打印
Generative AI, a technology that generates text, images, music, video, and more through large-scale data learning, is advancing at a frightening pace, with amazing results in areas of creativity that were once thought to be impenetrable to AI. This has led to a number of disputes between copyright holders and businesses providing generative AI technology.
Among them, the lawsuit between the New York Times, OpenAI, and Microsoft regarding LLMs in the U.S. and the lawsuit between Getty Images and Stability AI in the U.K. regarding image generation AI using a diffusion model are noteworthy. In this article, while focusing on these two cases, we will discuss the copyright law issues and implications to consider when utilizing AI.
1. LLM - The New York Times v. OpenAI
1.1 Copyright Law Issues
In late 2023, the New York Times filed a lawsuit against OpenAI and Microsoft in the U.S. District Court for the Southern District of New York, alleging copyright infringement on the grounds that the defendants' GPT-based services were created by copying massive amounts of New York Times articles for training and, under certain prompting, they produced outputs nearly identical to the articles used for training.
The issue of infringement of reproduction rights in collecting data for training was, in fact, somewhat settled by Authors Guild v. Google, Inc. 804 F.3d 202 (2d Cir. 2015). Although this case is not related to AI learning, Google's act of scanning books and storing the scanned copies thereof for its book search service was recognized as fair use, and we believe that the same can be applied to the collection and use of data for training of generative AI.
Unlike book search services, one possibility where fair use might not apply is when reproduction rights are infringed through near-exact reproduction of the learned data. OpenAI argued that the New York Times was able to output New York Times articles verbatim after numerous prompting attempts in a manner that violated its terms, and that this was merely an overfitting problem in AI technology. In other words, because GPT has a “substantial non-infringing use,” the development of, and provision to users of, GPT does not constitute aiding and abetting copyright infringement. The U.S. Supreme Court had previously held that Sony's manufacture and sale of Betamax videotape recorders did not constitute copyright infringement of motion pictures [Sony Corp. of America v. Universal City Studios, Inc. 464 U.S. 417, 456 (1984)].
1.2 Other Issues
Although not at issue in U.S. litigation, our copyright law recognizes the rights of database producers (hereafter referred to as “database rights” for convenience). The Copyright Act makes it an infringement of database rights to reproduce a substantial part of a database or to reproduce it repeatedly or systematically (Article 93 of the Copyright Act), applying mutatis mutandis fair use provisions (Article 94(1), Article 35-5 of the same Act). It is difficult to find a case law to refer to, but if the purpose of protecting databases under the Copyright Act is based on the selection, arrangement, and individual accessibility of data, it is reasonable to conclude that there is no infringement of database rights or that it constitutes fair use unless the set of collected data for learning has those characteristics.
Other interesting issues in the U.S. litigation include whether GPT's so-called hallucinations, in which it describes articles not written by the New York Times as if they were written by the New York Times, constitutes trademark dilution or tarnishment, and whether reducing traffic to the New York Times website by summarizing and showing the website's content in search results constitutes unfair competition.
2. Image Generation - Getty Images v. Stability AI
2.1 Issues
Among the image generation models, the diffusion model is designed to gradually add noise to an image for training and, conversely, learn to remove the noise based on a specific prompt, resulting in an image that satisfies a specific prompt from the noise.
Getty Images provides a service that allows users to search for images and then license and use the images they want, so it has a large collection of high-quality images and text data detailing the content of those images. Stability AI used the image and text data obtained by crawling the Getty Images website for training and made “Stable Diffusion,” which is now one of the most widely used diffusion models for image generation.
Getty Images filed a lawsuit against Stability AI in the United Kingdom in May 2023, and later filed a lawsuit in the United States when jurisdiction issues were raised. The lawsuit also deals with similar issues as the New York Times case. Whether Stability AI can claim fair use with the same logic as OpenAI and Microsoft will likely depend on the facts of the case, and the nature of the work at issue, which is images rather than text, may also have some influence.
2.2 Issues Outside the Lawsuit
While not highlighted in the above lawsuits, there is a widespread distribution of modules in diffusion model communities that are either clearly infringing or highly controversial, even under fair use doctrine. In addition, the widespread use of technologies that either very specifically constrain the conditions of the image to be output (e.g. ControlNet) or use the image itself as a prompt (e.g. IPAdapter) has made the “AI generated it automatically and didn't intend to infringe” defense less viable.
3. Implications
These two lawsuits are representative of the ongoing disputes between creators and AI, and their outcomes are expected to have a significant impact on the AI industry and the business operations that utilize it, warranting close attention.
As the technologies that enable generative AI rapidly advance, become less expensive, and become more popular, more companies are utilizing these technologies, and they need to understand exactly what the legal issues are and what steps they need to take to avoid the associated risks.
While not a legal issue, we also expect to see more social discussions on whether creating with AI can be recognized as art, and how we accept and utilize AI technology.
For complete coverage, please refer to the original article below. Thank you.
