A business that licenses its records to an AI developer is not selling a file. It is passing along a bundle of rights it may not fully hold, to a buyer who will turn the data into something it cannot give back. The agreement has to say which rights travel and which stay behind. This is the list, as the law stands in October 2026.
Who owns it, and who can license it.
Facts, measurements, and records are mostly not protected by copyright; the selection and arrangement of a database may be. What an owner actually holds is usually a mix: copyright in written work (transcripts, reports, documents), trade-secret protection in information kept confidential and commercially valuable, contractual rights against customers and vendors, and possession of personal data it is allowed to process for stated purposes.
Copyright can be licensed. A trade secret can be disclosed under confidentiality, but under US federal law it stays a trade secret only while the owner takes reasonable measures to keep it secret [23]. Personal data cannot be “owned” in the property sense at all; it can only be processed on a legal basis, and that basis has to cover the new use. A licence is therefore a copyright licence, a confidentiality undertaking, and a data-processing arrangement, stacked, with a warranty behind each.
The practical test is the chain: where each record came from, under what terms, and whether those terms allow this. Vendor data may carry a no-sublicensing clause. Employee-written material is usually the employer’s; contractor-written material often is not. The Data Provenance Initiative’s audit of more than 1,800 public text datasets found that over 70 percent were missing license information on the hosting sites where developers find them, and over 50 percent of the licenses that were listed were categorized wrongly [18]. Private business data is usually better documented, but only because someone did the work.
Consent and the terms your customers agreed to.
Most business data is about customers. Whether it can be licensed depends on what they were told, and the last three years have shown what happens when the answer is “not much.”
In August 2023, Zoom’s updated terms were read as granting it the right to train AI on meeting content. Within days the company amended section 10.4 to say it would not use customer audio, video, or chat to train its models without consent [1]. In May 2024, Slack drew the same reaction when readers noticed that its privacy principles allowed customer messages to train its machine-learning models by default, with opt-out available only by an email from a workspace owner; Slack clarified that the models in question were non-generative and that its paid Slack AI did not train on customer data [2]. In June 2024, Adobe’s revised terms prompted creators to fear their work would feed generative models; Adobe reissued the terms with an explicit statement that it does not train generative AI on customer content, Adobe Stock submissions excepted [3]. In September 2024, LinkedIn began training generative models on member data under an opt-out setting, while excluding members in the EEA, Switzerland, and the UK [4]. Meta paused training on EU users’ public posts in June 2024 after the Irish regulator raised concerns, and resumed on May 27, 2025 only after adding clearer notices, an in-app objection form, longer notice periods, and de-identification measures [5].
The pattern is consistent: broad language in a terms update, discovered, then narrowed. The lesson for a licensor is not “write broader terms.” It is that terms written for one purpose do not stretch to a new one, and the buyer will ask to see them. If they do not cover disclosure for model training, the options are a fresh notice and consent, a defensible legitimate-interest analysis where the law allows one, or de-identification to the point where the data is no longer personal.
Regulation that applies in 2026.
EU AI Act. Obligations for providers of general-purpose AI models apply from August 2, 2025. Article 53(1)(d) requires a public summary of training content, using the template the Commission’s AI Office published on July 24, 2025. The template has three parts (general information, a list of data sources, and data-processing aspects) and asks providers to disclose licensed and private datasets separately from web-scraped data [7]. The voluntary General-Purpose AI Code of Practice, published July 10, 2025, has chapters on transparency, copyright, and safety and security [8]. A serious buyer will need to name your dataset, or its category, in a public document.
GDPR. EDPB Opinion 28/2024, adopted December 18, 2024 at the request of the Irish authority, says three things that matter here: a model trained on personal data is not automatically anonymous and must be assessed case by case; legitimate interest can be a legal basis for training but only after a documented three-step test, never by default; and unlawful processing during development can affect the lawfulness of the model’s later use [6]. A licence should state the legal basis for the transfer and for the buyer’s processing; training does not wash the data clean.
US state law. Twenty states had comprehensive consumer privacy laws in force by January 1, 2026, with Louisiana and Oklahoma following on January 1, 2027, Alabama on May 1, 2027, and Vermont on January 1, 2028 [10]. Most of these treat a transfer of personal data for value as a “sale,” triggering opt-out rights and notice obligations. California’s updated CCPA regulations, operative January 1, 2026, add risk assessments for selling or sharing personal information, annual cybersecurity audits, and rules for automated decision-making technology [11]. Colorado replaced its 2024 AI Act with SB 26-189, signed May 14, 2026, which narrows the law to automated decisions in employment, housing, education, lending, insurance, healthcare, and government services, effective January 1, 2027 [9].
HIPAA. For healthcare data, the relevant rule is 45 CFR 164.514(b), explained in HHS guidance issued November 26, 2012. There are two routes. Safe Harbor removes 18 listed identifier categories (names, geographic units smaller than a state, dates other than year, contact details, record and account numbers, device and biometric identifiers, full-face images, and any other unique code) and requires no actual knowledge that the remainder could identify someone. Expert Determination lets a qualified expert certify that the risk of re-identification is very small, documenting the methods, which can preserve more detail such as month-level dates [12]. A licence for clinical data should say which route was used and attach the determination.
