The Most Private DNA Test Is One Where You Hold the Only Copy of the Data

The most private DNA test is one where the sequencing is performed under a contract that grants the lab no secondary-use rights, the raw files (FASTQ, BAM, VCF) are delivered to you, and no searchable relative-matching database exists on the other end. That is a structural property of the arrangement rather than a feature you can select in a settings panel. Consumer ancestry services are built the other way around: the matching database is the product, your consent is revocable only prospectively, and the company retains both the sample and the data derived from it. If privacy is your binding constraint, then the decision in front of you is not which consumer brand to trust but whether the genome ever enters a third party’s persistent store in a linkable form.

This page is about the mechanics of that decision. It covers what data gets generated and who holds which copy. It also covers what the consent documents permit, how re-identifiable each artifact is, and what you can do with the files once they are in your hands.

The threat model, stated properly

“Privacy” is too coarse a word to reason with, so it helps to separate the adversaries you might be worried about. The defenses against each one are different, and a test that performs well against one can be useless against another.

The first adversary is secondary commercial use, where the company you paid aggregates your data into a research cohort or licenses it to a pharmaceutical partner. This is the most common outcome by volume, and it is usually authorized by the consent you clicked through. Analyses of direct-to-consumer terms have repeatedly found that the documents permit broad use of samples and derived data. Consumers underestimate how far that use extends, and withdrawal of consent typically cannot reach data already shared with partners.¹ Recent legal work argues that the property and contract framework around consumer genetic data is what governs these transfers, and that it favors the holder of the database rather than the person sequenced.²

The second is compelled disclosure and forensic search. A relative-matching database functions as a search index over identity. Forensic genetic genealogy works because a partial match to a third cousin, combined with public records, narrows a candidate pool down to one person. You cannot opt out of your cousins’ decisions. Work on forensic DNA phenotyping and its governance makes exactly this point: the privacy impact of genetic data extends past the person tested to biological relatives who never consented.³

The third is breach and resale after corporate failure. A bankrupt company’s database is an asset like any other, so the relevant question is not the company’s current privacy policy but what a successor in interest may lawfully do with the store.

The fourth is inference you never asked for. A genome supports inferences about ancestry, disease risk, and family structure regardless of what was ordered. Critical work on consumer health kits argues that the interpretive frames these products impose carry their own harms, independent of how the data is handled.⁴

The practical consequence is that you have to ask which adversary a given service protects you from. A service that encrypts data at rest and never sells to pharmaceutical partners may still maintain the relative-matching index that makes forensic search possible.

Genotyping arrays versus sequencing, and why it matters for privacy

The technology a test uses shapes both how sensitive the resulting data is and how useful it will be to you, so it is worth understanding the two main options.

Most consumer kits use single-nucleotide polymorphism arrays, which measure a fixed panel of common variant positions, typically 600,000 to 900,000 markers on an Illumina Global Screening Array. They are cheap, they are sufficient for ancestry and relative matching, and they are also sufficient to re-identify you. Somewhere between thirty and eighty independent common SNPs uniquely identify a person in any population, which means a small subset of an array file works as a fingerprint. Methods designed to flag privacy-sensitive content in genomic files work precisely because sensitive positions are a small, enumerable fraction of the data and can be located at high throughput.⁵

Whole-genome sequencing reads the entire genome instead of a fixed panel. At 30x coverage, meaning each position is read on average thirty times, it generates about 100 GB of FASTQ (the raw sequencing reads). It also produces a 40 to 60 GB CRAM or BAM file aligned to the GRCh38 reference genome and a gVCF listing 4 to 6 million variants. That is far more informative and therefore more sensitive, but it is also far more useful to you. Array files cannot call rare variants, cannot resolve structural variation, and cannot be reanalyzed against next year’s references without re-genotyping. If a copy of your genome is going to exist somewhere, we would rather it be a sequenced one under your own control than an array file sitting inside someone else’s matching index.

The useful asymmetry here is that sequencing does not need a matching database to be valuable. Interpretation runs locally against public resources such as ClinVar, gnomAD v4, PharmGKB, and GTEx, and none of those require you to upload anything.

