High Court Rules That Training an AI model on copyrighted material does not infringe copyright unless the model reproduces or stores those works.

Waboga David
Nov 7
4 min read

The High Court has dismissed Getty Images’ primary copyright claims and database right infringement (the latter abandoned at trial), but delivered landmark guidance on the legality of training and deploying generative AI models under UK copyright and trade mark law.

Key Holdings

AI Training
Training an AI model on copyrighted material does not infringe UK copyright unless the model reproduces or stores those works.
Trade Marks
AI-generated images containing protected marks or watermarks may infringe trade marks under sections 10(1)–(3) of the Trade Marks Act 1994 (TMA).
Model Weights
The Stable Diffusion model weights are not “infringing copies” under section 27 of the Copyright, Designs and Patents Act 1988 (CDPA)—thus, secondary infringement fails.

This is the first UK judgment addressing generative AI training and outputs in the context of copyright and trade mark protection.

BACKGROUND

Getty Images group (US, UK, Irish, and Canadian entities), owners and exclusive licensees of millions of copyrighted photographs, identifiable by their distinct “Getty” and “iStock” watermarks.

Stability AI Ltd, developer of Stable Diffusion, an open-source latent diffusion model trained on LAION-5B datasets comprising billions of web-scraped image–text pairs.

Training Location: Conducted outside the UK using AWS clusters — the training claim was abandoned due to jurisdictional limitations.

Claims pursued at trial:

Trade mark infringement (arising from AI-generated images containing Getty watermarks);
Secondary copyright infringement (alleging that the model itself was an “infringing copy”); and
Passing off.

COPYRIGHT – AI TRAINING AND MODEL WEIGHTS

❖ Training on Copyrighted Works

The Court held that training an AI model on copyrighted material does not constitute copyright infringement under the CDPA 1988, provided that no copies of the training data are stored or reproduced:

“Inference does not require the use of any training data and the model itself does not store training data.” – [§7]

No reproduction or storage: The model’s parameters (“weights”) are mathematical abstractions, not reproductions of underlying works.
Territoriality: Even if training had occurred in the UK, no “restricted act” under section 16 CDPA would have been committed.
Comparative approach: The decision resonates with US fair use reasoning (e.g., Authors Guild v Google) but is rooted in UK statutory interpretation, not fair use principles.

❖ Secondary Infringement (s.27 CDPA)

Getty’s argument that the Stable Diffusion model constituted an “article” containing “infringing copies” failed:

An “infringing copy” requires reproduction of the protected work, not retention of abstracted patterns or parameters.
Therefore, model weights are not infringing copies, even if trained on millions of Getty images.

Result: All copyright claims dismissed.

3. TRADE MARK INFRINGEMENT – WATERMARKED OUTPUTS

❖ Existence of Infringing Outputs

Getty demonstrated that Stable Diffusion could, during ordinary operation, generate synthetic images bearing “Getty Images” or “iStock” watermarks, even when prompts did not mention those brands.

Stability’s defence—that such outcomes were the result of user manipulation or contrivance—was rejected.

❖ Section 10(1) TMA – Identical Mark & Services

Infringement established where generated images reproduced identical Getty marks on photographic content.
Stability was held liable for providing the model to users whose outputs included such marks — constituting “use in the course of trade”.

❖ Section 10(2) – Likelihood of Confusion

A significant likelihood of confusion found: consumers could reasonably believe watermarked AI images were licensed or affiliated with Getty.
Given Getty’s strong brand reputation, the risk of confusion was heightened in the AI context.

❖ Section 10(3) – Reputation and Dilution

Tarnishment: Association with low-quality or synthetic imagery diminishes Getty’s premium brand.
Unfair advantage: Stability gained commercial benefit from the reputational value of Getty’s marks.
Result: Infringement found under s.10(1)–(3).

4. PASSING OFF

Goodwill: Undisputed, given Getty’s 30-year market presence.
Misrepresentation: Watermarked outputs misled consumers to believe images originated from or were licensed by Getty.
Damage: Loss of licensing income and erosion of brand distinctiveness.

✅ Passing off claim upheld (as an alternative basis to trade mark infringement).

5. PRACTICAL IMPLICATIONS

For AI Developers	For Rights Holders
✅ Training on public web data likely lawful (if no storage or reproduction).	✅ Register and protect watermarks and logos as trade marks.
⚠️ High risk if AI outputs reproduce protected marks or watermarks.	⚠️ Monitor AI platforms (e.g., Hugging Face, APIs) for misuse.
✅ Open-source distribution does not insulate from trade mark liability.	✅ Employ technical filters to detect and block watermarked outputs.
⚠️ Must implement watermark-blocking systems within 90 days (per order).	⚠️ Consider licensing agreements for dataset use (e.g., LAION).
✅ Maintain audit trails to show non-reproduction of training data.	✅ Pursue passing off where marks are replicated by AI outputs.

NEXT STEPS

Appeal anticipated, given the novel legal questions surrounding AI model architecture and copyright scope.
Injunction and compliance: Stability ordered to deploy watermark-filtering technology within 90 days.
Damages inquiry pending, focused on potential loss of licensing revenue and reputational harm.

🧩 COMMENTARY

This ruling provides long-awaited judicial clarity for the AI sector, affirming that training alone is not copyright infringement, while emphasizing strict liability for trade mark misuse in AI-generated outputs.

It marks a pivotal moment in the UK’s evolving AI jurisprudence, aligning technical realities with IP protection—and setting a persuasive precedent likely to influence EU, US, and Commonwealth approaches.

Read the full case