DataPacts

Scientific Data and Papers as AI Training Data

Checked

At a glance

Scientific content reaches AI developers in two forms that sell very differently: published papers, which publishers license in bulk, and unpublished research records, which company data programs now ask labs and companies for directly.

Independent reference. Every buyer is covered on the same terms, every figure carries its source and check date, and no buyer pays to be listed or placed higher.

Request an independent introduction

Tell us what data your company holds. We reply with the buyers that fit and what their terms leave out. Your details reach a buyer only after you agree to a named introduction.

IndependentYour details stay with us until you agreeNot selling is a valid outcome

Wiley's 10-K for the fiscal year to 2026-04-30 reports AI licensing revenue of USD 33.1 million in Research, while Mercor states, checked 2026-11-09, USD 50,000 to 875,000 to a biotech lab for its records.

Open licenses and research exceptions keep part of the paper market free, and sponsor ownership, database rights and personal data limit what a lab can license. This page sorts scientific content within AI training data licensing by kind, buyer, license term, exception and rights, and not licensing is a valid outcome.

DataPacts Editorial Team gives every figure on this page with its stating party and check date; how we source, date and check every figure sets out the rules.

What Counts as Scientific Data for AI Training

Scientific data for AI training covers five kinds of material: journal articles and books, preprints and open access papers, open research datasets, unpublished experiment and run records, and expert reasoning written for training. The table lists them as of 2026-11-16.

Kind Examples on record Who usually holds the rights Route to AI developers on record Source
Journal articles and research books journal back files, monographs publisher under author and society contracts direct license, collective license Wiley 8-K; copyright.com; cla.co.uk
Preprints and open access papers papers under an open license authors, under the license they chose the license or a research exception creativecommons.org; EUR-Lex
Open research datasets public high-value datasets public bodies and research projects free release under CC0 or CC BY 4.0 where the rule applies EUR-Lex
Unpublished experiment and run records electronic lab notebooks, run and build records, failed runs the lab or company, unless a sponsor or partner owns them company data programs Mercor listing feed
Expert reasoning and protocols scientific reasoning traces, lab protocols, demonstrations the company or the expert who writes them programs and expert contributor work prism.labelbox.com; joinhandshake.com; sieve.ai

Only the first two rows are papers, and the rest are records or data that never passed through a journal. The wider picture of data types AI companies buy places scientific content among the others.

Papers or Lab Records: What Buyer Programs Ask Researchers For

The company data programs on record ask for unpublished records, not papers: Mercor's lab listing wants electronic lab notebooks with 1,000+ runs, failed runs included, and lists data already published as not a fit. The table compares four programs by their own words, all checked 2026-11-09.

Program and listing Wants (as stated) Not a fit (as stated) Stated deal range to the company Source
Handshake AI Data Partnerships "lab protocols and research documentation" among other categories not stated not used here (headline and estimator figures differ) joinhandshake.com
Labelbox Prism "scientific reasoning traces" not stated program-wide: "around $50K" for smaller targeted datasets, "$1M+" for large partnerships prism.labelbox.com
Mercor, Biotech & Life Sciences Lab electronic lab notebooks or structured experiment records the organization owns; 1,000+ runs with inputs and measured outcomes, failed runs included; unpublished data software-log data; data already published; data owned by a sponsor or partner USD 50,000 to 875,000, "depends on volume and depth of records" Mercor listing feed
Mercor, "Chem / Mat Science R&D records: hypotheses, experiments, results and failed work" hypotheses, experiments, results and failed work (listing title) not recorded USD 10,000 to 300,000 Mercor listing feed
Mercor, Publisher or Education Company rights to license for AI training; long-form text, courses or video; hundreds of titles or hours content under someone else's license; content already open access USD 35,000 to 370,000, "depends on catalog size and rights" Mercor listing feed
Mercor, Semiconductor & Materials Company run or build records linked to measured performance; 1,000+ runs; failed builds kept; unpublished data customer- or foundry-owned data; published papers and datasheets only; licensed third-party IP USD 100,000 to 750,000 Mercor listing feed
Sieve partners "educational and scientific demonstrations" among video categories not stated none published sieve.ai

The ranges are program claims and not paid amounts, and they are never averaged or converted to a price per record.

Three Mercor listings exclude open access or published material, so a lab with only papers has no listing that asks for them; the company data partnership programs page sets out the programs and their dated terms.

How Published Papers Reach Training Sets

Published papers reach training sets through three routes: a direct license from the publisher, a collective license that pools many publishers' works, and open access copies used under their license or a legal exception.

Wiley describes its first 2024 project as a "content rights project for training GenAI large language models", and Informa states that its data access agreements give archive access "to help train and improve the relevance of outputs from Large Language Models". Neither publisher names the buyer.

Publisher deals, filings and author consent are covered under academic and scientific publishers licensing to AI. This page keeps only the data-type view: what is licensed, and what stays outside the deals.

On the collective route, the Copyright Clearance Center offers a voluntary, non-exclusive AI Systems Training License, and the Copyright Licensing Agency offers a Generative AI Solution for opted-in books, journals and magazines. Neither publishes fees (checked 2026-11-09).

