Introduction
British organisations buying AI tools built in the US should ask every supplier how its training data was licensed and sourced, and weigh the answer alongside price and accuracy, because the US Copyright Office has said that some commercial training goes beyond fair use and more than 70 American copyright cases are still unresolved.
By the numbers
US copyright cases counted in the UK government March 2026 report
News publishing share of announced AI licensing deals between March 2023 and February 2025, in research cited by the UK government report
UK academic publishers' commercial text and data mining licensing revenue in 2021, cited by the UK government report
Why a Washington report belongs in a British contract file
On 9 May 2025 the US Copyright Office released a pre-publication version of Part 3 of its report on copyright and artificial intelligence, the part dealing with generative AI training, and its website still says a final version will follow "without any substantive changes expected in the analysis or conclusions". Most British chief information officers will never read its 108 pages. They should know the conclusion, because the tools they license for marketing copy, product imagery, code and customer service are often developed, at least in part, in other countries, and companies training models in the US often cite the fair use principle the report tests. If a supplier's training turns out to fall outside fair use, the dispute does not stay in a Californian courtroom. It turns up in indemnity clauses and, now and then, in a letter to the customer.
My argument is narrow: training-data provenance has become a supplier risk that a UK buyer can price, so it belongs in due diligence next to security, data protection, uptime and cost. A supplier that can show licensed or well-documented data for content-generating uses is a safer bet than one that cannot, and the difference is worth paying for where the outputs go out under your brand.
The strongest objection is that American courts will decide this, not the Copyright Office
Here is the case against, at full strength. Washington's Copyright Office is an adviser, not a judge. Its report says so itself, noting that "Dozens of lawsuits are pending in the United States" and that it is "not possible to prejudge the result in any particular case". Britain's own March 2026 report counted over 70 cases involving copyright reportedly commenced in US courts, most of them about fair use, and called the questions unsettled. Some of those courts may decide that training is fair use across the board. If they do, a British buyer that spent months interrogating suppliers about datasets will have bought reassurance against a risk that never arrived, and may have passed over a cheaper or better tool in the process.
Much of that is true, and I would not tell anyone to pick a model on the strength of a regulator's opinion alone. A pre-publication report from 2025 carries persuasive weight, not precedent. Judges will reach different answers on different facts, and the litigation may run for years.
But a buyer's exposure does not depend on who wins the headline case. It depends on the outputs you publish and on the contract you signed. A customer whose generated image closely resembles a protected work has a problem whatever an appellate court later says about the training run, and the first question will be whether the supplier's indemnity responds. Provenance evidence also has value under every outcome. If fair use prevails, a licensed supplier has lost nothing. If it does not, the licensed supplier is the one still standing. That asymmetry is why asking now is cheap insurance rather than premature box-ticking.
It is worth adding that the Copyright Office never called training unlawful; it said several uses are "likely to be transformative", and that analysis or research uses are unlikely to substitute for the works used. Buyers who read it carefully get a sorting tool, not a ban.
EU training disclosure thresholds
Where the Office drew its line
What matters is a single sentence in the conclusion. Making commercial use of vast troves of copyrighted works to produce expressive content that competes with them in existing markets, "especially where this is accomplished through illegal access, goes beyond established fair use boundaries". Fairness, it adds, turns on "what works were used, from what source, for what purpose, and with what controls on the outputs". Read as a buyer, that is a four-part questionnaire.
Scale explains why the questions are hard to answer. Common Crawl, a widely used web archive, is described as holding over 250 billion pages spanning 17 years, a figure the Office quoted in its 2025 report. Elsewhere the report records that some developers license data from publishers and specialist providers, and that researchers at Cornell trained an image model on about 70 million Creative Commons licensed images and claimed comparable performance to a well-known model trained on a substantially larger dataset. So permissively sourced training is possible, at least for some tasks. Whether it suits yours is exactly the conversation to have with a supplier.
On remedies, the Office was cautious. Given the growth of voluntary licensing and the lack of support for statutory change, it said "government intervention would be premature at this time", that licensing markets should keep developing, and that where gaps are unlikely to be filled, extended collective licensing should be considered. In other words, it expects deals rather than legislation to settle most of this. That makes the licence trail, not the statute book, the thing a buyer can actually inspect.
All this came out of a long Copyright Office process. Its notice of inquiry, published in August 2023, received over 10,000 comments by December 2023. Whatever one thinks of the conclusions, they are not a casual aside.
Indemnities arrive with homework attached
Large suppliers have responded to buyer nerves with contractual promises, and they are genuinely useful. They are also conditional, which is the bit that tends to get lost between the sales deck and the signature. Microsoft's documentation, to take one clearly written example, describes its Customer Copyright Commitment as an obligation to defend customers against certain third-party intellectual property claims relating to output content. For Azure OpenAI customers, the documentation adds, coverage also requires that the customer has implemented the required mitigations, including a system prompt directing the model to avoid infringing output and a retained report of testing for reproduced third-party content, with mitigations listed as effective from 1 December 2023. If a claim is tendered, the customer must demonstrate compliance.
