Copyright and AI
Copyright and AI is the set of legal and policy disputes over whether and how copyrighted works may be used to train generative models, whether a model's outputs can infringe, and whether AI-generated material can itself be protected. The questions are unsettled and differ by jurisdiction, so this entry describes the arguments and the filed cases neutrally and does not predict how courts will finally decide them. The practical stakes are large for creators, publishers, model developers and anyone deploying generative tools, because training on web-scale text, images, code and audio almost always involves copying works whose rights holders never agreed to the use.
Several distinct legal questions are bundled together and are easy to confuse. The input question asks whether copying works to build a training set, and training on them, is itself infringement or is a permitted use. The output question asks whether a model's generated text or images reproduce or are substantially similar to protected works, or can be induced to regurgitate them, a phenomenon documented in research on extracting training data such as Carlini and colleagues' 2021 paper on large language models. A third question concerns authorship: the US Copyright Office has maintained that copyright requires human authorship, a position upheld in Thaler v. Perlmutter, and issued guidance in March 2023 on registering works that include AI-generated material. A fourth concerns removal of attribution and rights-management information. Different cases emphasize different questions.
Lawsuits began arriving in 2022 and 2023 and span text, images and code. A class action over GitHub Copilot, Doe v. GitHub, was filed in November 2022. Getty Images sued Stability AI in the United Kingdom in January 2023 and in the United States in February 2023, alleging that Stable Diffusion was trained on millions of Getty photographs without permission, with some outputs reproducing Getty-style watermarks. Visual artists including Sarah Andersen brought Andersen v. Stability AI against image-generation companies in January 2023. The Authors Guild and a group of authors sued OpenAI in September 2023. On 27 December 2023 The New York Times filed suit in the Southern District of New York against Microsoft and OpenAI, alleging that its articles were used to train GPT models without license and that outputs can reproduce its content; the defendants dispute the claims. Cases against other developers followed, and rulings and settlements continue to appear, so current status should be checked against court records.
In the United States the central argument is fair use under section 107 of the Copyright Act, a case-by-case balance of four factors. These are the purpose and character of the use, including whether it is transformative; the nature of the work; the amount used; and the effect on the market for the original. Developers point to precedents such as Authors Guild v. Google (2d Cir. 2015), which upheld the copying involved in Google Books search, and Campbell v. Acuff-Rose (1994) on transformative use, and argue that training extracts patterns rather than expression. Rights holders respond that wholesale copying of entire works to build a commercial product that competes with them weighs heavily on market harm, and cite the Supreme Court's narrowing of transformativeness in Andy Warhol Foundation v. Goldsmith (2023). Neither side has a settled answer as a general rule, and courts may reach different conclusions for different models and uses.
Other jurisdictions have taken statutory routes. The EU's 2019 Directive on Copyright in the Digital Single Market (2019/790) created text-and-data-mining exceptions, with Article 4 allowing commercial mining unless rights holders reserve their rights in an appropriate machine-readable way, and the EU AI Act, Regulation 2024/1689, requires providers of general-purpose AI models to maintain a copyright-compliance policy and respect such reservations. Japan's Article 30-4 of its Copyright Act is generally regarded as permissive toward machine-learning uses, while the UK's exceptions are narrower and have prompted consultation. Because rules differ, a model trained in one country and deployed in another can face different standards.
Several misconceptions cloud public discussion. One holds that a model simply stores copies of its training data, and another that it only ever learns abstract patterns; both oversimplify, as models typically generalize but can memorize some material, especially duplicated text. Another is that publicly accessible means free to use, whereas public availability on the web says nothing about license. Finally, many disputes are being resolved partly through business arrangements instead of rulings: the Associated Press and OpenAI announced a licensing agreement in July 2023, and other publishers have struck similar deals, which supports the view that provenance and consent are becoming operational requirements, as discussed under data-provenance.
Read copyright and AI as an old bargain being renegotiated in public: the statutes were written for copying by people and presses, and each lawsuit, licence and regulation is a test of how far they stretch to cover learning by machines.