United States: AI Training Data Law (Federal and State)

US law on AI training data is not one statute: copyright, sector privacy statutes, the FTC Act, wiretap and export rules apply at the federal level, and state privacy, biometric and recording laws add a second layer.

This page sets out 13 federal and 12 state instruments and 13 dated changes, as of 2026-11-09, each with its citation and date, and it is free to read. It covers the US side of AI training data licensing for holders and developers, and it keeps final law apart from bills and undecided cases.

What US Law Decides Before a Company Licenses Data for Training

US law decides four things before a data license: whether copyrighted content may be used, whether personal data in the set may be disclosed or sold, whether recorded or stored communications may be passed on, and whether confidential or controlled data may leave the company or the country.

• Copyright: fair use under 17 U.S.C. 107, decided case by case.
• Personal data: sector statutes such as HIPAA (45 CFR 164.514), the FTC Act (15 U.S.C. 45) and state privacy acts.
• Communications: the Wiretap Act and the Stored Communications Act (18 U.S.C. 2511, 2702).
• Confidentiality and export: the Defend Trade Secrets Act (18 U.S.C. 1839), the EAR (15 CFR 734.13) and the DOJ rule at 28 CFR Part 202.

No federal statute governs AI training data as such: Executive Order 14365 of 2025-12-11 (90 FR 58499) directs a recommendation to preempt conflicting state AI laws and sets no training-data rule.

Federal Statutes and Rules That Reach a Training Data Deal

Thirteen federal instruments shape a US training-data license, and the table gives each one's citation, its dates and whether it is final or pending. Rows run in deal order as of 2026-11-09, the bills row holds proposals only, and Executive Order 14365 has its own row.

How the Federal Layer Reaches the Seller and the Buyer

The table lists each instrument; the sections below show what each one asks of the holder and of the AI developer.

Copyrighted Content in the Set: Decided Case by Case in Court

No US statute permits or bars AI training on copyrighted works; 17 U.S.C. 107 leaves each use to a court weighing four factors, and statutory damages under 504(c), in force since 1978-01-01, run from USD 750 to 30,000 per work, up to 150,000 if willful. The records split: Bartz v. Anthropic (N.D. Cal. 3:24-cv-05417-AMO, Dkt. 231, 2025-06-23) held training fair use, the Third Circuit held ROSS's use of Westlaw headnotes not fair (No. 25-2153, 2026-09-29), and the Copyright Office's Part 3 report is only a pre-publication version of 2025-05-09. The factors, rulings and damages are set out on the page on fair use and AI training in US law.

Health, Financial, Student, Children's and Tax Records: Sector Statutes

Five federal statutes cover records a company may hold, and each allows licensing only on its own condition: de-identification, notice and opt-out, written consent or separate parental consent.

• HIPAA: de-identified data is outside the Privacy Rule (45 CFR 164.502(d)); Safe Harbor removes 18 identifiers, and licensing PHI for a fee is a sale needing an authorization stating the remuneration (164.508(a)(4)).
• GLBA: a financial institution gives notice and an opt-out before disclosing nonpublic personal information to a nonaffiliated buyer (15 U.S.C. 6802(b)(1)), and the buyer inherits reuse limits (16 CFR 313.11).
• FERPA: de-identified records may be released after a reasonable determination (34 CFR 99.31(b)(1)); coded data serves education research only.
• COPPA: disclosing children's data to train AI needs separate parental consent, with compliance from 2026-04-22.
• IRC 7216: preparers need signed written consent, valid one year by default (26 CFR 301.7216-3), and may not sell statistical compilations except with the business (301.7216-2(o)).

Removing identifiers is the common route, described on the page on PII removal and de-identification before licensing.

