AI development means increasingly involved human interaction in generative AI models, locating copyrights in that human interaction is key to promote copyright’s true intent, the promotion and dissemination of ideas.

Ensuring that copyright doesn’t protect ideas is not just embedded in the case law and the Copyright Act. It is a fundamental part of what makes copyright law constitutional. Along with the fair use defense, the Supreme Court tells us the idea- expression dichotomy prevents copyright from treading on the First Amendment And it ensures that copyright “promotes the progress of science” by letting people learn and reuse the ideas expressed.20 Indeed, if there are only a limited number of ways to express a particular idea, copyright denies all protection rather than risk fencing the idea in.21 That system worked well in a world where the expression of ideas was hard, and was therefore the thing copyright needed to encourage. No one should be entitled to own the idea of a painting of a comet appearing over the beach at sunset. But everyone is free to express that idea in their own way, and the variations in how they express the basic idea is the stuff of copyright. The hard work of creation— actually painting the brushstrokes, making the individual choices of color, shading and composition that bring the painting to life—the artist gets to own.22 AI automates much of that formerly hard work. I can ask Dall-E or Midjourney to generate a painting of a comet appearing over the beach at sunset and it will take care of the rest in seconds. If I want more specificity, I can ask it to vary what is in the background, or to change the colors, or to render it in a particular style, and it will adapt the painting accordingly. I have to come up with the basic idea and tell it what I want. But the AI does the bulk of the work that copyright traditionally exists to reward. Dall-E 3 will even turn my basic request into a better prompt in order to give you better images.23 And this isn’t just true of paintings. ChatGPT writes text for you. It will even develop plot devices and characters to fit the parameters of a story you give it. Song AIs are taking off, some generating AI covers of existing songs but others writing entirely new songs that feel and sound like the songs of existing artists—or even ones that feel entirely new. So it’s not just the literal elements of the work, but even many of the non-literal elements of the work, that are contributed by the AI. The result is that, increasingly, the things humans contribute in a collaboration with generative AI will be ideas and high-level concepts. AI will contribute the expression. That turns copyright law upside down. Perhaps it means that the AI should be treated as the copyright author, though as noted above current law doesn’t permit that, and none of the AI companies seem to be pushing for copyright ownership in the expression they generate. Or perhaps it means that no one owns the copyright in AI-generated works: prompters don’t own the ideas they contribute and AIs can’t own the expression they contribute. That might make sense as a policy matter—it is not obvious that generative AI needs the incentive of copyright law—but it is unlikely to be politically sustainable. The instinct Rochelle Dreyfuss has derided as “if value then right”24 is quite strong, and I suspect that as the works AI generates become more valuable people and companies will rush in to claim ownership of them. The most common and most plausible answer, I think, seeks to locate creativity not in the generation of outputs from AI but in the human structuring of the prompts that produce those outputs.25 On this theory, Midjourney or Dall-E might generate an original work of art, but it does so in response to specific instructions from the user. While some instructions are simple, often those instructions are iterated, with users refining the questions they ask ChatGPT/large language models or the directions they give image generation AIs until they get the output they are looking for.26 Sometimes they get it right the first time, but users will sometimes engage in “prompt engineering,” going through dozens of refinements to coax an AI to produce just the right output. And there is strong evidence that the way the prompt is phrased or refined can fundamentally influence the output of foundation models.27 Indeed, companies are now hiring “prompt engineers” with the skill of asking the right questions.28 And “jailbreaking” limits on chatbots by asking questions the right way has become a sport in Silicon Valley.