AI companies should pay creators for training data.

Copying creative work for profit needs permission. Under UK law, copying a work for text and data analysis without the owner's permission is allowed only for non-commercial research. Commercial AI companies copy books, articles, images, and music on a vast scale to build products they sell. Paying for that use is what any other business must do when it builds on someone else's work; indeed, the government said in March 2026 that any reform must ensure creators can be "fairly rewarded" for the value their work creates. (legislation.gov.uk, 2014; DSIT and IPO, 2026)

Creators' livelihoods are at stake. The UK's creative industries contributed an estimated £145.8 billion to the economy in 2024, 5.5% of UK gross value added (GVA). AI systems trained on creators' work can now produce text, images, and music that compete with those same creators for commissions and sales. Without payment, the people whose work made these systems possible would carry the losses whilst the companies keep the gains. (DCMS, 2026)

Paying at scale is possible. Licensing at scale is already happening. In July 2026 a US court gave final approval to a $1.5 billion settlement under which Anthropic will pay about $3,000 for each of 482,460 books taken from pirate libraries. Rights are already licensed in bulk through collective licensing bodies and image libraries, and India has proposed a blanket licence for AI training, with fees passed to rights holders through collective management organisations. A workable system therefore need not mean negotiating with every creator one by one. (Pearl Cohen, 2026; DSIT and IPO, 2026)

AI needs a steady supply of human work. Research published in Nature in 2024 found that AI models trained repeatedly on material generated by other AI models degrade, losing the variety of the original data. As a result, AI depends on a continuing supply of new human-made work. Paying creators helps to keep that supply going, whereas leaving them unpaid risks it drying up. (Shumailov et al, 2024)

Learning is not copying. AI models learn patterns from what they are shown, much as a person learns from reading. In November 2025 the High Court ruled in Getty Images v Stability AI that the Stable Diffusion model does not store or reproduce any of the copyright works it was trained on; its weights are "purely the product of the patterns and features" learnt in training. In the US, a federal judge ruled in June 2025 that training on lawfully bought books was fair use. (High Court, 2025; Barry Sookman, 2025)

Payment would be impractical and small. Models are frequently trained on billions of copyright works, many of them from authors who cannot be identified or traced. Even the largest settlement so far covered fewer than half a million books. Spreading any realistic payment across every creator involved would leave most with very small sums, whilst much of the money would go on administration and to large rights holders such as publishers. (DSIT and IPO, 2026; Pearl Cohen, 2026)

Other countries allow training. The UK government's own report notes that in the US, the EU, and Japan it is lawful, or argued to be lawful, to train on publicly available works under certain conditions, and that models are often trained wherever the rules are most flexible. If the UK alone required payment, developers could train their models elsewhere, and the UK would lose the investment without its creators gaining anything in return. (DSIT and IPO, 2026)

Payment would favour the biggest firms. Only the largest companies can afford licences for training data on the scale that models need. A duty to pay would therefore raise barriers for start-ups, universities, and open-source projects, and concentrate AI in a few wealthy firms. The government has also said that its evidence on how copyright affects AI development is limited and uncertain, which is a weak basis for a new obligation. (DSIT and IPO, 2026)