What Company Data Cannot Be Licensed for AI Training

Company data cannot be licensed for AI training when a law bars its sale or disclosure, when a contract with its source forbids relicensing, or when the rights already belong to someone else. Dated examples: 28 CFR Part 202 has barred licensing bulk US sensitive data to buyers linked to six countries of concern since April 8, 2025, even when anonymized, and Xero's terms bar sublicensing its data (checked 2026-11-09).

Which records stay out is the first legal question in AI training data licensing. De-identification lifts only some bars, and licensing only the cleared subset, or not licensing at all, are both valid outcomes.

DataPacts Editorial Team checked every rule and term here on November 9, 2026; the page on how we source, date and check every figure explains the method.

Three Reasons Company Data Cannot Be Licensed for AI Training

Every bar on licensing company data comes from one of three sources: a statute or regulation, a contract with the party the data came from, or a prior grant of the same rights. The table compares them by reason and cure, as of 2026-11-09.

The bars are checked record by record, after ownership is settled in a rights audit for the data you want to license.

Records Barred by Law Even Though the Company Holds Them

Sector laws bar or condition the sale of specific records, and licensing data for money usually counts as a sale under them; California's CCPA treats it as a sale with an opt-out right (Cal. Civ. Code 1798.120).

Health and Substance Use Records

A HIPAA covered entity or business associate may not sell protected health information without an authorization that states the disclosure results in remuneration (45 CFR 164.502(a)(5)(ii) and 164.508(a)(4)). Data de-identified under 164.514(a)-(b) falls outside the Privacy Rule (164.502(d)). Substance use disorder records follow 42 CFR Part 2, and Washington's My Health My Data Act (RCW 19.373, in force 2024-03-31) requires a signed authorization naming the buyer.

Tax Return Information Held by Preparers

Tax preparers and accounting firms may not license client return information for AI training without the taxpayer's specific written consent (26 CFR 301.7216-3). Knowing or reckless disclosure is a misdemeanor under 26 U.S.C. 7216(a), with a fine up to USD 1,000, up to one year, or both (text checked 2026-11-09).

Mercor states that its Accounting and Tax Firm listing asks for "client documents that can be anonymized for licensing" (checked 2026-11-09). The regulation states that a preparer may not sell or exchange for value a statistical compilation, except with the transfer of assets on a sale of the business (26 CFR 301.7216-2(o)(2)). No conclusion about Mercor or any firm follows.

Financial, Bank Secrecy and Telecom Customer Records

Financial institutions must give notice and an opt-out before disclosing nonpublic personal information to a nonaffiliated buyer, and some records cannot be disclosed at all (15 U.S.C. 6802(b)(1)). Aggregate data without identifiers is not covered, and a recipient's reuse is limited to the exception it received the data under (16 CFR 313.11). Suspicious activity report records cannot be disclosed (31 U.S.C. 5318(g)(2)), and carriers need customer approval to disclose identifiable customer information (47 U.S.C. 222).

Children's, Student and Biometric Data

Data about children, students and biometric identifiers carries the narrowest rules, and some of it may not be sold at all.

• COPPA: disclosing a child's personal information to train AI needs separate consent (312.5(a)(2); 90 FR 16950, compliance 2026-04-22).
• FERPA: a de-identified release needs removal of all personal information plus a reasonable determination of non-identifiability, "whether through single or multiple releases" (34 CFR 99.31(b)(1)).
• BIPA: Illinois prohibits selling, leasing, trading or profiting from biometric identifiers (740 ILCS 14/15(c)).
• MODPA: Maryland bans selling sensitive data and the data of consumers known to be under 18 (in force 2025-10-01).

Recordings Made Without the Required Consent

A recording made without the consent the law requires cannot be licensed, because disclosing or using the contents of an unlawful interception is itself a violation (18 U.S.C. 2511(1)(c)-(d)). California's Penal Code 632(a) requires the consent of all parties to a confidential communication.