The Law Times
View original article▼
https://www.lawtimes.co.kr/LawFirm-NewsLetter/201296
Among them, the lawsuit between the New York Times, OpenAI, and Microsoft regarding LLMs in the U.S. and the lawsuit between Getty Images and Stability AI in the U.K. regarding image generation AI using a diffusion model are noteworthy. In this article, while focusing on these two cases, we will discuss the copyright law issues and implications to consider when utilizing AI.
1. LLM - The New York Times v. OpenAI
1.1 Copyright Law Issues
In late 2023, the New York Times filed a lawsuit against OpenAI and Microsoft in the U.S. District Court for the Southern District of New York, alleging copyright infringement on the grounds that the defendants' GPT-based services were created by copying massive amounts of New York Times articles for training and, under certain prompting, they produced outputs nearly identical to the articles used for training.
The issue of infringement of reproduction rights in collecting data for training was, in fact, somewhat settled by Authors Guild v. Google, Inc. 804 F.3d 202 (2d Cir. 2015). Although this case is not related to AI learning, Google's act of scanning books and storing the scanned copies thereof for its book search service was recognized as fair use, and we believe that the same can be applied to the collection and use of data for training of generative AI.
Unlike book search services, one possibility where fair use might not apply is when reproduction rights are infringed through near-exact reproduction of the learned data. OpenAI argued that the New York Times was able to output New York Times articles verbatim after numerous prompting attempts in a manner that violated its terms, and that this was merely an overfitting problem in AI technology. In other words, because GPT has a “substantial non-infringing use,” the development of, and provision to users of, GPT does not constitute aiding and abetting copyright infringement. The U.S. Supreme Court had previously held that Sony's manufacture and sale of Betamax videotape recorders did not constitute copyright infringement of motion pictures [Sony Corp. of America v. Universal City Studios, Inc. 464 U.S. 417, 456 (1984)].
1.2 Other Issues
Although not at issue in U.S. litigation, our copyright law recognizes the rights of database producers (hereafter referred to as “database rights” for convenience). The Copyright Act makes it an infringement of database rights to reproduce a substantial part of a database or to reproduce it repeatedly or systematically (Article 93 of the Copyright Act), applying mutatis mutandis fair use provisions (Article 94(1), Article 35-5 of the same Act). It is difficult to find a case law to refer to, but if the purpose of protecting databases under the Copyright Act is based on the selection, arrangement, and individual accessibility of data, it is reasonable to conclude that there is no infringement of database rights or that it constitutes fair use unless the set of collected data for learning has those characteristics.
Other interesting issues in the U.S. litigation include whether GPT's so-called hallucinations, in which it describes articles not written by the New York Times as if they were written by the New York Times, constitutes trademark dilution or tarnishment, and whether reducing traffic to the New York Times website by summarizing and showing the website's content in search results constitutes unfair competition.
2. Image Generation - Getty Images v. Stability AI
2.1 Issues
Among the image generation models, the diffusion model is designed to gradually add noise to an image for training and, conversely, learn to remove the noise based on a specific prompt, resulting in an image that satisfies a specific prompt from the noise.
Getty Images provides a service that allows users to search for images and then license and use the images they want, so it has a large collection of high-quality images and text data detailing the content of those images. Stability AI used the image and text data obtained by crawling the Getty Images website for training and made “Stable Diffusion,” which is now one of the most widely used diffusion models for image generation.
Getty Images filed a lawsuit against Stability AI in the United Kingdom in May 2023, and later filed a lawsuit in the United States when jurisdiction issues were raised. The lawsuit also deals with similar issues as the New York Times case. Whether Stability AI can claim fair use with the same logic as OpenAI and Microsoft will likely depend on the facts of the case, and the nature of the work at issue, which is images rather than text, may also have some influence.
2.2 Issues Outside the Lawsuit
While not highlighted in the above lawsuits, there is a widespread distribution of modules in diffusion model communities that are either clearly infringing or highly controversial, even under fair use doctrine. In addition, the widespread use of technologies that either very specifically constrain the conditions of the image to be output (e.g. ControlNet) or use the image itself as a prompt (e.g. IPAdapter) has made the “AI generated it automatically and didn't intend to infringe” defense less viable.
3. Implications
These two lawsuits are representative of the ongoing disputes between creators and AI, and their outcomes are expected to have a significant impact on the AI industry and the business operations that utilize it, warranting close attention.
As the technologies that enable generative AI rapidly advance, become less expensive, and become more popular, more companies are utilizing these technologies, and they need to understand exactly what the legal issues are and what steps they need to take to avoid the associated risks.
While not a legal issue, we also expect to see more social discussions on whether creating with AI can be recognized as art, and how we accept and utilize AI technology.
For complete coverage, please refer to the original article below. Thank you.
The Law Times
View original article▼
https://www.lawtimes.co.kr/LawFirm-NewsLetter/201296