GenAI: Copyright or Copywrong?
Despite its apparent value, there’s much about the future of AI that remains uncertain. No one knows the precise pace and range of the impending technological shift, and the social and economic impact are equally murky. But even without a crystal ball, we know one factor which will absolutely shape how generative AI is used: copyright law.
Generative AI brings with it a welter of thorny what-ifs concerning the legal status of written output. But while the specifics are tricky, the fundamental questions are simple. Inquiring copywriters want to know “Is material written with AI plagiarism?” and “Is material written with AI protected from plagiarism?”
The answer: it depends where in the world you are — and also when you’re reading this. Until the dust settles, the uncertain copyright status of works written with AI assistance means that caution is in order. That being said, a peek at the present state of things can be the basis of an educated guess on where the law is headed.
The United States
Let’s begin with the U.S. Here, the relevant authority is the U.S. Copyright Office (USCO). In early 2023, USCO launched a public consultation initiative to examine copyright questions raised by AI. It’ll use this initiative as a basis for a series of reports. The first, released on July 31, 2024, concerns the use of AI to impersonate others in voice or appearance. Future reports will cover other aspects, presumably including AI-generated text.
In case you can’t wait that long, the office has issued a policy statement. In a snub to robots everywhere, the office begins by saying, “it is well-established that copyright can protect only material that is the product of human creativity.” The cases that establish this aren’t all technology related: a 1997 decision held that copyright doesn’t apply to work “authored by non-human spiritual beings”, and a 2018 decision held that monkeys cannot register copyrights for photos they take. Truth is stranger than fiction.
That’s not what we really want to know though. What level technological help is compatible with copyrightable work? Citing an old photography court case as a precedent, the office says that it will consider whether AI output is part of an author’s “original mental conception” rather than mere “mechanical reproduction.” That’s vague, but the office takes us through some examples:
No Mere Prompts: If a prompt was the only human input, that doesn’t cut it.
Technical Assistance: Digital tools can be part of the creative process: that was true for Adobe Photoshop, and it’s still true for AI.
Duty to Disclose: When you’re filing for copyright, you need to provide an explanation of what the human involvement was.
Clearly, this is just a starting point. The USCO statement emphasizes that judgments “depend on the circumstances” and is “necessarily a case-by-case inquiry.”
The United Kingdom
Moving to the other side of the Atlantic, the U.K. body responsible for IP rights is the Intellectual Property Office, or IPO. After a March 2023 “Pro-Innovation Regulation of Technologies Review” by one Sir Patrick Vallance, the office boldly declared it would develop a “code of practice.” It’s now over a year later, and no such code has appeared. The U.K. government stance on generative AI is best characterized as spectacularly non-committal.
The BBC editorial guidelines on AI are a bit more fleshed out. AI must be used with “informed human oversight,” must never undermine audience trust, and must be used consistently with editorial guidelines of impartiality, accuracy, fairness, and privacy. Above all, it must be transparent: audiences must be told where AI was used, and why.
India
In India, the foundation of intellectual property law is the 1957 Copyright Act. This law was amended in 1994 to handle computer-generated work. While the act contains no precise definitions for “reproduction” or “copy”, it provides a good rule-of-thumb, and a starting point from which to consider AI-written work.
One of the biggest differences between the U.S. and Indian copyright law is the role of originality. In the U.S., originality is an important, if imprecise, criteria for what can be granted copyright status. In India, “sweat of the brow”, or effort, is required, as is “significant input” from a human being. However, originality, though present, is not emphasized in copyright law as much as it is in Indian patent law.
Will India be a legal refuge for our robot friends? In 2020, India recognized an AI as a co-author of a copyrighted image based on a Van Gogh painting, a copyright that was denied in the U.S. It remains to be seen whether this is a mere registration error, or a precedent that may one day extend to written works.
Japan
When it comes to AI copyright, Japan is marching to the beat of its own drum. According to the draft “Approach to AI and Copyright” released in early 2024, it’s fine to include copyrighted work within training sets — even if the work is pirated! However, you cannot pirate material for “enjoyment.” To be safe, those assembling training sets while in Japan should make sure to frown.
That’s just the status of training data. According to Japan’s General Understanding of AI and Copyright released in May 2024, the chief criteria to consider are “similarity to” and “dependence on” previously existing copyrighted work. These must be present jointly. Having the work of a visual artist in your training set is fine, but it’s a problem if your output looks just like some particular copyrighted item in the training set. This remains true if the person providing the prompt or using the generated image doesn’t notice the similarity.
The European Union
The EU AI Act came into force in August 2024. It’s very broad, and categorizes AI applications by risk levels, from the banned, dystopian category of “unacceptable-risk” to the unregulated “minimal-risk”, with “high” and “limited” risk sandwiched in-between. Generative AI is categorized as minimal-risk. While this may change, it’s unlikely to be moved to the high-risk category, which is reserved for applications which use data to assess aspects of citizen’s personal lives.
The EU follows a familiar pattern: while their framework is quite general, what it means in practice for copyright is fleshed out by individual court cases. The first EU ruling on the copyrightability of AI generated material was a Czech court case. The claimant had generated an image with Open AI’s DALL-E, and accused a defendant of plagiarizing this AI-generated image. The court ruled that the image was “not the unique result of the creative activity of a natural person and therefore the claimant may not be recognized as the author of the AI-generated output.” If this ruling sets a precedent that a prompt alone is not enough, the same will likely hold for AI-generated writing.
International Treaties
Since the copyright status of AI-generated material isn’t yet settled at the level of individual nations (or unions thereof, in the case of the EU), it’s no surprise that there is nothing like an international consensus.
The World Intellectual Property Organization has released a draft summary on IP and AI. Framed as a Q&A, many of its recommendations are fairly similar to those above. Humans, rather than machines, should hold copyright. Having copywritten data within a large training set doesn’t constitute an infringement of copyright.
Some other highlights from the summary:
Though the draft discusses human authorship, it says nothing about human originality.
The draft recommends that there should be a “separate sui generis system of protection” for artistic and literary works generated by AI.
It says that IP policy should consider new rights in relation to data in order to ensure fair competition.
The draft reads more like a to-do list than a policy. This is to be expected, given that it’s the start of a conversation, rather than a worked-out position. Still, you’ve gotta start somewhere!
Conclusion
Common sense is the continuing thread in the nascent legal frameworks emerging worldwide. Details vary, but there’s broad agreement that precisely mimicking some specific copyrighted work isn’t protected, nor is the automated result of one or two lines of prompt. For a work to be copyrightable, some combination of effort, originality, and human oversight is necessary. In the face of these emerging legal structures, the watchword is transparency. If you carefully document and explain your workflow, you are probably less likely to run afoul of nasty legal surprises.
