Opinion | ChatGPT and other LLMs have created a copyright crisis

Copyright law governs how that is done now. The most powerful large language models depend on vast data sets that encompass, basically, the whole public internet — which, in turn, contains troves of copyrighted material. This raises the question: Does gobbling up this material to train artificial intelligence systems infringe on creators’ intellectual property rights? And if not, should it?

The writers and artists argue that the companies behind these models didn’t ask permission before hoovering up countless people’s life work. Without that work, the models would have trouble mimicking, say, a romance story when queried or depicting the week’s political dramas in the style of a manga. Worse still, many of these models will end up competing with the same humans on whose opuses they were trained. Striking Hollywood writers this summer worried that streaming services would rely on AI to churn out bingeable rom-coms. The writers won that negotiation, putting their crisis on pause. But individual creators won’t always be able to win concessions from employers planning to use artificial intelligence — much less to win them from makers of large language models with whom they have no relationship.

The companies argue back that copyright law is on their side. It isn’t necessarily a copyright violation if someone fails to get permission before using an artist’s oeuvre to produce something new. This is “fair use”: Reproducing a creative work can be legal, so long as the reproduction is also, in its way, creative. Criticism, comment, parody and scholarship, for instance, have always been regarded favorably under the law. More generally, fair use depends on whether the manner in which copyrighted material is adapted is transformative — whether, as the U.S. Copyright Office puts it, the adaptation would “add something new.”

The addition of something new is at the core of the defense of large-language-model makers. They contend that they’re ingesting artists’ writing, drawing and strumming to unleash a wave of innovation. Yes, DALL-E, an AI image generator, can create a faithful riff on Peppa Pig, but it can also generate fantastical landscapes that even the surrealist for which it was named didn’t dream up. The purpose of these models, their overseers insist, isn’t to rewrite Jonathan Franzen novels but, with users’ help, to draft something original or accomplish something functional. Every book by every author is just raw material for building that engine.

All this means that it is likely to be easy for creators to challenge specific AI works that are manifest copycats of their portfolios or, as lawyers say, “derivative works” — that Peppa Pig imitation, for instance. But those creators will have a harder time bringing cases against systems that were trained on their work yet don’t closely imitate it. A judge’s recent decision dismissing part of a copyright lawsuit brought by comedian Sarah Silverman and others against Meta bears this out.

The Copyright Office is undergoing a review of AI systems that could clarify the rights authors have. And companies are, on their own, establishing creator protections. DALL-E 3 won’t respond to prompts asking it to copy a living artist’s style by name, and artists can request to have their work excluded from training data for models to come. Maybe model-makers will also voluntarily license at least high-profile work to avoid the possibility of a lawsuit down the road. A promising example is Axel Springer’s recently signed deal with OpenAI by which the model-makers will pay to use content from the publishing behemoth’s properties, such as Politico and Business Insider, as well as link to those sources when it employs them to answer a question.

According to the Constitution’s Framers, intellectual property rights exist to “promote the progress of science and useful arts.” The latest marvels in coding and computing have introduced unseen forms of creation to humanity. But by producing works faster and cheaper, they could reduce demand for human-painted portraits or human-drafted newspaper editorials — the fuel for the AI engine. That, in turn, could lead to a decline in artistic progress.

A broader rethinking of copyright, perhaps inspired by what some AI companies are already doing, could ensure that human creators get some recompense when AI consumes their work, processes it and produces new material based on it in a manner current law doesn’t contemplate. But such a shift shouldn’t be so punishing that the AI industry has no room to grow. That way, these tools, in concert with human creators, can push the progress of science and useful arts far beyond what the Framers could have imagined.