Key takeaways
- The Third Circuit in Ross accepted that unlicensed AI training can harm a developing licensing market, which strengthens rights holders’ arguments.
- The decision concerned a non-generative tool, so it does not decide how courts will treat large language model training.
- A training data deal should fix the grant, output limits, warranties and indemnity, term and deletion, and audit and privacy terms.
- Keep a provenance file for every dataset: contracts, versions, ingestion dates, source terms and opt-out signals honoured.
The Third Circuit’s decision in Thomson Reuters v. ROSS Intelligence did more than reject one company’s fair use defence. It treated the market for licensing material as AI training data as a market that copyright owners can claim to be harmed in, even where the owner had not yet licensed its content for that purpose. That reasoning strengthens the hand of anyone selling data and raises the stakes for anyone buying it.
This guide is for US AI companies that train models on third-party content. It explains what Ross implies for licensing, why a licence is worth having even where fair use might apply, the clauses that matter in a training data agreement, how deals are priced and where the ruling’s limits lie. It is written by an Indian law firm for founders with cross-border data: US copyright and contract points must be confirmed with US counsel.
What Ross signals about the licensing market
In affirming on 29 September 2026, the Third Circuit weighed the four fair use factors and found that three favoured the copyright owner. Commentary on the decision says the court identified harm to two markets: the market for legal research tools, and a developing market for licensing legal content as AI training data, a market the court treated as relevant even though the owner had not yet licensed for that use (Ballard Spahr; Mondaq; Jenner & Block).
The practical point is that a court has now accepted, at the appellate level, that unlicensed training can harm a licensing market that is only starting to exist. Publishers, data brokers and rights holders can point to Ross when they describe their licences as a real market, and each new deal adds evidence that the market exists.
The decision is limited. ROSS’s tool was not generative and the court distinguished generative AI, so it does not decide how a court will treat training of a large language model. Other US decisions have gone differently on their facts: in 2025 courts held training on lawfully acquired books to be fair use in some circumstances, while treating a pirated library as a separate problem, and the Bartz v. Anthropic class settlement of about 1.5 billion dollars shows how costly the pirated-data issue can be. The law is developing, which is a reason to contract carefully, not to wait.
Why license when fair use might apply
Fair use is a defence raised after a claim is filed, decided on the facts, and expensive to litigate. A licence moves the question from litigation to contract. It also gives the company certainty for investors and customers, who increasingly ask for training-data provenance in diligence and enterprise contracts.
A licence can also carry commercial value beyond legal cover: access to clean, structured, updated data; metadata and labels; permission to use the publisher’s brand; and exclusivity over a domain. For specialist models, a licensed proprietary corpus can be a competitive advantage that a scraped one cannot match.
There is a tension to manage. Paying for a licence may be cited by a counterparty as evidence that a market exists, which can weaken a later fair use argument over similar data. US counsel can advise on how to word the recitals, for example by stating that the licence is entered without admission that a licence is required.
Investors and enterprise customers are the other audience for a licence. Diligence questionnaires now routinely ask where training data came from, whether any of it is scraped, and what contractual or legal basis supports its use. A company that can produce a licence, a provenance file and a clear policy answers quickly and keeps the deal moving. A company that cannot may face delay, price pressure or a request for a special indemnity.
The clauses that matter in a training data deal
Grant and scope. State exactly what the company may do: train a base model, fine-tune, evaluate, build embeddings, use retrieval at inference time, or do all of these. Training and retrieval raise different issues, because retrieval can display or summarise content to users. Define the licensed data precisely, including versions and updates.
| Clause | What to settle | Common pressure point |
|---|---|---|
| Grant and scope | Training, fine-tuning, evaluation, embeddings and retrieval, and for which models | Whether retrieval and display to users are covered |
| Outputs | Whether outputs may reproduce licensed content, in what quantity and with what attribution | Verbatim regurgitation limits |
| Warranties and indemnity | Ownership, lawful collection and contributor licences, backed by an indemnity | Cap on liability and carve-outs |
| Term and deletion | Perpetual rights for trained models versus deletion of raw data | Technical difficulty of removing data from a trained model |
| Audit and privacy | Audit of use, security controls and personal data terms | Cost and scope of audit rights |
Outputs. Say whether outputs may reproduce licensed content, in what quantity, and with what attribution. Many licensors allow training but restrict verbatim regurgitation. Agree on technical measures such as filters, and on a process for fixing problems.
Warranties and indemnity. A buyer should ask the licensor to warrant that it owns or controls the rights, that the data was collected lawfully and that licences from contributors permit this use, backed by an indemnity for third-party claims. Licensors often resist uncapped indemnities, so the realistic outcome is a cap tied to fees, with carve-outs for fraud and for the licensor’s knowing breach.
Term, deletion and survival. Decide whether the licence is perpetual for models already trained. Deleting data from a trained model is technically hard, so many deals let the buyer keep trained models while stopping new training and requiring deletion of the raw dataset. Spell out what happens on expiry, breach or insolvency.
