The UK dropped its TDM opt-out plan
The March 2026 report abandoned the broad text-and-data-mining exception with rightsholder opt-out. Here is the UK status quo that actually applies now.
For most of 2025 the UK creative sector organised around a single fight: a proposed text-and-data-mining exception broad enough to cover commercial AI training, with an opt-out mechanism that would have put the burden of reservation on rightsholders. Publishers built campaigns around it. Trade bodies modelled the cost of running an opt-out register nobody had specified.
The report published on 18 March 2026 ended it. The government abandoned the broad exception with rightsholder opt-out and maintained the status quo.
If you produce content in the UK or buy model capacity from a UK entity, the practical question is what the status quo actually is, because "we kept things as they were" is doing a lot of work in that sentence.
This is a policy summary, not legal advice.
What was proposed, and what happened to it
The proposal was a commercial TDM exception with a reservation-of-rights mechanism: training would be permitted by default, and rightsholders who did not want their work used would have to say so through some machine-readable signal. That model was familiar from elsewhere in Europe, and it was the option the consultation was widely read as favouring.
It did not survive. The 18 March 2026 report dropped it, and no replacement broad exception was legislated in its place. Section 9(3) of the CDPA, the computer-generated works provision, also survived the process untouched and remains contested in its application to generative AI.
Two things follow that are easy to conflate and should not be.
The status quo, stated plainly
There is no general UK exception permitting commercial AI training. The exception that exists is section 29A of the CDPA, which permits copies for text and data analysis for non-commercial research, by a person who already has lawful access to the work. That is narrow by design and was narrow long before anyone was training diffusion models on it. A commercial model trained on UK-hosted copies is not covered by it.
What that leaves is the ordinary position: training on copyright works in the UK either needs a licence or needs to fit an exception, and the exceptions available are small. There is no opt-out register to check, no reservation signal you are required to honour, and no reservation signal you can rely on others honouring.
| Question | UK answer as of August 2026 |
|---|---|
| Broad commercial TDM exception? | No. Proposal abandoned 18 Mar 2026. |
| Rightsholder opt-out register? | No. Never legislated. |
| Non-commercial research TDM? | Yes, s.29A, requires lawful access. |
| Court ruling on training legality? | None. Not decided. |
| Copyright in AI output? | s.9(3) on the books, contested and untested. |
Why "no exception" is not the same as "rightsholders won"
This is the part the celebratory coverage got wrong. Maintaining the status quo removed a proposed permission. It did not create a remedy, and the one UK case that could have supplied one did not.
Getty Images v. Stability AI delivered judgment in the England and Wales High Court on 4 November 2025. Getty dropped its training and output claims mid-trial for want of UK territorial evidence, the secondary-infringement claim failed, and what survived was an extremely limited trade mark finding on watermarks. UK training legality was never decided.
The reason matters more than the result. Training happens where the compute is, and for most frontier models the compute is not in the United Kingdom. A UK claimant has to establish that infringing acts occurred within the jurisdiction before the substantive question is even reached, and that evidential problem sank the claims before the court got to them. An absent exception is only useful to a rightsholder who can get to a merits ruling, and in the UK nobody has.
So the accurate summary is: no permission granted, no liability established, and a territoriality problem sitting between the two.
What actually changed for a UK studio
Almost nothing about your own outputs, and that is worth saying clearly because the news cycle implied otherwise.
- Your ownership position is unchanged. The route to a protectable asset is still human authorship: selection, arrangement and modification recorded as you go. The TDM question was about inputs to model training, not about your rights in what you generate, and the AI copyright and safety guide is the checklist that has not moved.
- Your disclosure obligations are unchanged by this report. The UK has no blanket duty to disclose AI use in advertising. The ASA's position is that the test is whether omission misleads, and advertisers are fully responsible for AI-produced content either way. Your EU-facing obligations are a separate track entirely, covered in Article 50 of the EU AI Act and sponsored content.
- Your supplier diligence is where the exposure sits. If a model's training data is challenged in a jurisdiction where a claimant can reach the merits, the fallout lands on the provider, not on you. But it lands on your production schedule if the model is withdrawn. Licensed training data as a buying criterion is the framing that has held up best.
- If you licence your own catalogue, you have more leverage than the campaign implied. No exception means no default permission. Content owners negotiating with model providers are negotiating against a background rule of "licence or don't", which is a stronger starting position than an opt-out regime would have given them.
Where the pressure moves next
Three places, none of them Westminster in the short term.
Brussels. GPAI copyright-policy and training-data-summary duties under the EU AI Act applied from 2 August 2025, and the Commission's enforcement powers start on 2 August 2026. That is a transparency regime rather than a licensing one, but a published training-data summary is the input every future licensing negotiation and every future claim will start from.
US appellate courts. Thomson Reuters v. Ross held in the District of Delaware on 11 February 2025 that a non-generative legal research tool's use was not fair use, on a direct market substitution analysis. The Third Circuit heard argument on 11 June 2026 and the decision is pending. It will be the first US appellate word on AI-training fair use, and it will be quoted in every jurisdiction including this one.
The licensing market. The deals being signed now are setting the price of catalogue access without waiting for anyone's legislature. That is the mechanism actually resolving this, and a label deal is not a cleared model explains why signing one does not clean a model retroactively.
For a working studio, the practical posture has not moved: pick models you can choose deliberately rather than by default across the catalogue, keep the provenance record on every asset, and treat any provider's training-data claims as something to get in writing rather than something to infer from a blog post.
FAQ
Does this mean AI training on my work is illegal in the UK now?
It means there is no exception permitting it and no ruling establishing liability. Those are two different gaps. The 18 March 2026 report removed a proposed permission; it did not create a cause of action or resolve one. Anyone telling you the UK now prohibits AI training is describing an outcome no court has reached.
Is there an opt-out signal I should be publishing?
There is no UK-mandated one, because the opt-out regime was never enacted. Publishing machine-readable reservations on your own site is still reasonable practice for other reasons, including EU-facing arguments and simple evidence of intent, but do not treat it as a UK legal step, and do not assume it obliges anyone.
Did the report change anything about copyright in AI-generated output?
No. Section 9(3) survived the consultation unamended and unclarified. Its application to generative AI remains contested and no reported decision has applied it to a generative model. Ownership questions for your own work still run through human authorship and through your contract, not through that section.
If Getty lost, does that mean UK training claims are hopeless?
No, and reading it that way overstates the judgment. The training and output claims were dropped for want of territorial evidence rather than decided against Getty on the merits. The lesson is evidential: a claimant needs to show infringing acts inside the jurisdiction, which is hard when the compute sits elsewhere. A better-evidenced claim against a defendant with UK-based operations is a materially different case.