Professional Secrets and Privileged Client Files

Client and patient secrets held by lawyers, physicians, accountants and insurers are bound by professional secrecy, which a data license does not release. Germany's StGB s.203 carries up to one year or a fine, and up to two years where disclosure is for remuneration (s.203(6)); SRA Code of Conduct for Firms 6.3 (version of 2025-04-11) keeps clients' affairs confidential unless law permits disclosure or the client consents.

Export-Controlled, Infrastructure and Government Information

Engineering archives can hold defense technical data, controlled dual-use technology, critical infrastructure information or government information, and each stays out unless the controls allow release. Defense technical data cannot reach foreign persons, including a buyer's foreign staff, without authorization (22 CFR 120, a deemed export); controlled technology follows 15 CFR 734.13, government-originated information 32 CFR 2002.4(h) and critical energy infrastructure information 18 CFR 388.113(c)(2).

Bulk Sensitive Data and Buyers Linked to Countries of Concern

Since April 8, 2025, licensing bulk US sensitive personal data to a buyer in or linked to China, Cuba, Iran, North Korea, Russia or Venezuela is prohibited, even when the data is anonymized (28 CFR Part 202). The thresholds in 202.205 include genomic data over 100 US persons, biometric data over 1,000, health and financial data over 10,000, and covered identifiers over 100,000.

Special-Category Personal Data in the EU

Health, genetic, biometric and other special-category data may not be processed in the EU unless an Article 9(2) exception applies, in addition to an Article 6 legal basis (GDPR, applying from 2018-05-25). The legal basis comes from the rules on GDPR and AI training data.

Data the Company Licensed In: Third-Party Vendor Data and Relicensing

Data a company licensed from a vendor can be relicensed for AI training only if that vendor's license allows sublicensing for that use, and most data and content licenses do not. This is a contract rule, not a statute: the upstream license decides, and Xero's terms state that a customer "must not repackage, resell, or sublicense any leads or data accessed through the services" (cl. 56, checked 2026-11-09).

Mercor's listings name typical licensed-in data it does not take (checked 2026-11-09):

• Redistributed exchange and market feeds the firm does not own.
• Stock or licensed imagery, and content under someone else's license, including non-commercial or no-derivatives licenses.
• Licensed third-party IP, and data owned by a customer or foundry partner.

Data Locked by Platform Terms or Earlier Promises

Workplace platforms limit what can be extracted through their interfaces and used for AI training, and the privacy promises a company made when it collected data limit its later use. Each term below is stated as of 2026-11-09.

• Slack API Terms (effective 2025-10-10): apps distributed outside the organization may not use API Data to train a large language model.
• Google Workspace API user data policy: using or selling user data to train or improve an AI model is prohibited, beyond a user's personalized model.
• QuickBooks Online: training AI models on information derived from the platform without authorization is a prohibited use; its reach over a customer's own content is ambiguous.
• WhatsApp Business Platform (modified 2026-09-23): s.4.1(b) bars selling or licensing Platform Data in any form, and s.4.7 bars AI training except fine-tuning for the business's own use.

FTC staff stated on 2024-02-13 that quietly and retroactively amending terms or a privacy policy to use data for AI training may be unfair or deceptive (15 U.S.C. 45).

Staff messages add employee notice and consent when licensing internal communications.

Data Already Licensed Exclusively to Someone Else

Data licensed exclusively to one buyer cannot be licensed again for the same use during the exclusive term, and buyers ask sellers to warrant that no such grant exists.

• micro1: its Data License Master Terms (updated 2026-11-08) have the licensor warrant in s.5.1 that no third party holds the same rights on substantially similar terms.
• Labelbox Prism: states that exclusivity is the default and that trained models "cannot be unlearned" (checked 2026-11-09).
• Mercor: states that it does not take medical data "already licensed exclusively through a broker" (checked 2026-11-09).

The exclusivity and warranty clauses of an AI data licensing agreement turn an earlier grant into a breach if the data is licensed again.

What Removing Names Does Not Fix

De-identification lifts some bars, such as HIPAA's for properly de-identified data, but several rules and contracts apply to anonymized, aggregate or derived data as well.

