AI Explained · Why We Fear AI, Part 2
By Dr Peter Fields
THE 2-MINUTE SUMMARY
Commercial training, misuse of someone’s likeness and competition for the same work can all trouble an artist. Each raises a different question.
I favour licensing as the starting point for the commercial use of protected works in training, and meaningful consent for commercial replicas of a person’s face or voice. Those are positions about how things should work. The law gives only partial protection in some cases.
People may enjoy work generated by AI and still care how it was made. Good disclosure gives them the chance to decide what matters to them.
Two people looking at the same picture
A software developer and an illustrator look at an image made by AI. The developer sees a way for someone with little money or technical training to turn an idea into a picture. The illustrator sees a service that can compete for a commission. Suppose some of the illustrator’s work was used to train it. The finished image might contain no recognisable fragment of those pictures, yet the illustrator has a reasonable question about the use made of them.
Training, likeness and competition
Begin with training. I favour licensing as the starting point when protected works are used to train a commercial model, with enough disclosure for rightsholders to exercise their rights. Rights clearance at scale is a serious objection: if every small developer needs to negotiate with thousands of rightsholders, the companies with the largest budgets may be the only ones able to do it. That is a reason to make licensing workable and preserve lawful exceptions. One limited UK exception covers text and data analysis for non-commercial research where the researcher has lawful access and the statutory conditions are met.1
The House of Lords Communications and Digital Committee recommended a licensing-first approach in March 2026. In its own report that month, the government dropped its preference for a broad exception with an opt-out and committed to further work. The committee’s recommendation remains a policy proposal.2,3
A separate concern is a voice or likeness. In my view, licensing a performer’s voice for a commercial replica should require meaningful consent. UK law has no general right to control one’s likeness. Passing off can help in some cases of false endorsement. Performers’ rights and data protection may also be relevant, depending on what was used and how. The government said it would consider whether a digital replica or personality right would help.3
Then there is the lost commission. An imitation can compete with an artist’s work while reproducing no substantial part of a particular picture. Copyright protects expression. An artist’s style, by itself, is generally outside that protection. Licensing the works used for training could address one grievance while leaving the pressure on commissions in place.2
Geek Box · What is in the model?
The files used in training, the numerical weights learnt from them and a generated output are three different things. A model’s weights encode patterns. Some models also memorise material they can reproduce. In Getty Images v Stability AI, the High Court dismissed the secondary copyright infringement claim concerning the models in issue. The primary training claim had been abandoned at trial. The secondary claim is on appeal. In a different case, about song lyrics, the Munich Regional Court held that memorised lyrics had been reproduced in the models themselves. The cases raise related questions on different facts and under different laws.4,5,6
Does it matter who made it?
Porter and Machery found that non-expert readers struggled to tell generated poems from human-written ones. With authorship undisclosed, readers tended to give the generated poems higher ratings. Describing poems as AI-generated lowered ratings in the study. I can see why a reader might like a poem on first reading and feel differently once they know who or what produced it. Both responses are part of the experience.7
The same tool can help a disabled artist realise an idea or let an independent filmmaker try a scene before committing money to shooting it. What I would ask of the person publishing the result is an honest account of any role for AI that would matter to the audience. Content Credentials, developed by the Coalition for Content Provenance and Authenticity (C2PA), can provide a signed record of parts of a file’s history. A missing credential tells you little on its own. Can You Tell It’s AI? and Images, Voice and Video discuss what these tools reveal and where their evidence stops.8
It is possible to value a finished image and care how it came to be. The next part takes that same distinction into a classroom, where the process may be the thing the teacher most wants to see.
Sources and notes
1 Copyright, Designs and Patents Act 1988, section 29A, copies for text and data analysis for non-commercial research.
2 House of Lords Communications and Digital Committee (March 2026), AI, Copyright and the Creative Industries, HL Paper 267, especially paras. 70–75.
3 UK Government (18 March 2026), Report on Copyright and Artificial Intelligence, sections C and J.
4 Getty Images (US) Inc and others v Stability AI Ltd [2025] EWHC 2863 (Ch), especially paras. 543–609.
5 Court of Appeal, CA-2026-000007, Getty’s appeal skeleton argument, paras. 1–3. This records the appellant’s case and the scope of the appeal.
6 Munich Regional Court I (11 November 2025), official account of GEMA v OpenAI, 42 O 14139/24.
7 Brian Porter and Edouard Machery (2024), AI-generated poetry is indistinguishable from human-written poetry and is rated more favorably, Scientific Reports 14, 26133.
8 Coalition for Content Provenance and Authenticity, Content Credentials technical specification, version 2.5 (October 2026).