Audit, security and privacy. Licensors usually want to audit use and require security controls. If the data includes personal information, add data protection terms. If any data originates in India, the DPDP Act, 2023 will apply to its processing from 13 May 2027, and Indian transfer, consent and notice rules may be relevant.
Pricing, structure and compliance records
Common structures include a flat fee for a fixed dataset, recurring fees for updates, usage-based fees tied to queries or tokens, revenue shares on products that use the data, and mixed models with a minimum guarantee. Exclusivity usually commands a premium and should be limited by field and time, to avoid competition law concerns and commercial lock-in. We have not cited market price levels, because reported figures are inconsistent and vendor-sourced.
Most-favoured-customer clauses, step-down pricing for large volumes, and audit rights over royalty reporting are worth negotiating. If you are the licensor, protect your brand with approval rights over how it is displayed, and keep the right to license the same content to others.
Whatever the deal, keep a provenance file: contracts, dataset versions, date of ingestion, source terms and any opt-out signals honoured. This record helps in diligence, in responding to a rights-holder complaint, and in showing good faith if a dispute arises. For data you do not license, record why you consider the use lawful.
Smaller companies often assume they cannot afford to license. Options exist short of a large bespoke deal: licences from collecting bodies and aggregators, openly licensed datasets whose terms have been checked, synthetic or self-generated data, and partnerships in which the content owner receives equity or a revenue share instead of a large upfront fee. Whatever the route, read the licence for restrictions on commercial use, derivative models and sharing with customers, because open licences are not all the same.
Limits, remedies and quick answers
Ross does not make every unlicensed training run unlawful, and it does not set a price for licences. It does mean that arguing that no licensing market exists is harder in at least one federal circuit. Before you train, sort your data into three groups: licensed, public domain or openly licensed, and unlicensed. For the third group, take US advice on fair use risk. If a rights holder sends a demand, preserve records, pause use of the disputed data if it is easy to do so, and respond through counsel. Possible outcomes include a retrospective licence, retraining on clean data, or litigation. For founders with Indian operations, align the licence with Indian law on assignments, stamp duty and cross-border payments, and consider whether India’s own debate on a mandatory licensing model for AI training, discussed in the DPIIT working paper of December 2025, could affect how you work with Indian content. A well drafted licence limits that uncertainty by fixing rights in writing rather than leaving them to a court.
Related reading: Thomson Reuters v. Ross and legal tech startups, copyright in legal headnotes and AI training, the October 2026 AI music copyright suits, AI training data licensing with Indian publishers, the DPIIT AI copyright working paper, AI copyright in India vs the US and the DPDP Act for US SaaS companies; also AI training data provenance audits after Bartz.
Quick answers
Does Thomson Reuters v. ROSS mean all AI training needs a licence?
No. It concerned a non-generative tool and specific headnotes, and the court distinguished generative AI. But it accepted that a developing licensing market for AI training data can be harmed, which strengthens rights holders’ arguments.
What should an AI company ask for in a data licence?
A clear grant covering training and any retrieval use, warranties on ownership and lawful collection, an indemnity with a realistic cap, rules on outputs, deletion and survival terms, and audit and data protection provisions.
Can a licensor use our deal as evidence of a licensing market?
Possibly. That is why recitals should state that the deal is made without admitting a licence is required, subject to advice from US counsel.
Does a licence admit that a licence was needed?
Not necessarily. Recitals can state that the licence is entered without admitting one is required, and US counsel can advise on the wording because a licensor may cite your deal as evidence that a market exists.
What if we cannot afford a large licence?
Options include licences from aggregators or collecting bodies, openly licensed datasets whose terms you have checked, synthetic or self-generated data, and revenue-share or equity arrangements instead of a large upfront fee.
Final word
Halverton & Co. is an Indian law firm advising US companies, founders, investors and law firms on Indian law, including cross-border technology contracts, DPDP compliance and Indian-law aspects of data licensing for AI companies. We practise in Jharkhand, Maharashtra and before the Supreme Court of India, and work alongside US counsel, who advise on US law. Halverton & Co.: Where tech needs law! If you have a question about Indian law, write to us at office@halvertonandco.com, or get in touch.
This article reflects developments reported up to late October 2026. It is for general information only, is not legal advice, and does not create an attorney-client relationship. Halverton & Co. is an Indian law firm and does not advise on US law; US-law points should be confirmed with US counsel.
Sources
Related practice area
Technology & AI Law
Contracts, data protection and AI policies for products that handle other people’s data.
View the practice areaComments
No account needed. Comments are reviewed before they appear, so yours will show here once approved. Please do not share confidential details or describe your own legal matter: replies are general discussion, not legal advice, and do not create an advocate-client relationship. For your own situation, write to office@halvertonandco.com.
