A British AI copyright ruling that pleases nobody
The High Court has ruled that an AI company was not liable for copyright infringement for training its image model on material that included content from a major media library. Artists expected a different outcome. AI firms wanted clear protection. Neither got what they wanted.
The case was one of the first UK attempts to address whether training on copyrighted works is lawful. The judgment suggests current law can be read to permit training in some circumstances, but it stops short of offering a stable rulebook. That leaves policy to Westminster.
Why rights-holders are frustrated
Creators see their work ingested, style learned, and market value diluted-often without consent or payment. A narrow win for a single developer does not answer questions about large-scale copying to assemble datasets, or whether specific outputs can infringe. The path to reliable licensing revenue remains unclear.
Why AI companies are uneasy
The ruling is not a blanket permission slip. It does not resolve exposure around outputs that are substantially similar, brand use, moral rights, passing off, or database rights. Nor does it settle obligations for transparency or provenance when models are trained on mixed datasets.
Where UK copyright law feels thin
- Training vs. copying: Is ingesting works to learn statistical patterns a restricted act, and if so, under what conditions can it be excused?
- Text and data mining: The UK exception is limited to non-commercial research. Most foundation model training is commercial. That gap invites litigation and policy action.
- Transient/technical copies: Caching and intermediate copies may be incidental, but mass scraping and dataset creation can exceed that boundary.
- Outputs: Even if training is non-infringing, outputs can still cross the line where they reproduce protected expression.
- Licensing signals: Opt-outs, standard licences, and machine-readable terms are inconsistent or ignored, making compliance hard to verify.
What government should clarify
- Create a commercial text-and-data-mining framework with clear conditions (lawful access, honouring technical measures, audit trails).
- Mandate transparency: dataset source categories, high-level dataset documentation, and a process for rights-holders to query and challenge.
- Codify an opt-out/opt-in mechanism that is technically enforceable and legally meaningful.
- Provide safe harbours for compliant developers and statutory damages or fee schedules for non-compliance to deter free-riding.
- Encourage collective licensing via CMOs so smaller creators can get paid without bespoke negotiation.
Action list for legal teams now
- For developers: Maintain a catalogue of dataset sources, licences, and scraping methods. Record access rights at the time of collection. Implement style-similarity and image-matching filters to reduce output risk. Contractually pass compliance duties to vendors and model providers.
- For rightsholders: Publish machine-readable terms, deploy content protection and watermarking, and register works with fingerprinting services. Offer clear licensing paths for training to convert unauthorized use into revenue. Monitor outputs; focus claims on high-similarity cases.
- For platforms and enterprises: Update procurement terms to require lawful data acquisition, indemnities, audit rights, and dataset provenance summaries. Set review thresholds for high-risk use cases (e.g., style replication, stock-photo lookalikes).
What to watch next
Expect more test cases on output similarity and database rights. Watch for government movement on a commercial TDM exception and a code of practice that standardizes documentation and opt-outs. Until then, contracts and technical controls will carry most of the weight.
Copyright, Designs and Patents Act 1988 and the UK's TDM exception for non-commercial research frame much of the current analysis.
If your team needs fast AI literacy to pressure-test vendors and policies, explore role-based training here: Complete AI Training - Courses by Job.
Your membership also unlocks:
