AI, Creative Work, and Intellectual Property, Part I
While my interests focus more on ethical questions about the treatment of AI, a great deal of the current discourse focuses on the potential effects AI can have on creative work — art, writing, acting, and similar activities. In particular, many artists and writers argue that using artwork that they have created (and may have a copyright on) to train AI infringes their rights.1 Also, many creative professionals argue that placing limitations on AI-generated creative works are important in order to preserve professional opportunities for human artists; for example, members of the Writers’ Guild of America and SAG-AFTRA have argued that their current strikes are necessary to prevent TV and movie producers from using AI tools to eliminate most or all of their jobs.2 To say that I’m not an expert on IP law is a serious understatement — it’s not an area of law that I have any formal training in. Moreover, while courts have started to engage with the issues of generative AI and IP, we’re still at a point where it is too early to say where they will end up with any great confidence. However, I still want to give my perspective on the issues generative AI poses for creative work.
In this essay, I argue that the use of art in training a generative AI model does not inherently infringe the rights of the copyright holders/creators of the art, though the outputs of generative AIs can infringe copyright (or other IP rights) in specific cases. I will follow up with an essay tomorrow arguing that the primary limitation on the use of generative AI to displace human creators will be quality of output and cost, not any IP law or contract agreements between creators and producers or publishers, and that to the extent that this transforms the creative industry, we should think of this in terms of other economic displacements due to new technology, not as something unique and special.
Generative AI Does Not Inherently Infringe Because It Is Not a Collage Tool or a Blurry JPEG of the Web
The core argument in claims that the use of generative AI trained using copyrighted material without a license inherently infringes IP rights is that generative AI is a “collage tool” or, to use Ted Chiang’s memorable metaphor, “A Blurry JPEG of the Web.”3 Under this theory, a generative AI text-to-image tool like Stable Diffusion or a Large Language Model (LLM) like Chat-GPT works by storing the training data used in a lossy format — one that doesn’t maintain a perfect copy, but still stores a copy — and then regurgitating a copy or a collage of copies of several different works in response to a prompt. Therefore, any response infringes the copyright of one or more pieces of the training data, because it’s just an imperfect copy of the training data or a collage of a couple of different works in the training data.
As I’ve argued before, this is misunderstanding what these models do. They use the data that they are trained on to learn concepts like what images of a dog look like, what “the style of van Gogh” means, and what words are likely to follow after each other in a paragraph about cars. With an insufficiently large training set, too much duplication, or insufficient computational power, their understandings of these terms can be very similar to specific, pre-existing examples. But as the training set improves and the power and capabilities of the model increase, its ability to generate novel responses — responding correctly to novel prompts by understanding the prompt and producing a new, correct response — improves correspondingly.
In this regard, training a generative AI model is fundamentally similar to the process of human creators consuming artwork to develop their own creative capabilities. Novelists read novels, and painters train by copying or creating new works in the style of existing great works. The screenwriters I know consume TV and movies in vast quantities. And of course novelists also learn from movies, and visual artists learn from written texts. All of that work goes into the blender of the human brain to develop connections between neurons, some of which will then be used to generate new art. The craft of creation is learned both through practice and through consumption, analysis, and frequently mimicry of other creations. When human artists stand on the shoulders of giants, that doesn’t inherently mean that they are infringing the copyright of the artists whose work inspires their creativity.
The Output of Generative AI Can Infringe Copyright (or Other IP Rights) and Raise Some Special Concerns
That doesn’t mean, of course, that the output of a generative AI can’t infringe copyright. A human author can read books by other authors and then produce writing that might be sufficiently different from existing works to be non-infringing or conversely that might copy sufficient elements to constitute infringement. So, too, a specific work of a generative AI might infringe copyright. If I ask ChatGPT 5 to write a space opera, and it turns out a copy of the script of Star Wars with the names changed that I then publish, that’s copyright infringement. How close is too close is the subject of all of the legal wrangling that takes place over allegations of copyright infringement involving human creators. As another example, some people have argued that because tools like Stable Diffusion will occasionally include bogus watermarks in their output, they are inherently acting as a collage tool and thus all of their output is infringing. Instead, that should be understood as recognizing a common feature in its inputs and then replicating that feature, in a way that may make that specific output infringing (although perhaps infringing of trademark, by claiming a false origin, more than copyright, by copying the watermark).
The unique questions raised by generative AI, then, have more to do with specific use cases and outputs than with the fact that they are trained on corpuses that include copyrighted materials. Humans can have a clearer understanding of whether their artistic works are copied from existing works than generative AI has yet demonstrated — though human understanding is highly imperfect, as demonstrated by the fact that Paul McCartney was concerned for a long time about the question of whether he had subconsciously plagiarized the melody of “Yesterday,” before being convinced when enough people told him that they did not recognize the tune that it was an original composition. Current generative AIs are likely to return some infringing works and some non-infringing works in response to prompts that do not inherently ask for infringing outputs.
Moreover, how should we view explicit requests that generative AI produce derivative works? I have asked ChatGPT to write fanfic, and it has without resistance produced (very poor) fanfic. Should that be understood as fair use of the source material if I’m using a tool to produce a derivative work for my own non-commercial consumption? Should that be understood as OpenAI producing infringing derivative works for pay, because my family pays for the ChatGPT 4 service and I am then using it to have ChatGPT produce works that would be infringing if I hired someone to write them for me? Should the quality of the creative work ChatGPT currently produces (low) and the degree of creativity that it employs beyond incorporating creative suggestions from the prompt (also low) matter? We can easily imagine a near-future world where AI can create high quality, bespoke sequels to existing copyrighted works that compete directly with commercial derivative works authorized by the copyright holders. While that may not come to pass — so far, AI’s ability to produce creative work is low — that would raise serious questions of whether those uses infringe copyright and if so who is liable for that infringement. Note, however, that this is all related to the specifics of the output — it’s not that any output by a generative AI is inherently infringing on anything used in its training corpus, but rather that outputs that would be infringing if produced by a human under similar circumstances are similarly infringing if produced by an AI. And while many creators would want extremely broad protection of their art, art done in the style of a specific artist, but not copying specific works, would often not be infringing if done by another human artist trained on the artist’s work, and therefore should not be infringing if done by an AI similarly trained on the artist’s prior work. Copyright law has always sought to balance the interests in protecting creators’ specific creative output with the interests of other creators to produce new creative outputs that have some similarity to the prior art without being infringing. Concepts like the idea-expression dichotomy (copyright protects specific expressions, not the idea embodied in the expression) and fair use serve to operationalize that balancing, and those considerations remain at least as important when applied to AI generated work.
Likewise, a representation of a specific human model or actor in AI generated work may infringe a right to publicity. But that’s cold comfort for actors earning a living (or part of one) as extras, because if an AI uses a corpus of video to learn to create background actors and extras who look like real people, but not like specific real people, it would not be infringing any right to publicity. And if its work would not infringe the copyright those actors might have in their performances, then it likely should not be actionable. But all of this will be very fact specific — how much does that artificially generated character resemble Tom Cruise? To what extent is its acting mimicry of a specific actor’s technique, and is that mimicry enough to infringe a copyright if done by a human actor?
I Leave Open Whether Incidental Copying During Training Infringes Copyright
Finally, I do want to note that there are questions raised about incidental copying in the training process that I leave open and cannot really address without digging deeper into the nitty gritty of copyright law than I care to. Two basic claims can be made that all unlicensed uses of training data is infringing. The first, that I addressed above, is that the output is infringing — because it is derived from the input, it inherently infringes all of the unlicensed inputs that affect its outputs. I believe that is wrong. The second is that in order to use the training data, it needs to be copied and transmitted. The argument goes that downloading a local copy of Wikipedia, for example, for the purpose of submitting that data as a training input (perhaps after creating an edited version by de-duplicating blocs of text that are repeated multiple times) represents an infringing act of duplication. The counter-argument is that where that data is deliberately made available for free access or otherwise permissibly accessed, incidental local copying in the course of processing that data is fair use. As a non-expert, that seems likely correct to me, but the contours of fair use are sufficiently unclear to me that I don’t want to express a firm opinion.4
Tomorrow, I’ll turn to discussing the broader economic implications of the potential for generative AI to compete with human creative production.
1See, e.g., David Simon, creator of The Wire, “When I sold 150 hours of television to Time Warner HBO, nothing in the contracts suggested the material could be mixed with material from other writers and vomited back as product. Fuck AI.” https://twitter.com/AoDespair/status/1682436852584685571 (July 21, 2023); Andersen v. Stability AI Ltd., 23-CV-00202-WHO (N.D. Cal.) (civil litigation alleging that Stable Diffusion and Midjourney infringe the copyright of artists by functioning as collage tools that reproduce artists’ work).
2See, e.g., “Justine Bateman on AI, Labor, and the Future of Entertainment,” Justine Bateman interviewed by Justin Hendrix, Tech Policy Press (July 23, 2023), available at https://techpolicy.press/justine-bateman-on-ai-labor-and-the-future-of-entertainment/ (Q: “[T]o some extent, you imagine the executives who are running these media and film companies [thinking]… the perfect price of content is zero?” A: “Yes. And it’s never been closer to doable.”; “I believe 2023 is the last time that any of [the DGA, the WGA, or SAG-AFTRA] will have leverage on their own.”)
3Ted Chiang, “Chat GPT is a Blurry JPEG of the Web”, New Yorker, https://www.newyorker.com/tech/annals-of-technology/chatgpt-is-a-blurry-jpeg-of-the-web (Feb. 9, 2023). I’ve written about and argued against Ted Chiang’s position before, at https://ahmorse.medium.com/llms-and-artificial-general-intelligence-part-v-counter-arguments-the-argument-from-design-and-28b427afa08d (July 12, 2023).
4For what it’s worth, my opinion that the collage tool arguments are wrong but the incidental copying/fair use arguments are more challenging seems to match Judge Orrick’s tentative ruling in Andersen. See Copyright Lately, post of July 20, 2023, available at https://www.linkedin.com/feed/update/activity:7087569348604694528/. That said, even if Judge Orrick issues a definitive ruling in this case, it will be the first of many rulings before these issues become settled as a matter of U.S. law, let alone more broadly.