What Scientific Content Has Earned on Record

Scientific content has public figures from two kinds of source only: listed publishers' filings, which report aggregate AI licensing revenue, and buyer programs' stated deal ranges, which are claims and not paid amounts. The table lists the filed items by date, as of 2026-11-16, with the program claims in the last row.

Date Stating party What Figure as stated Kind Source Target Status
2024-06-13 Wiley first GenAI content rights project, buyer unnamed USD 23 million filing (8-K Ex. 99.1) sec.gov Wiley's AI licensing deals pending
2024-09-05 Wiley second project, buyer unnamed USD 21 million project; Q1 FY2025 USD 16 million (release) vs USD 17 million (presentation), both stated filing sec.gov none n/a
FY2024 Informa (Taylor & Francis) data access revenue, buyers unnamed USD 75m+, "non-recurring" results (RNS) investegate.co.uk Taylor & Francis AI licensing deals pending
FY to 2026-04-30 Wiley AI licensing revenue by segment Research USD 33.1 million; Learning USD 16.0 million 10-K sec.gov none n/a
2026-06-01 Wiley acquisition of Emerald Publishing, partly for "proprietary content advantage" in AI preliminary consideration USD 462.7 million (acquisition, not a license) 10-Q sec.gov none n/a
2026-10-17 to 2026-10-31 Mercor lab, R&D, semiconductor and publisher listings ranges as in the buyer criteria table program claim, checked 2026-11-09 Mercor listing feed none n/a

The filed figures have different bases, so they are not added, and Wiley's separate USD 49 million total has a different base from its two segment values.

No per-paper or per-journal price is on record in a filing, docket or signed license, and the Informa label "non-recurring" says the 2024 revenue was not expected to repeat.

License Terms on Open Research Data and Open Access Papers

Open research data and open access papers come with license terms that decide commercial training use, and some public research data must by law be released free of charge. The table sets four instruments side by side, as of 2026-11-16.

Instrument What it says Effect on scientific data for AI training Source
Implementing Regulation (EU) 2023/138, in force 2023-02-09 listed high-value public datasets free of charge, machine-readable, via API, under CC0, CC BY 4.0 or less restrictive terms; Recital 6: the free-of-charge rule does not apply to libraries, including university libraries, museums and archives public research data in these categories is released, not sold EUR-Lex
Directive (EU) 2019/1024, Recital 55 documents in which third parties hold IP rights are excluded from re-use a public body cannot release a publisher's or author's rights EUR-Lex
Creative Commons, own statement licenses apply only where copyright permission is required; ND content not to be used as training data under a cautious reading where an exception covers training, the CC terms do not decide creativecommons.org, checked 2026-11-09
ODC-BY 1.0 s.2.4; CDLA-Permissive-2.0 s.3.1 ODC-BY covers the database, not its contents; CDLA places no restriction on results such as trained models the license of a research dataset may not cover the papers or images inside it opendatacommons.org; cdla.dev

Mercor's publisher listing treats content already open access as not a fit, so an open license removes a paper from that program.

An audit of more than 1,800 text datasets found license omission above 70 % and errors above 50 % on hosting sites (arXiv 2310.16787, 2023-10-25), which is why the stated license of a dataset needs checking; dataset licenses for AI training explains CC, ODC, CDLA and custom terms.

Research Exceptions: When Mining Scientific Content Needs No License

Research exceptions let research organizations mine papers and data without a license, which leaves a publisher's licensable value mainly in commercial training. This section is information, not legal advice, and the table is as of 2026-11-16.

Jurisdiction and provision Who may mine Commercial use Contract override Source
EU, Directive (EU) 2019/790 Art. 3 (transposition by 2021-06-07) research organizations and cultural heritage institutions, for scientific research no (research purpose) not possible, Art. 7(1) EUR-Lex
EU, Directive (EU) 2019/790 Art. 4 anyone with lawful access, unless the rightholder reserves it; online reservation machine-readable, Art. 4(3) yes Art. 7(1) does not protect Art. 4 EUR-Lex
UK, CDPA 1988 s.29A (in force 2014-06-01) anyone with lawful access, non-commercial research only no unenforceable, s.29A(5) legislation.gov.uk
Germany, UrhG § 60d (implements Art. 3) research purposes; retention with safeguards as long as needed for research or verification, § 60d(5) no per Art. 7(1) gesetze-im-internet.de

The Article 3 exception cannot be switched off by contract, so a publisher cannot license away a research organization's right to mine. The commercial route sits under Article 4, explained in the EU text and data mining exception article together with the opt-out.

Clearing Research Data Before It Is Licensed

Three rights layers decide whether research data can be licensed at all.

Sponsors, Partners and Co-Authors

A lab can license only research data it owns, and data owned by a sponsor or partner is the first exclusion buyers name.

Mercor's lab listing excludes data owned by a sponsor or partner, its semiconductor listing excludes data owned by a customer or foundry partner, and its publisher listing excludes content under someone else's license.

A rights audit before licensing checks funding agreements, collaboration contracts and co-author or society terms first, because one sponsor clause can remove a whole notebook archive from a deal.

