Alternatives to Selling Data to AI Labs, Including Not Selling at All

A company that does not want to license its data to an AI lab can keep it, use it in its own models, release it openly, share it with researchers, license a narrower slice, or delete it, and not licensing at all is a valid outcome. The paths keep or give up different rights, and only some can be undone later. A license grant is often irrevocable, as in micro1's master terms, and Labelbox states that "Models already trained on the data cannot be unlearned" (checked 2026-11-09).

Each path is one outcome of AI training data licensing decisions, and open release and deletion are the hardest to reverse. DataPacts takes no cut of any sale.

DataPacts Editorial Team checked every figure on this page against its primary source on November 9, 2026; the page on how we source, date and check every figure sets out the rules.

Six Paths Other Than a Full License, Compared

The six paths differ in what the company keeps, what it risks and whether the choice can be reversed later. The table compares them as of 2026-11-09.

Keeping the Data and Not Licensing It at All

Keeping the data means doing nothing with it for AI buyers, which leaves every right, every later option and every confidentiality protection with the company.

• Benefit: no grant exists to recall, and trade-secret protection continues while the company takes reasonable secrecy steps (Directive (EU) 2016/943, in force 2016-07-05).
• Risk: none from a buyer; the internal duties to protect the data stay.
• Downside: no license income.

Licensing revenue is uneven, so the income forgone is uncertain too. Reddit's 10-Q for the quarter ended 2026-09-30 states that substantially all of its licensing contract value comes from two partners, not named, whose agreements are "approaching the end of their respective current terms". Shutterstock's Data, Distribution, and Services revenue was USD 15.8 million in Q3 2026 against USD 65.7 million a year earlier, down 76 %, and Wiley's AI license revenue was USD 13.7 million in Q1 FY2027 against USD 28.9 million (all from 10-Qs). Keeping the data now does not rule out a license later, since no grant has been made.

Keeping the Data for the Company's Own Models

A company can keep its data out of any license and use it to train or adapt models of its own, as AI developers do with their users' data.

• Benefit: the data's value stays inside the company, and no one else's model learns from it.
• Risk: the privacy exposure is the same as under a license, because the EDPB's Opinion 28/2024 of 2024-12-17 finds that AI models trained on personal data cannot in all cases be considered anonymous, and the FTC ordered Everalbum on 2021-05-07 to delete models and algorithms developed using users' photos and videos.
• Downside: the company carries the cost and the legal duties itself; no public figure exists for that cost.

SpaceXAI, formerly xAI, states in its Grok 4.7 training summary of 2026-09-21 that user interaction data was used. A provider that places a general-purpose AI model on the EU market owes the duties of Article 53 of Regulation (EU) 2024/1689, which apply from 2025-08-02. The open-source exemption in Article 53(2) covers only points (a) and (b), so the copyright policy and the training summary still apply. This page does not explain how to train a model. The duties are set out under the EU AI Act Article 53 training data summary.

Releasing the Data Openly

Releasing data openly means publishing it under an open license or none, so anyone, including every AI developer, may use it without paying.

• Benefit: public reach, citation and community use.
• Risk: the release cannot be recalled, and open licenses do not control uses that need no permission.
• Downside: buyers stop paying for what is already public.

Creative Commons states that its licenses apply only where copyright permission is required (checked 2026-11-09). Common Crawl states that its WARC files are immutable and that affected URLs are only removed or filtered from subsequent crawls and public indexes (2025-11-04). Common Crawl also states, citing a 2024 Mozilla Foundation study, that at least 64 % of 47 LLMs published from 2019 to 2023 used it (checked 2026-11-09). Mercor states that its software listing wants "Private repositories that have never been public, mirrored or open-sourced" and does not fit "Forks or copies of open-source projects", with USD 50,000 to 500,000 to the company (checked 2026-11-09). Release terms are compared under dataset licenses for AI training such as CC, ODC and CDLA.

Sharing the Data With Researchers Under Research-Only Terms

Sharing data with a university or research group under research-only terms keeps commercial training rights with the company while letting the data be studied.

• Benefit: scholarly use without a commercial grant.
• Risk: research outputs and released subsets can travel further than the agreement intends; DCLM-Baseline, for example, is public under CC BY 4.0 with "research purposes only" on its dataset card (checked 2026-11-09).
• Downside: little or no income, and no public terms of company-to-university AI data agreements exist to compare.

The Institutional Data Initiative, a research initiative at Harvard Law School Library, open-sourced Institutional Books 1.0 on 2025-06-12: 983,004 volumes of 1,075,899 scanned. On Hugging Face the set carries noncommercial early-access terms, and commercial use including AI model training requires contacting the initiative (checked 2026-11-09). Once data is lawfully in a research partner's hands, statutory exceptions apply whatever the agreement says: Article 3 of Directive (EU) 2019/790 lets research organisations and cultural heritage institutions mine lawfully accessed content for scientific research, with no opt-out and no override by contract, and section 29A of the UK Copyright, Designs and Patents Act 1988 allows non-commercial research and makes contract overrides unenforceable. Research arrangements are covered for universities and research groups licensing data to AI companies.

Licensing Less: Collectives, Crawler Controls and Narrower Terms

A company that wants some income without a full grant can license less: through a collective, by setting terms for crawlers on its public content, or by keeping exclusivity and scope narrow.

