Data and AI in contracts: what may your counterparty do with your data?

In 1981 Lynn Goldsmith photographs the musician Prince. Three years later she licenses the photo to a magazine for 400 dollars, as an “artist reference” for an illustration, expressly for one-time use. The illustrator the magazine hires is Andy Warhol. He delivers more than the commissioned illustration; the photo becomes 16 works. When the Warhol Foundation licenses one of them to Condé Nast for 10,000 dollars in 2016, Goldsmith learns of the series for the first time. In 2023 the US Supreme Court rules in her favor, but only on the 2016 license to Condé Nast. The purpose of that use, the first factor of the fair-use test, does not favor the Warhol Foundation.¹

This is not art history but the base case of every data contract. One side hands over material, the other makes more of it than was agreed, and the conflict surfaces only years later. With AI the same thing happens much faster. A model enriches data in a short time, rebuilds them, and turns them into new products.

The setting: your data, their model

The pure case is the data license. A provider licenses market data, maps, or industry figures; the licensee enriches them, builds products, and licenses on to end customers. The real subject of negotiation is what the licensee, its affiliates, and its sublicensees may do: only use, or also modify, blend, analyze with AI, and pass on?

The same questions now sit in contracts nobody calls a license. In the machinery supply contract, the equipment generates operating and sensor data, and the manufacturer wants them for maintenance, product improvement, and the training of its models. In the AI or SaaS service, the provider wants to keep inputs and outputs to improve its systems. In the development contract, both sides feed in data and later dispute who owns the enriched version.

And the roles switch. In the data license, the data user is the customer. In the machinery purchase, it is the supplier. With SaaS, the provider. Whether the contract calls someone “buyer” or “seller” says nothing about who is using whose data; all that counts is the direction of the data flow. My first question in every negotiation is therefore the same: which data are you handing over, and what may the other side do with them?

Diagram: your data flows to the counterparty, which uses AI and turns it into products passed on to third parties. Two marks flag the open questions: may the counterparty train on the data, and is the result an adaptation or a transformation?

Your data

Counterparty + AITraining?

Products, distributionStill an adaptation?

Since the Data Act, this is no longer a matter of negotiation alone

For connected products the starting position has shifted. The Data Act (Regulation (EU) 2023/2854) has applied since 12 September 2025 and inverts the default: connected products placed on the market after 12 September 2026 must be designed so that the product data is accessible to the user easily, securely, free of charge, and in a machine-readable format by default (Article 3(1)). Where the user cannot access the data directly, the data holder must give the user access to the readily available data without undue delay (Article 4(1)). A simple request is enough. Even before the sale, rental, or lease, the seller, rentor, or lessor must disclose what data the machine generates, whether continuously and in real time, where it is stored, and how the user gets to it (Article 3(2)).

More important for drafting is Chapter IV. It subjects clauses on data access and data use that one company imposes unilaterally on another to a fairness review of their own: if they are unfair, they do not bind (Article 13(1)). For several clauses the Regulation presumes that they are unfair. A data owner often uses three of them. The party that imposes them can rebut the presumption in the individual case. The presumption covers a clause that gives the imposing party access to and use of the other side’s data in a way that seriously harms that side’s legitimate interests, in particular where trade secrets or protected rights are involved (Article 13(5)(b)). It equally covers a clause that prevents the other side from making reasonable use of the data it provided or generated (point (c)), and a clause that denies that side a copy of that data during the term or within a reasonable period after it (point (e)).²

What matters is whether a clause is unilaterally imposed. That is the case where one party supplies it and the other cannot influence its content despite trying to negotiate; the burden of proving otherwise falls on the party that supplied it (Article 13(6)). If you present your data clauses as a non-negotiable block, they are subject to the review in a dispute. Why the counter-clause, that everything was individually negotiated, does not help under German standard terms law, I set out in the article “No individual agreement”. In my view the same applies under the Data Act, because a pre-drafted confirmation does not prove that the other side could influence the content.