What “private” looks like in the contract, not the marketing

Privacy claims in marketing copy are rarely specific enough to evaluate, so the document that matters is the contract. Read it for the following clauses, roughly in this order.

Start with sample retention and destruction. Ask whether the biological specimen is destroyed after sequencing and whether you receive written confirmation that it was. A retained sample can be re-assayed with any future technology, so if the answer is that samples are “retained indefinitely for quality purposes,” you should treat the test as permanent.

Next, look at secondary use and de-identification. “De-identified” offers no meaningful protection for a whole genome, because the genome is itself the identifier. Look for language that prohibits secondary use outright rather than permitting it once identifiers have been stripped. Research on genome privacy in study settings is candid that de-identification alone does not prevent re-identification, which is the reason cryptographic and access-controlled approaches exist at all.⁶

Then check relative matching. Does the service maintain such an index, can it be disabled, and does disabling remove you from the index or merely hide the results from your own view? Those are two different systems with very different consequences.

Look closely at deletion semantics as well. Ask what deletion does to derived files, to backups, and to copies already transferred to partners. Withdrawal of consent is almost always prospective only.

Bankruptcy and acquisition deserve their own reading. Look for a clause committing any successor to the same restrictions, and treat the absence of such a clause as the default answer.

Finally, consider the clinical order pathway. A physician-ordered test conducted in a CLIA laboratory, meaning one certified under the Clinical Laboratory Improvement Amendments, sits under HIPAA and under state medical records law. Together those form a stronger regime than a consumer contract. It also keeps a clinician in the loop, which matters for any finding with medical implications. If your sequencing surfaces a variant in a gene such as BRCA2 or LDLR, interpretation and any resulting decision belongs with a genetic counselor or physician rather than with you and a language model.

One further caution about vendors is warranted. Claims about what genetic data can tell you are frequently overstated, and the collapse of BritainsDNA after sustained scientific criticism is a documented case of a testing company making assertions its data could not support.⁷ Privacy claims deserve the same scrutiny as interpretive ones.

Handling your own files

Suppose you now hold a 30x genome delivered as FASTQ plus a CRAM and a gVCF on an encrypted drive. The goal from here is to keep the only copy under your control while still being able to analyze it properly.

Storage comes first. Keep the CRAM rather than the BAM, since a CRAM referenced against GRCh38 is roughly 40 percent smaller. Encrypt the volume with LUKS on Linux or FileVault on macOS, and keep a second encrypted copy offline. For files you need to move between machines, age is simpler to use correctly than GPG:

age-keygen -o key.txt
age -r <recipient-pubkey> -o sample.cram.age sample.cram

Verify integrity on every transfer with sha256sum -c. A truncated CRAM will fail silently in some tools and produce wrong calls in others, so the check is worth the few seconds it takes.

For local analysis, use bcftools wherever you can. To extract a specific gene region without touching the rest of the genome:

bcftools view -r chr13:32315474-32400266 -Oz -o brca2.vcf.gz sample.g.vcf.gz
bcftools csq -f GRCh38.fa -g Homo_sapiens.GRCh38.gff3.gz brca2.vcf.gz

For annotation, run Ensembl VEP or snpEff in a container with the cache downloaded locally. VEP with --offline --cache --assembly GRCh38 makes no network calls at all, which is the entire point of running it that way. Avoid web annotation portals, because pasting a variant list into a browser form is a data transfer, and a handful of positions can be enough to identify you.⁵

Third-party uploads deserve the most thought. Uploading a VCF to an interpretation site is the single largest privacy decision you will make after choosing the lab. If you decide to do it, upload the narrowest possible slice rather than the whole file. Approaches exist for running clinically useful tests without exposing the full genome, including a system that performs pharmacogenomic interpretation on encrypted genotypes so that the interpreting party learns only the result.⁸ Those designs are not yet common in commercial products, but they demonstrate that full disclosure is not technically necessary.

AI agents raise a similar question. If you want an agent to reason over your data, run the model locally or use an API with a no-retention agreement, and feed it derived summaries rather than raw genotypes wherever you can. An agent given a table of annotated variants with allele frequencies and ClinVar assertions performs nearly as well as one given a VCF, and the table is far less identifying.

