Buyers name these record families on their own pages, and the first limit on each is who owns the records and under what consent they were collected.
Programs are listed alphabetically, never by fee, and every buyer figure is a dated claim, checked 2026-11-09 unless marked, with the Mercor lab listing re-read on 2026-11-17. Sources are buyers' own pages, SEC filings, the Federal Register and statute text. Lab records are one sector of AI training data licensing.
What a Biotech or Life Sciences Lab Holds That AI Buyers Ask For
Life sciences labs hold six record families that AI buyers name, and the one asked for most precisely is the unpublished run record: inputs, measured outcomes and failed runs.
Mercor's lab listing, read 2026-11-17, asks for electronic lab notebooks or structured experiment records the organization owns, with 1,000+ runs and their inputs and measured outcomes, failed runs included, and data that has not been published. The table compares the six families, as of 2026-11-17.
| Record family | Typical contents | Who names it (stated by) | First limit | Source |
|---|---|---|---|---|
| Lab notebooks and run records | electronic lab notebooks, structured run records, inputs and measured outcomes, failed runs | Mercor lab listing; tools named: Benchling, LabArchives, Dotmatics, LIMS | ownership; unpublished data only | Mercor listings feed, research-lab |
| Assay results and analysis files | measured results, analysis files, in-house assay and analysis tools | Mercor (included; nice to have) | sponsor or partner ownership | Mercor listings feed, research-lab |
| Clinical trial documentation | de-identified trial documentation; clinical study reports with patient-level data, tables and analysis plans; Medidata Rave named | Mercor | consent and de-identification; sponsor ownership | Mercor listings feed, research-lab |
| Human omics data | genomic data collected with research consent; NIH's draft lists genomic, epigenomic, proteomic and transcriptomic data | Mercor (genomic only); NIH draft | consent; DOJ bulk thresholds | Mercor listings feed; 90 FR 59131 |
| Medical and microscopy imaging | multimodal medical records; imaging in Gradient Health's list | Labelbox Prism; Gradient Health | IRB approval or qualified de-identification | prism.labelbox.com; gradienthealth.io |
| Protocols, write-ups and reasoning | lab protocols and research documentation; scientists' write-ups; scientific reasoning traces | Handshake AI; Mercor; Labelbox Prism | employee and contractor terms | joinhandshake.com; prism.labelbox.com; Mercor listings feed |
Published papers and the licenses on them are a separate data type. Papers and open research data are covered under scientific data and papers as AI training data.
Buyer Programs That Name Life Sciences Labs
Six programs and listings on record name life sciences, lab records or research organizations on their own pages, and Mercor runs a listing written for biotech and life sciences labs. The table compares them alphabetically, as of 2026-11-17, with each value the stating party's own claim, checked 2026-11-09 unless marked.
| Program | What it names for labs | Threshold stated | Range stated (claim) | Source |
|---|---|---|---|---|
| Gradient Health data partners | research organizations among eligible holders; imaging, EHR, pathology reports, multimodal patient-level data | not stated | revenue share per use; no amount used here | gradienthealth.io/data-partners |
| Handshake AI Data Partnerships | "Pharmaceuticals & life sciences" among 11 industries; lab protocols and research documentation | 20+ full-time employees and 3+ years (license and referral pages); no minimum (Data Partnerships page) | not lab-specific; not used here | joinhandshake.com/ai/data-partnerships |
| Labelbox Prism | scientific reasoning traces; multimodal medical records; IRB approval or qualified de-identification for medical data | no size rule | program-wide, not lab-specific; not used here | prism.labelbox.com/faq |
| Mercor Biotech & Life Sciences Lab listing | notebooks or structured records the organization owns; 1,000+ runs, failed runs included; unpublished data | 10 to 1,000 employees; anywhere | USD 50,000 to 875,000; "Depends on volume and depth of records" (re-read 2026-11-17) | Mercor listings feed, research-lab |
| Mercor Chem / Mat Science R&D records listing | hypotheses, experiments, results and failed work (listing title) | not stated (re-read 2026-11-17) | USD 10,000 to 300,000 | Mercor listings feed, company-rd-research-records |
| Protege data providers | healthcare among listed domains | not stated | revenue share per use; percentage not public | withprotege.ai/data-providers |
Mercor's lab listing names biotech and biopharma, CROs, genomics and diagnostics labs, consumer genomics companies and academic spin-outs as its industries, and it calls biotechs that have wound down and labs with years of archived work a strong fit (read 2026-11-17).
Handshake AI shows two thresholds for the same program, and both appear in the table. BioStack sells healthcare and pre-clinical datasets to biotech companies, so it sells to the sector rather than buying from it.
Each range is a claim, not a paid amount, and ranges are never averaged or ranked. The terms of every company data partnership program sit on the program page.
Who Owns a Lab's Records: Sponsors, Partners, CROs and Spin-outs
A lab can license only records it owns outright, and Mercor lists data owned by a sponsor or partner as not a fit and asks whether the company owns failed runs too. Four situations need a contract check before any offer:
- Sponsor-funded studies: Mercor's listing excludes data owned by a sponsor, so the study agreement decides.
