AI and Copyright---Let's Strike An Ac-chord

Yes, we used AI to create this image. Yes, we are aware of the irony. No, we are not sorry. The following article was published by our friend Dana Rao on 30 April 2026, and with his permission, we are sharing it here. Dana is a former General Counsel, engineer, and AI policy leader who helped found the Content Authenticity Initiative and C2PA, built Adobe’s AI ethics engineering programme, and has testified before the US Senate on AI and copyright. He is one of the few people writing on this subject who has shaped policy from the inside. Regular listeners will recognise Dana from our podcast episode last September, where we covered copyright, competition, content authenticity, and Dana’s proposals for a new impressionistic right for creators. This piece develops that thinking further, responding to two live policy moments: the White House’s recent AI framework, which left the question of training data and copyright to the courts, and the UK’s decision to scrap its opt-out system for AI training following creator pushback. Dana’s argument, that the legal focus should sit on outputs rather than inputs, and that a new copy-similar standard could protect artists from competitive harm without stifling AI development, is one of the more coherent and practically grounded positions we have come across in what remains a genuinely unsettled area of law. Over to Dana. The pursuit of balance between rewarding creativity and promoting progress of science in the age of AI remains elusive. Recently, the White House issued a Framework on AI, but left the decision of whether AI can train without permission on copyrighted works up to the courts...Also recently, the UK scrapped its opt out system to allow for AI training due to creator complaints, concluding there was “no consensus on how these objectives should be achieved.” Artists are left on their own to find ways to protect themselves against AI theft, This week, Taylor Swift recently took out novel ‘soundmark’ trademarks of her voice and image to prevent AI likeness thefts. It shouldn’t be this complicated. Artists should be rewarded if people are trading off their works (or likeness) for their own economic gain. If anyone is going to benefit in a new model, it should be the ones being copied, not the copiers. That’s always been the law and it should become part of the law of AI. In practice, that means the legal focus of AI copyright law should remain on the output of generative AI, the part that does or does not create competitive harm to the artist. And not on the input, the training data, which can be used by AI for a myriad of non-copyrighted relevant use cases. That’s the balance that rewards the artists in the traditional way that we expect while allowing AI technology to move forward for general societal good (or bad, that’s not the point of this essay :))---but I do believe AI used correctly can bring tremendous value). How should artists be protected? On the output side, hold the LLMs strictly liable if they generate copyrighted work or work that is substantially similar to the copyrighted work. Substantial similarity is already the test for infringement in US copyright law. If an AI model generates work that infringes this test (like LLMs are apparently able to do right now for Harry Potter novels at least---https://arxiv.org/abs/2601.02671, “in some cases, jailbroken Claude 3.7 Sonnet outputs entire books near-verbatim”), the LLMs should be liable, no intent required. The LLM providers are certainly capable of investing in the filters required to police their output (which they already do in many cases), and LLMs should have that obligation since the LLM providers are choosing the data on which they are training. I personally would expand the penumbra of substantial similarity to include works that create competitive harm to the artist because the copied work is intentionally trading off on the artist’s look and feel, or style (not copyrighted, but ‘copy-similar’) (I just made that up!). Competitive harm was one of the key elements of the fair use test used in the Supreme Court’s Andy Warhol decision and I would rely upon it as part of a new AI copyright analysis for copy-similarity. If a work is so similar to an artist, author, or musician’s work that consumers are buying it to ‘enjoy’ that artist, that feels like an area that traditional intellectual property law was designed to protect against. First, you should have to prove ‘copying’, in this case, that the AI was trained on the artist’s works. Second, you should have to prove intent, that the creator was attempting to pass off the AI work as being derived from the artist. Prompts can be very informative evidence of intent (”create a song that sounds like Taylor Swift”). Third, let the jury decide if the AI work is really a mere attempt to pass off as the artist’s own or was something original. Juries have to make this difficult decision today in copyright cases; no reason to think they can’t do so for copy-similar cases. And you should limit the application of copy-similar protection to competitive harm. If you use AI as a parlor trick for personal fun, no problem. But if you use it to create a ‘copy-similar’ Tayor Swift song that sells millions? That feels like Ms. Swift should get a piece (or a preliminary injunction). There are existing proposals for massive licensing bodies and complicated (and ultimately inaccurate) math to derive artist royalties from AI training data but these seem far-fetched to ever occur in practice and are not focused on the output itself. That feels like the wrong philosophical approach. After all, every artist builds on the influences of others. For humans, those influences are stored in the artist’s brain to be extrapolated and enhanced upon when they create their own original work. That’s how art evolves. If an AI does the same, creates an original work even though its training data includes the work of artists, no royalties should be owed to the AI creator. As long as the song is sufficiently original, it shouldn’t matter if the original song creator was a human or a machine. Once a few cases are decided, just like in every other area of the law, the norms of ‘sufficiently original’ and ‘attempting to pass off’ will be defined for creators and AI providers to understand, protect against, and license, copy-similar works. With these two principles, of promoting the progress of science and preserving the rights of artist to be free of unfair competitive harms, we should be able to fashion a compromise to allow both AI and creativity to flourish. That compromise, however, has to be driven by the policymakers. Inventors invent, creators create, and the government is supposed to find the rules that guide societal interests into a productive place for all. We can’t and shouldn’t leave these complex policy-making decisions to the court, a few enterprising artists, or the ether, merely because consensus is hard. Removing ambiguity and uncertainty is in everyone’s interests; new industries like predictability and artists deserve a roadmap to help them compete in this new age. Thanks for reading Law://WhatsNext! Subscribe for free to receive new posts and support our work.