Privacy Promises Made at Collection and Section 5 of the FTC Act

The FTC treats using data for AI training against the privacy promises under which it was collected as a possible unfair or deceptive practice under 15 U.S.C. 45. A staff post of 2024-02-13 warns that a surreptitious, retroactive change to terms or a privacy policy to allow training may be unfair or deceptive. A post of 2024-01-09 recalls orders to delete models built on unlawfully obtained data. For a knowing violation of a trade regulation rule, 16 CFR 1.98(d) allows up to USD 53,088 per violation, an amount the FTC notice of 2026-09-15 (91 FR 58446) keeps for 2026.

Stored Messages and Recorded Calls Under ECPA

Federal law lets one party consent to recording a call, but selling a recording made without lawful consent is itself a violation. Under 18 U.S.C. 2511(2)(d), one party's consent suffices unless the purpose is criminal or tortious; disclosing or using contents known to be unlawfully intercepted is a separate violation (2511(1)(c)-(d)), punishable by up to 5 years (2511(4)(a)).

Stored messages follow 18 U.S.C. 2702: a provider of a service to the public may not divulge stored content without the lawful consent of a sender or recipient (2702(b)(3)), while a company's internal systems are generally not public services. California, Florida and Washington require every party's consent, as the state table shows, and the comparison of call and meeting recording consent rules covers other jurisdictions too.

Confidential Archives, Controlled Technology and Foreign Buyers

Three federal regimes decide whether confidential or controlled data may go to a buyer at all: trade-secret law, export controls and the DOJ rule on bulk sensitive data.

• Trade secrets: under the DTSA, in force since 2016-05-11, disclosure by someone bound to secrecy is misappropriation (18 U.S.C. 1839(5)), and claims run 3 years from discovery (1836(d)).
• Export controls: releasing controlled technology to a foreign person, including a buyer's staff, is a deemed export (15 CFR 734.13); published datasets are not subject to the EAR (734.7), and government CUI stays out unless controls permit (32 CFR 2002.4(h)).
• Bulk data: since 2025-04-08, 28 CFR Part 202 bars licensing bulk US sensitive data, even de-identified, to China, Cuba, Iran, North Korea, Russia, Venezuela or covered persons, above thresholds such as genomic data on more than 100 or health data on more than 10,000 US persons (202.205).

Confidentiality and export terms then go into the license, as described in AI data licensing agreements.

State Laws Compared: Sale Rules, Sensitive Data, Biometrics and Recording

Twelve states have a page here, and the table shows for each one its privacy act, its rule on selling sensitive data, its biometric or health-data rule, its recording consent rule and any rule written for AI training. Rows run A to Z as of 2026-11-09; "not recorded" means no primary source was verified on that date, not that the state has no such law.

Where State Law Changes the Deal: Sale Opt-outs, Bans and Training Notices

State law changes three things in a data deal: whether licensing personal data counts as a sale, whether sensitive data may be sold at all, and whether the developer must disclose what it trained on. Under the CCPA, licensing personal information for money is a sale that consumers may opt out of (Cal. Civ. Code 1798.120), and regulations effective 2026-01-01 require a risk assessment before selling. Utah and Virginia define a sale as an exchange for monetary consideration, and Utah counts synthetic data as deidentified data (Utah Code 13-61-101).

Sensitive data meets four different rules across the twelve states.

• Consent: Virginia (59.1-578) and Texas (ch. 541).
• Notice and opt-out: Utah (13-61-302).
• Outright ban: Maryland, which also bars selling data of consumers known to be under 18 (14-4707(a)(5)).
• Signed authorization naming the buyer, valid one year: Washington, for consumer health data (RCW 19.373).

Biometrics and training add a third layer. Illinois BIPA requires a written release before collecting face or voice templates (740 ILCS 14/15(b)) and bars profiting from them (15(c)), with damages of USD 1,000 or 5,000 and a single-violation rule under Public Act 103-0769, effective 2024-08-02. Texas HB 149, in force since 2026-01-01, exempts biometric identifiers used to train AI unless the system identifies a specific person. California's AB 2013 makes developers publish whether training data was purchased or licensed (per the category list in an amicus brief in the appeal) and stays in force while X.AI v. Bonta is on appeal (9th Cir. No. 26-1591). Connecticut privacy notices must say from 2026-07-01 whether personal data trains large language models (P.A. 25-113), Colorado SB 26-189 requires training-data categories from 2027-01-01, and Oregon's attorney general (2024-12-24) bars retroactive notice changes.