The line on which control hangs

Where that line runs, I show by the example of US copyright, the venue of the current AI litigation. As long as the counterparty adapts your data, you as the data owner are in the stronger position, to the extent the data are protected by copyright; bare facts and raw data are not, and a compilation is protected at most in its selection and arrangement. The adaptation (derivative work) is, in US law, the copyright owner’s exclusive right (17 U.S.C. § 106(2)): whatever you do not permit in the contract stays reserved. In addition, there is no protection for any part of the adaptation in which your data have been used unlawfully (17 U.S.C. § 103(a)).³ The danger lies in transformation. If the use is transformative, the fair-use limitation can apply, and copyright control then ends. It is exactly this line that the Supreme Court shifted in favor of rights holders in Warhol v. Goldsmith: the transformation needed for transformative use must go beyond what already makes a work derivative, and where both works serve the same commercial purpose, a new message alone is not enough. I set out the doctrine in detail in the reference article “IP clauses in commercial contracts”.

For the contract this means two things. First: “modifications” is the most contested word of the data contract. Define it, and allocate the modifications expressly, through reassignment, license, or prohibition. Be careful with the word ownership: German law knows no ownership of data, and the debate about it has expressly run the other way.⁴ What gets allocated are powers of use, duties to hand over, and duties to delete. Second, the contract has a limit. It binds only the parties. Against your own counterparty, a modification bar works beyond the copyright line as a contractual duty. The chain of affiliates, end customers, and sublicensees is reached only if you oblige the data user in the contract to bind those parties in turn. In practice this duty is called flow-down. The data user must impose restrictions on its customers that are no less restrictive than its own, up to and including bans on compiling competing databases from the data or publishing representations that enable third parties to scrape them.

These clauses are needed. The statute alone does little here. The sui generis database right applies to distributed access only where the users act in conscious and deliberate concert, and repeated retrieval of insubstantial parts does not infringe it as long as it is not aimed at reconstituting the database as a whole or in substantial part. And in the scraping case, even a contractual bar on automated retrieval was not enough for the Federal Court of Justice to treat the querying of freely accessible data as unfair obstruction.⁵

Training and competing products: the new core

The biggest AI question is not enrichment but training. A model, once trained, preserves the value of your data permanently, even after termination and deletion of the raw data. And the US courts showed in 2025 that copyright does not reliably stop training:⁶

This line is fact-dependent, ended in one case by settlement, so it is not settled law. But Thomson Reuters v. Ross shows clearly that fair use fails most readily where the data are used to train a competing product, for the same market purpose. That is also where the contractual line has settled in large data license agreements: whether training is permitted is negotiable. What should not be negotiable is the bar that trained systems must not produce products that compete with the data owner’s data.

In Europe, what counts is a formality. Text and data mining of lawfully accessible works is permitted as long as the rights holder does not expressly reserve it. For works accessible online, the reservation must be machine-readable (Section 44b(2) and (3) of the German Copyright Act, Urheberrechtsgesetz, UrhG). How much that reservation achieves is still open. In 2025 the Higher Regional Court of Hamburg accepted in principle a reservation declared by a licensee. But the rights holder must show that machines could read and evaluate the reservation at the time of the use; for a reservation in natural language and a use in 2021, the photographer who brought the claim failed to do so.⁷ The case is now before the Federal Court of Justice. Declare the reservation, oblige your counterparty in the contract to pass it on with every onward transfer, and also apply it technically, precisely because the law is unsettled.

What does work are two other routes, and you have to secure both in the contract. Machine-generated data can be protected as a trade secret, but only where it is secured by reasonable steps to keep it secret – access restrictions, marking, confidentiality duties down the chain. And a data pool can attract a database right of its own where the investment in obtaining, verifying, and presenting it is substantial. That right does not apply, however, to data obtained from or generated by connected products and related services within the scope of the Data Act (Article 43).⁸

The clause map

Both roles need the same catalog, with the signs reversed.

As the data owner:

1. Define the permitted uses exhaustively; whatever is not expressly permitted stays reserved.
2. Define “modifications” and allocate them: reassignment, license, or prohibition. Not through ownership of data, but through powers of use, handover, and deletion.
3. Address AI training expressly: prohibit it, or permit it as a separate, separately priced license track, always with a bar on competing products. Declare the reservation under Section 44b(3) UrhG and oblige your counterparty to pass it on.
4. Bind the chain: affiliates, sublicensees, and end customers (flow-down no less restrictive than the main contract), plus audit and reporting rights.
5. Settle the end of the contract: deletion duties that cover modifications, and a clear end to further use in products, at most with a defined tail period.