- Partnered programs: the same exclusion covers data owned by a partner, and the partnership agreement decides.
- CRO work for sponsors: Mercor names CROs among its industries and Medidata Rave among its tools, and who owns the trial data a CRO produces is a question for the service agreement; this page states no general rule.
- Spin-out records licensed from a university: the spin-out license decides what the company may license onward; this page states no general rule.
Trade secrets add a second test. Under 18 U.S.C. 1839(3), information is a trade secret only if its owner has taken reasonable measures to keep it secret, and under 1839(5), added by the Defend Trade Secrets Act (in force 2016-05-11), misappropriation includes disclosure by a person who acquired the secret under a duty to maintain secrecy.
The ownership check comes first. Sponsor and partner contracts are the first layer of a rights audit.
Funder Terms on Publicly Funded Lab Data
NIH states in a Federal Register notice of 2025-12-18 (90 FR 59131) that its Data Management and Sharing Policy (DMS Policy) "encompasses all types of scientific data, including genomic data."
Under its proposed Genomic Data Sharing Policy revision, human genomic data not shared within 6 months of data cleaning and quality control follows DMS Policy timing: "publication date or end of the award period, whichever comes first."
Mercor excludes data that has already been published, so records a funder's plan commits to share are read against that line. That reading is a check, not a legal conclusion.
The same notice proposes a draft NIH Controlled-Access Data Policy, with comments closing 2026-03-18 and its final status not checked for this page. In the draft:
- Scope: it applies to NIH-supported research generating human data or deriving data from human data, cell lines or biospecimens, and not to research that only generates and shares non-human data.
- Protected types: genomic, epigenomic, proteomic and transcriptomic data, personal health data and individual level clinical trial data, among others.
- Open sharing: allowed only with informed consent that explicitly states open sharing (plus institutional review) or where federal law requires it.
Ownership terms sit in each award's own documents, so those are read first. Institutional grant and ethics terms are covered for universities and research groups licensing data.
Patient and Participant Data in Lab and Trial Records
Patient-level and genomic records are the part of a lab archive buyers ask about first: Mercor's scoping question reads "Is any of it patient-level or genomic data, and under what consent?" (read 2026-11-17). Labelbox Prism states that medical data needs IRB approval or qualified de-identification (checked 2026-11-09).
Under HIPAA, covered entities and business associates may not sell protected health information without authorization (45 CFR 164.502(a)(5)(ii)), while data de-identified under 45 CFR 164.514(a) to (b) falls outside the Privacy Rule (164.502(d)). The HIPAA routes are set out under medical and clinical data for AI training.
NIH's proposed revision of its Genomic Data Sharing Policy (notice of 2025-12-18) would require that human genomic data from biospecimens or cell lines created or collected after 2015 have consent "consistent with 45 CFR 46.116(d) (the Common Rule)."
It points to 45 CFR 46.102(e) and 164.514(b) for de-identification. This is a proposal, and the page does not restate the Common Rule's text.
GDPR Article 9(1) prohibits processing genetic data and health data unless an Article 9(2) exception applies, and the Regulation applies from 2018-05-25.
The DOJ rule at 28 CFR Part 202, effective 2025-04-08, sets bulk thresholds in any 12 months: more than 1,000 U.S. persons for human 'omic data, more than 100 for human genomic data and more than 10,000 for personal health data.
Selling or licensing bulk data above those thresholds, even de-identified, to buyers in or linked to countries of concern is prohibited, and buyer checks follow the DOJ bulk sensitive data rule.
Life Science Data Deals in Filings
Five transactions in SEC filings involve a life science company supplying data, and only the 2025 Tempus order form with Pathos AI states that the data serves foundation model training. The table lists them by date, as of 2026-11-17, with each value as the filing states it.
| Date | Data supplier | Counterparty | What changes hands | Value as stated | Source |
|---|---|---|---|---|---|
| 2023-10-30 | 23andMe Holding Co. | GSK plc | one-year non-exclusive license to a genetic and phenotypic database, for drug target discovery and research, not AI training | USD 20 million upfront (23andMe 8-K, 2023-10-30) | sec.gov |
| 2025-04-17 | Tempus AI, Inc. | Pathos AI, Inc. | de-identified multimodal oncology dataset, "for the sole purpose of" training a foundation model | USD 200 million in data license fees over three years, incl. USD 50 million upfront (Tempus 10-K FY2025, filed 2026-02-24) | sec.gov |
| 2025-07-14 | 23andMe Holding Co. and debtor subsidiaries | TTAM Research Institute | substantially all assets incl. the genetic database; bankruptcy sale, not an AI-training license | USD 305.0 million cash plus assumed liabilities (23andMe 8-K filed 2025-06-30) | sec.gov |
| 2026-10-14 | Tempus AI, Inc. | Bristol-Myers Squibb Company | de-identified multimodal oncology records plus Lens access and services; use with the Pathos platform allowed | USD 137.5 million aggregate fixed payments through 2031-12-31 (Tempus 10-Q, quarter ended 2026-09-30) | sec.gov |
| 2026-10-22 | Tempus AI, Inc. | Pathos AI, Inc. (sublicense to BMS) | 500,000 de-identified neurology records, 100,000 a year over five years; purpose "to support therapeutic product development", model training not stated | up to USD 55.0 million: USD 50.0 million at USD 100 per delivered record plus USD 5.0 million for updates (Tempus 10-Q, quarter ended 2026-09-30) | sec.gov |
The Tempus filings show three different structures: license fees over three years for the 2025 Pathos order, fixed payments to 2031 for the BMS statement of work, and a per-record price for the neurology order.