Enacted, Pending and Proposed: US Changes as of November 2026

Thirteen dated items changed or may change US rules, and the table keeps what is in force apart from what is a bill, a proposal or an undecided case. The rows run newest first, with status as of 2026-11-09.

Bills are not law, the Copyright Office report remains a pre-publication version, the appeal in X.AI v. Bonta does not suspend AB 2013, and MDL 3143 has no summary judgment ruling. Court filings on these questions are followed in the tracker of AI copyright lawsuits over training data.

Benefits, Risks, Downsides and Alternatives of Licensing Data Under US Law, Including Not Selling

US law lets a company license data it holds the rights to, often on contract alone for non-personal data, but each federal and state layer adds conditions, and not licensing remains a valid choice.

Benefits

• Non-personal and properly de-identified data licenses on contract terms (45 CFR 164.502(d)).
• A written license documents lawful acquisition, while the Copyright Office report (p. 52) says knowing use of pirated works weighs against fair use.
• Trade-secret protection survives a license that keeps the data confidential (18 U.S.C. 1839).

Risks

• Recording liability under 18 U.S.C. 2511 and the all-party states (California, Florida, Washington).
• BIPA damages per person in Illinois.
• FTC action that can end in deletion of models and algorithms.

Downsides

• Sale notices, opt-outs and risk assessments before income (CCPA and its 2026 regulations).
• AB 2013 summaries can show that a developer used purchased or licensed data of the seller's category.

Alternatives

• License only de-identified or non-personal data.
• License only to US buyers where Part 202 or export rules apply.
• Wait for the pending items in the change table.
• Do not sell or license at all; the options are compared on the page on alternatives to selling data to AI labs.

Beyond the US Layer: Other Jurisdictions, Topic Comparisons and the Sale

US law stops at the border and at the contract; the pages below cover other jurisdictions, topic comparisons and the sale.

• AI training data law by jurisdiction.
• Canada rules for licensing data.
• EU rules: GDPR, AI Act, TDM exception and Data Act.
• copyright and AI training across jurisdictions.
• employee data and AI training rules.
• web scraping for AI training.
• US bankruptcy data sales.
• selling data to AI companies from outside the US.
• sell or license company data to AI companies.
• privacy law and AI training data compared.
• biometric data in AI training.
• trade secrets in AI data licensing.
• voice and likeness rights in AI training data.

How This Page Is Sourced and Dated

Every rule here cites the U.S. Code, the Code of Federal Regulations, a state code or a court record, with the date it was checked. The sourcing rules are set out in how we source, date and check every figure. This page is general information about US law, not legal advice for any specific deal.

Frequently Asked Questions

Is there a federal law on AI training data?

No single federal statute governs AI training data; copyright, sector statutes, the FTC Act, ECPA and export rules apply instead. Senate bills S. 2455 and S. 2367 were referred to committee in July 2025.

Is licensing customer data for AI training a "sale" under state privacy law?

Yes, in California licensing personal information for money is a sale, and Utah and Virginia define a sale as an exchange for monetary consideration. Maryland bans selling sensitive data outright.

Can a company license recorded calls for AI training in the US?

Federal law accepts one-party consent, but California, Florida and Washington require all parties to consent. Disclosing a recording known to be unlawfully made is a separate federal violation (18 U.S.C. 2511(1)(c)-(d)).

Do AI developers have to disclose licensed training data in the US?

No federal duty exists, but California's AB 2013 requires developers to say whether training data was purchased or licensed. That category is listed in an amicus brief in the appeal, and the law applies from 2026-01-01 while the challenge is pending.