As the data user:

1. Does the license cover the plan: use, modification, AI analysis, future use cases? Purpose-bound licenses end sooner than roadmaps do.
2. Secure ownership of your own value added: US law protects only what you added yourself (17 U.S.C. § 103(b)), and there is no protection for any part of the work in which preexisting material has been used unlawfully (§ 103(a)). With AI-generated enrichment comes the additional protection gap of missing human authorship, which you can close by contract only as against your counterparties; German law faces the same question for inventions, where according to the Federal Court of Justice a human inventor nonetheless remains the rule.⁹
3. Provenance and indemnity: warranties on the chain of rights in the sourced data and an indemnity for infringement cases; how the indemnifying party confines it, I set out in the article “The indemnity clause in commercial contracts”.

Two model building blocks that cover the core. The highlighted terms are placeholders for the data-providing and the data-receiving party and for references to other sections. Replace them with the contract’s own designations and section numbers.

Conclusion

Lynn Goldsmith received 400 dollars and won before the Supreme Court decades later, but only on the 2016 license to Condé Nast. Do not rely on that route. It takes years and is expensive. It is open to you only as long as the counterparty adapts your data and does not transform them. And the contract must be clear. If you hand over data without settling training, modification, and the chain, your counterparty can use them to build a product that replaces yours.

Notes

1. Andy Warhol Foundation for the Visual Arts, Inc. v. Goldsmith, 598 U.S. 508 (2023); the details of the 1984 license per the syllabus of the decision. ↩

2. Regulation (EU) 2023/2854 (Data Act), OJ L of 22 December 2023; Articles 3(1) and (2), 4(1), 13(1), (5), and (6) with recital 62, and 50 (application from 12 September 2025; Article 3(1) applies to connected products placed on the market after 12 September 2026). The German version of Article 13(5) reads “gilt als missbräuchlich” (is deemed unfair); the English version and recital 62, in the German version as well, provide for a rebuttable presumption. For context Hennemann/Steinrötter, Der Data Act, NJW 2024, 1. ↩

3. 17 U.S.C. §§ 101, 102, 103, 106, 107; on facts not being protected Feist Publications, Inc. v. Rural Telephone Service Co., 499 U.S. 340 (1991); foundational Campbell v. Acuff-Rose Music, Inc., 510 U.S. 569 (1994). ↩

4. On the rejection of data ownership Hoeren, Datenbesitz statt Dateneigentum, MMR 2019, 5; Kühling/Sackmann, Irrweg “Dateneigentum”, ZD 2020, 24. ↩

5. BGH, judgment of 22 June 2011 – I ZR 159/10 (Automobil-Onlinebörse), headnotes a and b on Section 87b(1) UrhG; BGH, judgment of 30 April 2014 – I ZR 224/12 (Flugvermittlung im Internet), on the fair-trading assessment of automated retrieval despite contrary terms of use; set aside and remanded there. ↩

6. Thomson Reuters Enterprise Centre GmbH v. Ross Intelligence Inc. (D. Del., 11 February 2025); Bartz v. Anthropic PBC (N.D. Cal., 23 June 2025), settlement announced in September 2025; Kadrey v. Meta Platforms, Inc. (N.D. Cal., 25 June 2025). ↩

7. OLG Hamburg, judgment of 10 December 2025 – 5 U 104/24, paras. 99 and 103 (reservation by a licensee), paras. 112, 117, and 119 (machine-readability), appeal on points of law pending as BGH I ZR 281/25; lower court LG Hamburg, judgment of 27 September 2024 – 310 O 227/23. ↩

8. On the protection of machine-generated data under the German Trade Secrets Act (Gesetz zum Schutz von Geschäftsgeheimnissen, GeschGehG) Hessel/Leffer, MMR 2020, 647; on the data pool as a trade secret Krüger/Wiencke/Koch, GRUR 2020, 578; on the sui generis database right Sections 87a et seq. UrhG, on the exclusion for data from connected products Section 87b(3) UrhG. ↩

9. Thaler v. Perlmutter, 130 F.4th 1039 (D.C. Cir. 2025); U.S. Copyright Office, Copyright and Artificial Intelligence, Part 2: Copyrightability (January 2025); on the human inventor where AI is used BGH, order of 11 June 2024 – X ZB 5/22 (DABUS), paras. 38, 40, 44, and 47; Gotthardt, Regelungen des Einsatzes von künstlicher Intelligenz in Forschungs- und Entwicklungsverträgen, GRUR 2024, 1870. ↩

Reference: Poleacov, P. (2026). Data and AI in contracts: what may your counterparty do with your data?. INN.LAW. https://inn.law/en/perspectives/data-ai-contracts/

---