I think that is a reasonable bargain, and other suppliers will set their own terms. For a British buyer the lesson is practical. An indemnity is only as good as your evidence that you met its conditions, so someone in your organisation needs to own the configuration and keep the test reports. Better to settle who that is before the first claim arrives than after.
The UK government's own guidance points the same way. Its AI Playbook, published in February 2025, tells public bodies to consider at the outset "how the balance of risk and liability should be determined between the parties, as this will be relevant to any claims for infringement of third party intellectual property". Private buyers can borrow that sentence wholesale.
AI licensing deals involving UK content owners
Disclosure rules are giving buyers something to read
Until recently, asking a supplier about training data often produced a polite paragraph and little else. That is changing on two fronts. In Europe, every provider of a general-purpose AI model must publish a public summary of training content using a template the European Commission issued on 24 July 2025. According to the Commission's guidance, the summary must list the top 10% of scraped internet domains, or for SMEs the top 5% or 1,000 domains, whichever is lower, and from 2 August 2026 non-compliance can bring fines of up to 3% of worldwide turnover or EUR 15,000,000, whichever is higher. Models placed on the market before August 2025 have until 2 August 2027.
California has gone further on licensing status. Its Training Data Transparency Act, signed on 28 September 2024, took effect on 1 January 2026 and reaches systems released or substantially modified since 1 January 2022, requiring developers to say whether their datasets are copyright-protected, purchased, licensed or in the public domain. The UK report summarises the same law as requiring disclosure of whether datasets were purchased or licensed by the developer.
Neither regime was written for British buyers. Both happen to produce documents British buyers can ask for by name. A supplier selling into the EU and California that cannot point you to its published summaries is telling you something.
| Document | When | What it says |
|---|---|---|
| European Commission | 24 July 2025 | every provider of a general-purpose AI model must publish a public summary of training content |
| Training Data Transparency Act | 1 January 2026 | requiring developers to say whether their datasets are copyright-protected, purchased, licensed or in the public domain |
| AI Playbook | February 2025 | how the balance of risk and liability should be determined between the parties, as this will be relevant to any claims for infringement of third party intellectual property |
Britain has chosen to watch the market rather than set the rules
At home, the government has stepped back from legislating. Its consultation drew 11,520 responses, only 3% of respondents supported the opt-out exception it had originally preferred, and the March 2026 report said a broad exception with opt-out was no longer the preferred way forward. On licensing it was blunt: "We propose not to intervene in the licensing market at this stage."
That leaves the market doing the work, and the market is real. The same report cites research showing news publishing accounted for 68% of announced AI licensing deals between March 2023 and February 2025, with 14 deals in 2024 involving UK-based content owners. UK academic publishers reported £335 million of commercial text and data mining licensing revenue in 2021, and in 2024-25 UK collective management organisations represented over 600,000 right holders and distributed over £2.5 billion in royalties. Licences exist to be bought, and some suppliers have bought them.
One line in the report is worth pinning to the wall of any AI steering group. It says AI users, including SMEs and individuals, "also have an interest in ensuring that a model they use has been produced in accordance with the law, and in understanding the provenance of training data". When the government says users have that interest but declines to regulate for it, the job of asking falls to the users.
I am less sure how far to push this for internal tools. A model summarising your own meeting notes produces nothing that competes with anyone's book or photograph, which is close to the analysis-type use the Copyright Office thought unlikely to substitute for training works. Provenance matters most where outputs are expressive and public. Buyers should weight accordingly rather than apply one rule to everything.
What UK technology leaders should ask AI suppliers before 2027
Start with an inventory of which AI tools generate content that leaves the building, from marketing images to shipped code. Those are the uses the Copyright Office's line is most likely to touch. For each, ask the supplier four questions taken almost straight from the report: which works were used, from what sources, under what licences, and with what controls on outputs.
Then ask for documents rather than assurances. That means the EU public summary where one exists, the California disclosure, any licensing arrangements the supplier will describe, and the full text of the indemnity with its conditions. Map those conditions to named owners in your team and keep the evidence, because a defence you cannot document is worth little.
Finally, put provenance into the scoring. Give it real weight for public-facing generation and less for internal analysis. By August 2027 even general-purpose models already on the EU market before the rules began should have published a summary, which gives British buyers a natural date to revisit suppliers that could not answer the first time. Whether US courts will have settled fair use by then is anyone's guess, and that uncertainty is the whole reason to ask now.
An indemnity is only as good as your evidence that you met its conditions, so someone in your organisation needs to own the configuration and keep the test reports.
Where we land
We will keep writing these as we find them. If any of this lands close to a problem you are working on, the team is always happy to talk it through.