• DOJ bulk data rule: 28 CFR Part 202 covers bulk sensitive data even when anonymized or de-identified.
• WhatsApp Business Platform: s.4.1(b) reaches anonymous, aggregate and derived forms.
• Tax compilations: 26 CFR 301.7216-2(o)(2) bars selling statistical compilations.
• Spirit Airlines sale: the court record excludes transaction systems, refunds, inflight purchases and Wi-Fi sales in any form, "deidentified or otherwise" (Bankr. S.D.N.Y. 25-11897, Dkt. 1684, 2026-10-05); the sale was not approved as of 2026-11-09.

Deleting names cures none of these. See what PII removal and de-identification can achieve where it does apply.

What Buyers Themselves Refuse to License

Buyers publish their own exclusions, and Mercor's company-data listings name the records it does not take, checked November 9, 2026. The table lists 18 of its 25 listings that state such an exclusion, as of 2026-11-09.

These are one buyer's program rules, not law, and another buyer may differ; the terms and history sit in Mercor's data buyer profile.

Keeping Barred Data Out: Benefits, Risks, Downsides and Alternatives

Excluding barred records shrinks what a company can license but removes the exposure that follows them into every copy and trained model.

Benefits

• No breach of an upstream license or exclusive grant.
• No exposure under provisions such as 26 U.S.C. 7216.
• A written exclusion list a buyer can check.

Risks

• A record missed in sorting travels with the dataset, and Labelbox states trained models cannot unlearn it.
• A wrongly cleared recording stays a violation whatever the contract says.
• Rules and vendor terms change.

Downsides

• A smaller licensable dataset.
• Staff time to trace each system's source and terms.

Alternatives

• License only the cleared subset.
• Seek written consent or authorization where the law provides for it.
• Wait until an exclusive term ends.
• Do not license at all.

The wider question is whether to sell or license your data at all.

How to Sort Records Into Licensable, Conditional and Excluded

Sorting starts from the inventory and ends with a written exclusion list that the license warranties refer to.

1. List each system and source that holds records.
2. Tag the records by the three reasons: law, contract, prior grant.
3. Mark conditional records with the consent or authorization they need.
4. Carve out the rest before any sample leaves the company.
5. Keep the written exclusion list for the buyer's warranty.

A record with no cure stays excluded; a conditional one moves only when the paper trail exists.

Where the Exclusion Check Sits in the Seller Process

The exclusion check follows the data inventory and the rights audit and precedes consent, de-identification and the data manifest. It is one step in the sequence to sell or license company data to AI companies.

1. Data sale readiness
2. Data inventory for an AI data sale
3. Rights audit, then this exclusion check
4. Customer consent and confidentiality clauses
5. De-identification, then the data manifest
6. Alternatives to selling data to AI labs, including not licensing at all

Rules by Jurisdiction and Industry Cited on This Page

Each rule above has its own page. The hubs are AI training data law by jurisdiction, data licensing by industry and data types AI companies buy. The US bars have pages on HIPAA and AI training data, IRC Section 7216, Maryland MODPA ban on selling sensitive data, ECPA and the Wiretap Act, export controls (ITAR, EAR, CUI) and the DOJ bulk sensitive data rule. This page is general information about the law, not legal advice for any specific deal.

Frequently Asked Questions

Can a company relicense data it bought from a vendor?

Only if the vendor's license allows sublicensing for AI training; Xero's terms, for example, bar repackaging, reselling or sublicensing data accessed through its services.

Does anonymizing data make it licensable?

Sometimes: properly de-identified health data falls outside HIPAA, but the DOJ bulk data rule and some platform terms also cover anonymized or aggregate data. The legal effect of anonymization vs pseudonymization differs by regime.

Can a company license customer data for AI training?

Only where its customer contracts, privacy notices and the sector rules above allow it. The contract side is covered by customer consent and confidentiality clauses.

Is any company data banned from AI training everywhere?

No single rule bars all company data; the bars attach to specific records, holders and recipients, such as Maryland's ban on selling sensitive data. Closed companies stay bound too, because selling data from a defunct company does not lift these bars.