Questions people also ask

Is AncestryDNA getting sued? Ancestry and 23andMe have both faced privacy litigation and regulatory attention, including class actions following 23andMe’s 2023 credential-stuffing incident that exposed profile and relative-matching data for millions of accounts. Consult current filings rather than any summary, since the posture changes over time. The structural point stands independent of any single case: the legal framework governing consumer genetic data is contract and property law, and it has generally favored the database holder.²

Can you do a DNA test discreetly, or anonymously? You can order sequencing under a pseudonym with prepaid payment and a delivery address that is not your home, and some labs will accept this. Full anonymity is weaker than it sounds, because the genome itself is an identifier and any future relative in a public database narrows you down. Pseudonymity reduces the chance that a breach links your name to your data, but it does not make the data non-identifying.³

Do home DNA tests hold up in court? Generally no for legal purposes such as paternity or immigration, because those require a documented chain of custody with witnessed collection at an accredited laboratory. A consumer kit collected at your kitchen table has no chain of custody. Courts may still admit genetic evidence obtained through other means, which is a separate question from the accuracy of the assay itself.

Does Ancestry sell your DNA to law enforcement? Major services state that they require a warrant or court order, and they publish transparency reports of the requests they receive. The larger exposure is forensic genetic genealogy on databases that permit law-enforcement matching, such as GEDmatch and FamilyTreeDNA, where a distant relative’s upload can identify you without any participation on your part.³

What about HIPAA-compliant consumer testing? HIPAA applies to covered entities such as providers, health plans, and their business associates. A physician-ordered CLIA test is covered by it. A direct-to-consumer kit sold to you is generally not covered, whatever the marketing says, and the governing document is the terms of service you accepted.¹ The broader ethical debate here is well documented. It includes the question of whether consent obtained through a click-through can be meaningful for data this durable.⁹

Oak builds longitudinal molecular profiles of individuals: whole-genome sequencing, RNA sequencing, proteomics, blood biomarkers, and continuous glucose data, integrated into one model of you. Build your profile.

Footnotes

1. Emilia Niemiec, Heidi Carmen Howard. Ethical issues in consumer genome sequencing: Use of consumers’ samples and data. Applied & Translational Genomics, 2016. https://doi.org/10.1016/j.atg.2016.01.005 ↩ ↩²

2. Anya Prince. Genes For Sale!. 2025. https://doi.org/10.2139/ssrn.5357175 ↩ ↩²

3. Nathan Scudder, Dennis McNevin, Sally F. Kelty, et al. Forensic DNA phenotyping: Developing a model privacy impact assessment. Forensic Science International: Genetics, 2018. https://doi.org/10.1016/j.fsigen.2018.03.005 ↩ ↩² ↩³

4. Sav Schlauderaff. Re-Imagining Futurity for Fat, Disabled and “unhealthy” Bodyminds: A Response to 23andMe’s Health + Ancestry Genetic Testing Kits. Fat Studies, 2019. https://doi.org/10.1080/21604851.2019.1651124 ↩

5. Vinicius V. Cogo, Alysson Bessani, Francisco M. Couto, et al. A High-Throughput Method to Detect Privacy-Sensitive Human Genomic Data. Proceedings of the 14th ACM Workshop on Privacy in the Electronic Society, 2015. https://doi.org/10.1145/2808138.2808139 ↩ ↩²

6. Shuang Wang, Xiaoqian Jiang, Dov Fox, et al. Preserving Genome Privacy in Research Studies. Medical Data Privacy Handbook, 2015. https://doi.org/10.1007/978-3-319-23633-9_16 ↩

7. Debbie Kennett, Adrian Timpson, David Balding, et al. The Rise and Fall of BritainsDNA: A Tale of Misleading Claims, Media Manipulation and Threats to Academic Freedom. Genealogy, 2018. https://doi.org/10.3390/genealogy2040047 ↩

8. Paul J. McLaren, Jean Louis Raisaro, Manel Aouri, et al. Privacy-preserving genomic testing in the clinic: a model using HIV treatment. Genetics in Medicine, 2016. https://doi.org/10.1038/gim.2015.167 ↩

9. Salvador Macip, Christopher Willmott. Challenges in Biomedical Areas. Bioethics, 2023. https://doi.org/10.1201/9781003269885-11 ↩