• Benefit: income or control with a smaller grant.
• Risk: collective and crawler payouts are not published.
• Downside: it is still a license, so the grant and trained-model questions remain.

The four routes are listed alphabetically and compared as of 2026-11-09.

Cloudflare blocks AI crawlers by default for new domains since 2025-07-01, and on 2026-09-15 it announced separate Search, AI training and AI agents controls; it states that 17 % of its sites restrict AI training. Article 4(3) of Directive (EU) 2019/790 requires a reservation for online content to be machine-readable. License, sale, revenue share and data-as-a-service as models are compared on a sibling page. The controls to allow, block or charge AI crawlers are covered separately.

Deleting the Data

Deleting the data ends both its value and its risk for the company, and it is the only path that cannot be followed by any other.

• Benefit: no breach, privacy or license exposure from the deleted records, including records an erasure request under Regulation (EU) 2016/679 Art. 17 would reach.
• Risk: deleting records the law requires the company to keep, since legal retention duties apply.
• Downside: the step is permanent, with no later license, research use or own use.

Deleting after a license does not reach trained models. micro1's master terms s.4.4 require deletion of the data within 120 days after delivery completes, or on termination if earlier, with no duty to delete or retrain Licensee Materials and backups excluded, and Labelbox states that models already trained on the data cannot be unlearned (checked 2026-11-09). Regulators have ordered model deletion against developers, as in the FTC order above; a seller deleting its own copy does not achieve it. What survives deletion is governed by model retention rights after a license ends.

Weighing a Non-License Path: Benefits, Risks, Downsides and Alternatives

Choosing a path other than a full license trades possible income for control, and the trade is easiest to reverse when the company keeps the data.

Benefits

• No irrevocable grant is made, unlike micro1's master terms grant.
• Trained models never hold the data.
• Confidentiality and trade-secret status are preserved, whereas a license protects trade secrets only by contract.
• No buyer counterparty risk.

Risks

• Privacy duties stay with the company on every path except deletion.
• An open release cannot be recalled.
• Research exceptions apply once data is shared.

Downsides

• Income is forgone where offers exist; a listing range such as Mercor's is a claim, not a paid amount.
• Own models cost the company itself.
• Deletion is permanent.

Alternatives, including not selling or licensing at all

• License after all, once rights and personal data are cleared.
• License a narrower, nonexclusive slice (micro1 grants nonexclusive; Labelbox is exclusive by default).
• Wait, since no grant has been made.
• Do nothing at all.

A no today can become a license later, but an irrevocable grant, once signed, cannot be reversed.

Where the Alternatives Sit Before the Process to Sell or License Company Data

The alternatives come before step 1 of the seller process, and every later step can still end in one of them. The full process to sell or license company data to AI companies runs from readiness to payment.

• should you sell or license your data.
• AI data monetization models.
• selling direct or through a broker.
• data sale readiness.
• data inventory for an AI data sale.
• rights audit.
• choosing a buyer for your data.
• what company data is worth to AI companies.

Alternatives When a Company Is Closing or Insolvent

A closing company has the same paths, but vendors delete data on their own schedules, so doing nothing can turn into deletion by default. Microsoft documents that 365 for business data is deleted no later than 180 days after cancellation (page dated 2026-06-08), and Slack's free plan has deleted history older than one year since 2024-08-26. In a US bankruptcy, personal information covered by a privacy policy may be sold only if consistent with it, or after a consumer privacy ombudsman under 11 U.S.C. 332 and court approval after notice and a hearing (11 U.S.C. 363(b)(1)). micro1 states USD 100K to 3M+ company compensation in its defunct-company program (checked 2026-11-09). Deletion clocks are covered for data from a shut-down or defunct company. Officeholders handle data assets in insolvency.

Pages That Cover Each Alternative in Depth

Each path has its own page with the full rules and terms. This page is information, not legal advice.

• Keep and own models: GDPR and AI training data; enterprises buying data to train their own models; EU Trade Secrets Directive and licensed datasets; Defend Trade Secrets Act.
• Release openly: public domain and openly licensed material.
• Share and license less: EU text and data mining exception and the opt-out; collective licensing for AI training; pay-per-crawl and crawler licensing.
• Delete or close: deletion, return and machine unlearning clauses; exporting company data before shutting down.
• Programs cited above: micro1 data buyer profile and Mercor data buyer profile. Labelbox Prism, Cloudflare, RSL Collective and Copyright Clearance Center are cited without a profile.

Frequently Asked Questions

Can a company keep its data instead of selling it to AI companies?

Yes; keeping the data is the default, and no law obliges a company to license it to AI developers. Secrecy keeps trade-secret status under Directive (EU) 2016/943, and a license stays possible later.

Can openly released data still be licensed to an AI buyer later?

Only with difficulty: once data is public, a buyer can usually get it without a license, and some programs reject it outright. Mercor states that its software listing wants repositories that have never been public or open-sourced (checked 2026-11-09), and Common Crawl states its archives are immutable.

Does deleting data remove it from AI models already trained on it?

No; deleting the company's own copy does not change a model already trained on it. micro1's master terms s.4.4 impose no duty to delete or retrain Licensee Materials, and Labelbox states trained models cannot be unlearned.

Is not selling data to AI a valid choice?

Yes; not selling is a valid outcome, and it is the only choice that leaves every other path open. Licensing income is uneven, as Reddit's 10-Q for the quarter ended 2026-09-30 shows for its two main partners.