Copyright questions raised by generative AI
A medical writer using generative AI meets copyright questions about what may be collected to train a model, whether its output infringes, whether that output can itself be protected, and what happens when a published article is put into a model.
Three copyright questions
Generative AI raises three separate copyright questions: what data may be collected and used, whether the output infringes, and whether the output can itself be protected. Medical writers both create text and consume text produced by others, so copyright has a lot to do with their end-to-end workflows, and AI is affecting all of it.
One intellectual property lawyer divides where AI and intellectual property meet into three phases: using, typically, someone else's intellectual property as input to build the machine, protecting the machine that has been built, and what happens with whatever the machine outputs. Asking whether a potential problem lies before, during or after the machine is used gives at least a way of thinking about where the intellectual property question is. In the second phase, patent applications for AI-related inventions have been rising steeply. A practitioner who comes across something valuable in their field would do well to have an experienced intellectual property practitioner dissect whether any of it can be protected and made part of a product.
Collecting works to train a model
Large language models typically use vast quantities of books, journal articles, news, blogs, websites and other text, much of it protected by copyright. Some have questioned whether the technology requires the actual reproduction of those works, and the Copyright Clearance Center holds that it does: copying takes place when content is ingested and when it is tokenised, and the models retain training materials, store content in its original form and can recall it. So, in its view, the reproduction right is certainly involved. On the input side, a number of economic rights and probably moral rights are involved too.
The European Union's DSM directive sets rules for using data protected by copyright. The first step is to establish whether the works are protected at all, since only works that show individuality and personal expression can be protected by copyright. Where they are protected, the next step is to acquire the rights to them. The directive's exception then permits protected works to be used where they are needed in research and science. That exception should be seen as narrow: it is meant only for public research institutes and universities and cannot be used for economic research by enterprises. As of May 2024, many regions of the world had laws that make it necessary to clear rights when information is ingested into a model, and the laws affecting the input side came mainly from text-and-data-mining laws made just before AI came to drive the whole discussion.
Besides copyright, plagiarism issues arise on the input side. Trademarks might also become an issue, as those who have worked in dictionary publishing know, because including a trademark in a dictionary can get its publisher into trouble.
Regulation and litigation over training
As of May 2024 most countries had no AI-specific regulation. A major development was the European Union's recently approved AI Act, which emphasises the importance of copyright. Its recital 105 contains a reminder of the general principle that users need permission before using a copyrighted work unless a copyright exception or limitation applies. Its Article 53 requires AI providers to have a policy for complying with European copyright law and to make publicly available a sufficiently detailed summary of the content used to train their systems. The Act's transparency rules, still being worked out in detail in May 2024, reach the use of pre-existing copyright works to train a system. In the United States a bill titled the Generative AI Copyright Disclosure Act 2024 was pending in Congress and would, if enacted, mandate certain transparency requirements for companies offering generative AI trained on protected works.
While legislation was being considered, litigation multiplied: more than twenty cases were brought in the US district courts in the eighteen months to May 2024, over all sorts of content and not only text, and the plaintiffs in many of them claimed infringement on two grounds: the large-scale ingestion of protected content, and the systems' ability to produce output that may contain material identical or substantially similar to the originals. Decisions were awaited in cases such as the Authors Guild's against OpenAI and the New York Times's against Microsoft. A first decision had come earlier in 2024 from a Chinese court, which found a local AI company responsible for copyright infringement through the unauthorised reproduction of copyrighted images.
A model converts the content it ingests into numerical representations held in vector databases, from which it retrieves relevant information, and some see a similarity with what Google Books did with the content it had access to. Whether the two differ is ultimately for the courts to say, since the concept of fair use is at stake, in cases such as the New York Times's against Microsoft over the use of its content in ChatGPT and outputs that reproduce verbatim extracts of its news stories. The Google Books case took fourteen years to resolve, and the judge who found fair use excluded look-up material, such as cookbooks, science articles and maps, where the snippet displayed would itself give away what the user was looking for. The analysis is one of substitution, or interference with the market for the work in question. One lawyer holds that, by excluding look-up material, the Google Books decision is not as strong a precedent as it might seem at first.
Licensing content for training
AI providers can obtain permission in two main ways: directly from individual publishers and rights holders, as some had been doing, or collectively from an intermediary, a collective licensing organisation. The Copyright Clearance Center holds that collective licensing seems to make sense for rights holders and content users in the field of AI too, and expected developments there very soon. It also holds that copyright protection and AI innovation are not in confrontation: AI vendors should be very interested in not putting at risk the human generation of the high-quality content that copyright protects, because language models need such content, which helps with accuracy and with lack of bias.
One company library team that serves its medical writers is very concerned about the training of large language models and who has permission to use those materials, keeps its copyright tools up to date, and takes part in the Copyright Clearance Center's international debate over whether publishers would be able to license material for building AI models.
Infringement by the output
Output may infringe copyright where a generated result contains identifiable parts of other works that are not quoted. A model's sonnet that opens on a summer's day raises the question, although nothing turns on that one because Shakespeare holds no copyright. The problem is not new: in music, for example, two themes can be similar enough to constitute infringement, and it is not easy to say when the line has been crossed without looking at the detail. When AI is used to create works, it can happen very easily.
A different view holds that generative AI synthesises rather than parrots its sources, combining information into new content. On that ground some companies, Microsoft among them, guarantee to cover any plagiarism or copyright infringement arising from output created with their tools, since it will not overlap with the original content.
Text created by a language model arrives without its sources and without quotation, and that is a real risk for transparent scientific work and for publication. Using AI to support an author's own work is not really a problem, provided the part contributed by the natural person prevails, and that is really difficult to establish.
Protection for the output
As of May 2024, work generated by AI could not be protected by copyright in the European Union, because it could not be said to express individuality. The European Union was reassessing the DSM directive and would discuss whether some form of copyright protection should exist even for works of artificial intelligence.
One intellectual property lawyer likewise holds that, for now, the law requires a human inventor or author. Patents had been granted in Australia, China and South Africa although the inventor was ostensibly an AI tool, and he was not entirely sure these would work out, since South Africa runs no patent examination and a grant there may turn out differently once there is a dispute. Some in the patent and copyright fields strongly believe that machines should be registrable as authors and inventors, but he did not expect the law to go there within the next year, while warning to beware of predictions. Whatever one does, he advised keeping the human who will claim to be the inventor or the author, because for legal purposes protection needs at least a collaboration. His image is the union rule that kept a stoker on electric trains with no furnace, offered somewhat provocatively in the expectation that more and more creative work will be done with machine help in future. Where a writer who automates a lot in future generates output that is not protected by copyright, they will have to come up with a business model that gets them paid before any copying can take place.
In the United States, the Copyright Office issued its first report on the question in the summer of 2024, erring towards AI-generated images not being protectable by copyright. At least one graphic novel has text that could be protected by copyright and graphics that could not. The European Union was well ahead of many countries on the point. Until the question is settled, it is a potential concern for anyone using AI to generate any content, and for illustrations and derivative projects especially.
Feeding a model an author's own published work
One publisher cautions that putting content into a third-party platform that uses it to teach its language models compromises the copyright of a publisher that holds the copyright or is licensed to use the content, even while the content is still being developed before submission. That holds even for an author's own publication, as when an author uses AI on it to develop a graphical abstract to be submitted to the same journal. The publisher advises authors to avoid doing so and to talk to the journal editor, who can consult the publisher's ethics team.