Axel Scheffler considers legal action over Gruffalo copies

You must establish and document consent, provenance and purpose separately for both the source material used to train or prompt an AI system and the output it generates. A single approval to use protected creative work does not cover both.

The prompt is an analysis of 18 October 2026 of the commercial exploitation of AI-generated copies of protected characters, which argues that rights holders can pursue legal action against operators who offer AI-generated imitations of recognisable, copyright-protected work. The case concerns an illustrator considering action against those producing and selling AI copies of a character he created. In our assessment, this exposes a governance gap that affects any organisation using AI to generate images, text or branded content derived from existing creative works: the question of whether you hold valid consent to exploit recognisable output is separate from whether you hold valid consent to use source material in training, and you must be able to demonstrate both.

What makes this case different from the training-data question?

Public debate about AI and copyright has focused heavily on whether protected works appeared in training datasets. This case turns on a second layer: the commercial offering of output that imitates a recognisable character. Those are two distinct legal and contractual questions. The European Commission's explanation of the AI Act requires providers of general-purpose models to supply transparency about training content and copyright, but those rules address the input side only. They do not settle whether you may commercially exploit recognisable output, and they say nothing about imitation of a specific protected character.

The UK impact assessment on copyright and artificial intelligence describes why this distinction matters in practice. Rights holders rarely have visibility into whether and how their work ended up in a model. There is no uniform technical standard to reliably identify works, rights holders and necessary consents. Digital replicas form a separate emerging issue. That opacity means you cannot rely on training-data transparency alone to establish that commercial exploitation is permitted.

Which risks does this create for your workflows?

• Character imitation and passing off — generating and selling output that imitates a recognisable, protected character without consent from the rights holder.
• Scope creep in licensing — holding a licence to use a work in training but not in commercial output, or vice versa.
• Provenance reconstruction failure — inability to demonstrate after the fact what consent was obtained and for what purpose.
• Indemnity gaps — supplier contracts that do not guarantee provenance or permit recall if output infringes.
• Regulatory misalignment — treating compliance with AI Act transparency rules as sufficient to cover exploitation risk.

What must you document before you publish AI-generated output?

1. Identify the rights holder and scope of licence — establish who owns rights in any protected work your workflow touches and what that licence permits.
2. Record model identity and purpose — document which model generated the output, what prompts were used and the lawful basis for using source material.
3. Separate training consent from exploitation consent — obtain and record distinct approvals for using source material in training and for commercial use of generated output.
4. Create a versionable evidence trail — maintain records of prompts, model version, output, publication channel, human review and any notifications to rights holders.
5. Include provenance guarantees in contracts — require suppliers to warrant that output does not infringe and to permit recall on notification; require customers to indemnify you if they use output in breach of licence.

How should your board treat this uncertainty?

The outcome of any legal action is not yet settled, and no court has determined whether a particular AI system acted unlawfully. It is unwise to act now as if AI output automatically infringes or is automatically permitted. Instead, your board can draw up an internal decision that treats consent for training and consent for commercial exploitation as two separate approvals, each requiring its own documentation and sign-off. This approach aligns with the broader track of AI governance and responsible deployment.

What can tooling do, and what remains your responsibility?

Versionable evidence trails, contract templates and audit logs can all be automated or templated. What cannot be automated is the human judgement about whether a particular licence scope covers a particular use, whether a rights holder would consent to imitation, and whether the risk of legal action justifies the commercial benefit of a given campaign. That judgement rests with the responsible person themselves, not with the model or the system. Tooling can make that judgement more informed; it cannot replace it.