000 variants. A distillate has twelve.
ReadYourDNA privacy policy
What the optional genome assistant sends to a model, and that nothing is used for training.
ReadYourDNA methodology
The reference data a report is built from.
What is inside a 23andMe raw data file
readyourdna.com/guides/23andme-raw-data-file-format
The layout, the strand and the build of a 23andMe file.
What is inside an AncestryDNA raw data file
readyourdna.com/guides/ancestrydna-raw-data
The layout of an AncestryDNA file.
Where a raw DNA file can still be uploaded in 2026
readyourdna.com/guides/where-you-can-still-upload-a-raw-dna-file
Services built to read the file.
United States, 45 CFR 160.103, "Definitions" (the HIPAA rules)
Who counts as a covered entity: health plans, clearinghouses and health care providers.
National Human Genome Research Institute, "Genetic Discrimination" (its page on the act known as GINA)
What the act covers and the three kinds of insurance it leaves out.
GINA's health insurance protections do not cover long-term care insurance, life insurance, or disability insurance
Regulation (EU) 2016/679, the General Data Protection Regulation, Article 9
Genetic data as a special category of personal data.
Raw DNA Data Reports for 23andMe & AncestryDNA · About ReadYourDNA · FAQ · Methodology & Science · Pricing · Support and complaints · Legal & Contact · Health Traits · Pharmacogenomics · Carrier Status · Variant Explorer · The Human Karyotype · Privacy Policy · Terms of Service · Accessibility · Consumer Health Data Privacy · Your Privacy Choices · ReadYourDNA and Promethease compared · What is inside a 23andMe raw data file · What is inside an AncestryDNA raw data file · What a chatbot can and cannot do with your raw DNA file · Where a raw DNA file can still be uploaded in 2026
Written, maintained and operated by Martin Pechhacker, sole proprietor, Vienna, Austria. Trade register GISA 40064420. support@readyourdna.com