Anthropic trains its AI models on British creatives’ work without permission. We should not be celebrating this.

“Data centers and AI infrastructure projects are notoriously capital-intensive but not labor-intensive. They generate headlines and valuations, not necessarily jobs.” Exactly right. Data centers employ vanishingly few people long-term. Politicians suggesting otherwise are lying.

The UK government should only invite Anthropic to expand here if they (i) disclose their training data and (ii) license any of it that is copyrighted. Until then, the government should have nothing to do with them.

AI companies’ argument that they shouldn’t have to reveal the training data they use because it is a ‘trade secret’ is absurd. They are using other people’s work - other people’s intellectual property. People’s right to know when their property is being exploited commercially should *clearly* trump AI companies’ right to keep it secret. If AI companies had to reveal their training data, there would be 1000x more AI copyright lawsuits. That tells you all you need to know. New law is needed to make AI companies reveal the copyrighted work they use to build their models. Governments should enact this law asap. Without it, rights holders are at a huge disadvantage - their work is being exploited at scale behind closed doors, and defending their rights comes down to guesswork. Require AI companies to reveal their training data, and you will end much of the unfair exploitation of people’s life’s work overnight.

The reason it is so important that we introduce transparency laws over AI training data is that AI training is one of the only instances of mass copyright infringement that is mostly invisible to the rights holder. Usually, when someone uses your copyrighted work commercially without permission, you can see it happening. Maybe they modify your picture and use it on a magazine cover; maybe they pirate your film and resell it; maybe they steal your melody and use it in their own song. In each case, the infringement is out there in the open. That means (i) you are unlikely to miss it, and (ii) it is easy to hold the perpetrators accountable in the courts. AI training is different. It happens behind closed doors, and the result - the trained model - might not make the infringement obvious. Models trained on copyrighted work will sometimes output obviously copyrighted work, but, much of the time, the fact that your work has been used in training isn’t obvious. So people’s work has been copied and exploited for commercial purposes, but that copying is hidden. This is why we need new laws requiring AI companies to reveal the training data they use. Training commercial AI models on copyrighted work without permission is straight-up illegal in some countries (e.g. the UK), and there are strong arguments it is illegal in many cases in others (e.g. the US). But rights holders can't defend themselves under the law if the use of their work is kept hidden. The only reason these transparency laws aren't more widespread already is AI industry lobbying. Governments need to resist this, and introduce transparency requirements as a matter of urgency.