Database Right, Trade Secrets and Export Rules in Unpublished Records

Unpublished research records can carry rights that papers do not: a database right in the EU, trade-secret protection while they stay confidential, and export rules for controlled technology.

Directive 96/9/EC Art. 7(1) lets a maker with substantial investment prevent extraction of a substantial part, Art. 7(3) allows licensing, and Art. 10 sets a 15-year term from 1 January after completion, restarted by substantial new investment.

Under Directive (EU) 2016/943 (in force 2016-07-05), Art. 4(4) makes use or disclosure unlawful where the person knew or ought to have known the secret came from someone using it unlawfully.

On the US side, 15 CFR 734.7 takes published information out of the EAR, while 734.13 treats release of controlled technology to a foreign person as a possible deemed export. The EU database right page covers the sui generis right in full.

Personal Data in Research Records

Research records with patient or participant data stay personal data, and calling the work scientific research does not by itself make a license to an AI developer lawful. Health and genetic data are special-category data under GDPR Art. 9(1), applying from 2018-05-25.

The UK Data (Use and Access) Act 2025 s.67 defines scientific research to include research carried out "as a commercial or non-commercial activity", in force 2026-02-05 (S.I. 2026/82).

The EDPB listed Guidelines 1/2026 on scientific research with a consultation from 2026-04-16 to 2026-06-25. Regulation (EU) 2025/327 (EHDS) Art. 53(1) lists training of AI systems among secondary-use purposes, with an opt-out under Art. 71(1); Chapter IV, which holds both, applies from 2029-03-26.

Records are therefore stripped of identifiers before any delivery; see de-identification before licensing for the methods.

Disputes Over Research Content and Research Datasets

Scientific and educational publishers sue AI developers over training copies while licensing to others, and a research dataset built by a university is itself the subject of a copyright suit.

Elsevier, Cengage, Hachette, Macmillan, McGraw Hill and Scott Turow allege in a proposed class action that Meta infringed copyright in training Llama. It was filed 2026-05-05 in the Southern District of New York and transferred on 2026-09-25 to the Northern District of California as 3:26-cv-11415-VC; it is pending with no trial date as of 2026-11-09.

In the second case, EVOX Productions claims copyright infringement over vehicle images in a Stanford research dataset (N.D. Cal., 5:24-cv-09521-NW, filed 2024-12-31). A stipulation on Stanford's first amended answer was filed on 2026-10-13 (Dkt. 90), so the case is past the pleadings stage as of that date.

The EVOX v. Stanford research dataset claims page tracks that case and the related EVOX suits.

Licensing Scientific Data: Benefits, Risks, Downsides and Alternatives, Including Not Licensing

Licensing research records can earn from experiments already run and paid for, but delivered records usually cannot be recalled and the research exceptions keep part of the market free.

Benefits

  • income from experiment records that already exist, at the program ranges stated in listings checked 2026-11-09
  • a use for failed runs, which buyer programs name as wanted
  • aggregate revenue that listed publishers report for papers, as in Wiley's filing

Risks

  • revenue that does not repeat, as Informa labels its 2024 figure
  • trade secrets that lose protection once records leave confidentiality
  • litigation by peers over the same kind of training copies

Downsides

  • research use under Art. 3 and s.29A stays free
  • a training content summary under the EU AI Act shows licensed data only as Yes or No plus modality
  • a sponsor or partner can block records the lab produced

Alternatives

  • a collective route that pools publishers' works
  • open release under a license the holder chooses
  • licensing a subset or a sample only
  • not licensing at all, covered under whether to sell or license your data

Each route depends on who holds the rights and what the buyer's terms say.

Neighboring Data Types and Routes for Research Holders

This page covers scientific content as a data type; books, clinical records, genomic data and course content have their own pages, and so do the holders who license them. The neighbors are:

Holders have their own pages: universities and research groups licensing data, governments and open data, collective licensing for AI training and the tracker of AI copyright lawsuits over training data. A research holder works through four steps:

  1. Read how to sell or license company data for the whole process.
  2. Describe the records with the data manifest, without handing them over.
  3. Check export controls on technical data and EU trade secrets rules for unpublished records.
  4. Compare offers against AI data licensing agreements and their clauses.

Frequently Asked Questions

Are research papers used to train AI models?

Yes: Wiley states its 2024 projects were for training GenAI models, and Informa states its data access agreements help train LLMs. Neither publisher names the buyer.

Can a lab sell its research data to AI companies?

Only data it owns can be licensed. Mercor's lab listing asks for unpublished notebooks with 1,000+ runs and excludes data owned by a sponsor or partner (checked 2026-11-09).

Is open access research free to use for AI training?

It depends on the license and the exception. Creative Commons states its licenses apply only where copyright permission is required, and Directive (EU) 2019/790 Art. 3 covers research organizations.

How much do AI companies pay for scientific data?

No per-paper price is on record. Mercor states USD 50,000 to 875,000 for lab records and USD 35,000 to 370,000 for publishers, checked 2026-11-09, as program claims. Valuation methods are explained under what company data is worth to AI companies.