The neurology order also carries a side letter limiting third-party licenses until 18 months after the later of delivery and 2027-10-22. Terms are set out under Tempus AI clinical data licensing deals.
The GSK and 23andMe license was a drug-target license, not AI training, and the 23andMe asset sale was an insolvency sale.
Benefits, Risks, Downsides and Alternatives for a Biotech Lab
Licensing can earn income from experiments a lab has already run, including failed ones, but sponsor rights, consent limits and irrevocable license terms carry risks a fee does not offset, and not licensing is a valid outcome.
Benefits
- Failed runs count: Mercor states that failed experiments are as valuable as successful ones (read 2026-11-17).
- Demand is stated across programs: Gradient Health, Handshake AI, Labelbox Prism, Mercor and Protege each name life sciences or lab-adjacent data, as claims checked 2026-11-09.
- A non-exclusive grant leaves the lab free to keep using its records: Protege's provider terms (s.1.1) state a non-exclusive license, while Labelbox Prism states exclusive by default (checked 2026-11-09).
Risks
- Breaching a sponsor, partner or funder term, since Mercor excludes sponsor- or partner-owned data and funder plans can commit records to sharing.
- Trade-secret loss once records leave confidentiality, because the reasonable-measures test of 18 U.S.C. 1839(3) must still be met.
- Bulk-rule exposure when a buyer is in or linked to a country of concern.
- Security incidents at buyers: the California Attorney General report on Mercor lists unauthorized access through compromised LiteLLM versions, with breach dates 2026-03-24 and 2026-03-30, and the lead case is In re Mercor.io Corporation Data Breach Litigation, N.D. Cal. No. 3:26-cv-02821-WHO.
Downsides
- De-identification and consent review take work: Labelbox Prism requires IRB approval or qualified de-identification for medical data.
- Size thresholds exclude smaller archives, such as the 1,000+ runs in Mercor's lab listing.
- Trained models cannot unlearn: Labelbox Prism states that models already trained cannot be unlearned, and micro1's master terms (last updated 2026-11-12, checked 2026-11-17) state a nonexclusive, worldwide, royalty-free, irrevocable, perpetual license (s.1.2) with no duty to delete or retrain (s.4.4).
Alternatives
- License a subset of the archive.
- Share data under the funder's plan instead of licensing it.
- Sell through a wind-down route: Labelbox Prism states that distressed and wound-down companies are eligible.
- Not licensing at all, which is a valid outcome; see alternatives to selling data, including not selling at all.
Where Lab Licensing Meets Neighboring Data Types, Industries and the Seller Path
Biotech labs share patient-record questions with medical practices and the run-record question with materials companies, and the industry index sets out how each sector's limit differs. The index of data licensing by industry also covers medical practices licensing data and materials companies licensing process data.
Omics files follow genetic and genomic data for AI training, and microscopy and slides follow medical imaging and pathology data.
Health and genetic records are special category data (Article 9) in the EU, and confidentiality in the EU runs under the EU Trade Secrets Directive and licensed datasets. HIPAA, export controls and the EU database right also apply to some lab records.
- Check data sale readiness.
- Build the data inventory for an AI data sale.
- Complete PII removal and de-identification of trial and patient records.
- Describe the archive in the data manifest.
- Start evaluating an offer for your data.
- Negotiate the clauses of AI data licensing agreements.
This page is information, not legal advice.
Buyer, Deal and Law Pages Named on This Page
Each program, deal and rule above has its own page with dates and sources.
Buyer profiles:
Deals, cases and trackers:
- AI data licensing deals
- Bankruptcy estates selling data assets (23andMe sale)
- Mercor data breach class actions
Frequently Asked Questions
Can a biotech company license its lab data to AI companies?
Yes, for records it owns outright and has not published. Mercor's lab listing asks for 1,000+ runs, failed runs included, and excludes sponsor- or partner-owned data (checked 2026-11-09, re-read 2026-11-17).
How much do AI companies pay for biotech lab data?
The only lab-specific figure on record is a claim, not a paid amount. Mercor states USD 50,000 to 875,000 for lab records, "Depends on volume and depth of records" (checked 2026-11-09). The wider picture is under what company data is worth to AI companies.
Can a wound-down biotech sell its experiment records?
Yes, through whoever still holds authority over its assets. Mercor calls wound-down biotechs a strong fit and Labelbox Prism states that distressed and wound-down companies are eligible. The routes are under selling data from a shut-down company.
Can scientists get paid to train AI?
Individual scientists can do paid expert tasks for human-data platforms, a separate route from licensing the lab's records. Scientists' tasks are covered in expert contributor work by field.