What litigation has settled, and what it hasn’t.
In Thomson Reuters v. Ross, Judge Bibas held on February 11, 2025 that using Westlaw headnotes to train a competing legal-research tool was not fair use. The Third Circuit affirmed on September 29, 2026, calling it “no more than an ordinary copyright case,” and found that Ross had usurped Thomson Reuters’s opportunity in the “rapidly developing” market for licensing content as training data. The panel limited its holding to a system that cannot generate original expression [13].
In Bartz v. Anthropic, Judge Alsup ruled in June 2025 that training on lawfully purchased books was “quintessentially transformative” fair use, but that building a permanent library from pirated copies was not. The case settled for $1.5 billion, roughly $3,000 per title across about 500,000 works, with preliminary approval on September 25, 2025 [14]. In Kadrey v. Meta, Judge Chhabria granted Meta summary judgment on June 25, 2025, while writing that the ruling “does not stand for the proposition that Meta’s use of copyrighted materials to train its language models is lawful,” and that a plaintiff who proved market dilution could win [15].
New York Times v. OpenAI is unresolved. The motion to dismiss was largely denied in an opinion dated April 4, 2025; OpenAI was ordered to produce 20 million de-identified ChatGPT logs; cross-motions for summary judgment were filed September 4, 2026, with no trial date [16]. On September 1, 2026, the Department of Justice filed a statement of interest arguing that training-stage copying is fair use, while saying outputs must be analyzed separately [17].
Read together: how data was acquired matters as much as what was done with it; a licensing market, once it exists, counts against fair use; and no court has decided whether a generative model’s outputs dilute the market for its training data. None of those questions arises for a dataset the buyer licensed.
The clauses that decide the deal.
| Clause | What it settles | Owner’s interest | Buyer’s interest |
|---|---|---|---|
| Scope of use | Training, fine-tuning, evaluation, or all three; which models; whether derivatives count | Narrow, named uses; no sublicensing | Broad enough to cover the roadmap |
| Exclusivity | Whether the owner can license the same data to others, by field or outright [21] | Non-exclusive, or exclusivity priced in | Field-of-use exclusivity against competitors |
| Term and survival | How long the licence runs and what survives its end | Fixed term; data deleted at end | Trained models survive termination |
| Deletion and return | Whether and when the buyer must destroy copies, including backups, and certify it | Certified deletion with a deadline | Carve-out for models already trained |
| Audit | Owner’s right to verify use and, in revenue-share deals, revenue [21] | Annual audit at buyer’s cost if discrepancies | Limited frequency, confidentiality of findings |
| Warranties | Owner’s promises: rights held, consents obtained, data as described | Knowledge-qualified, capped | Unqualified, backed by indemnity |
| Indemnities | Who pays for third-party claims about rights or privacy | Cap tied to fees received | Uncapped for IP and privacy claims |
| Price mechanism | Flat fee, per-unit rate, or revenue share | Flat fee with no audit burden; share if upside is real | Pay for what is used; defer cost |
| Confidentiality | Treatment of the data as a trade secret during and after the term [23] | Secrecy measures that preserve the trade secret | Freedom to use outputs and learnings |
| Legal basis and de-identification | The GDPR basis, the HIPAA route, the state-law notices relied on [6][12] | Buyer’s obligations mirrored in writing | Clean data; no residual consent risk |
For scale, Reddit’s non-exclusive content agreement with Google was reported in February 2024 at about $60 million a year [22]. Private business data is priced on scope, exclusivity, and how clean the rights are.
Provenance as a document.
A licence is a promise about rights; a provenance record is the evidence. Datasheets for Datasets (2018) proposed seven sections: motivation, composition, collection process, preprocessing, recommended uses, distribution, and maintenance [19]. The Data Provenance Initiative added lineage: the source of each record, its creators, the chain of licence conditions, and subsequent use [18]. Content Credentials (C2PA) attach signed manifests to media, though C2PA clarified in January 2026 that training and data-mining permissions are not part of its core specification and belong to extensions [20]. The EU’s public-summary template now asks providers to disclose licensed datasets as a category [7].
Mikra’s Data Passport is our version of this record, written for one dataset at a time: origin, rights basis, consents and the terms relied on, de-identification methods, what was excluded and why, and the permitted uses. It travels with the data so the buyer can answer a regulator and the owner can answer a customer.
Checklist before you sign.
- Inventory the rights: copyright, trade secret, contract, personal data, and for each, who holds it and under what document.
- Pull the customer-facing terms and privacy notices in force when the data was collected, and confirm they cover disclosure for this use [1][2][3][4].
- Name the legal basis for personal data (consent, legitimate interest with a documented test, or de-identification) and write it into the agreement [6].
- For healthcare data, choose Safe Harbor or Expert Determination and attach the determination [12].
- Check state-law “sale” definitions and opt-out obligations in every state where data subjects live [10][11].
- Define scope of use, exclusivity, and term precisely, and decide what happens to trained models when the term ends [21].
- Cap warranties and indemnities to what you can verify, and keep audit rights proportionate to the price mechanism.
- Ship a provenance record with the data, and keep a copy [18][19].
This is not legal advice.