AI is quickly getting better at generating images.
Copyright law protects your creations from being used or abused by another, without permission. At its core, it is territorial–legally but also psychologically. Humans want to protect our creations, whether that be out of genuine pride or commercial interest.
What happens when the robots get involved?
AI systems now mass digest human work each and every day, using our inputs (books, images, notes) to either produce new media, or train its own models. Copyright rules were built for human authors, yet everything after that initial input into a generative AI system involves no traditional, human author. The High Court’s decision in Getty Images v Stability AI exposed that tension. The real question it leaves behind is not whether the law was wrongly applied, but whether the law and our human instincts are still describing the same thing.
The Copyright, Designs and Patents Act 1988 (CDPA) is the principal statute protecting intellectual property, covering all sorts of work (literary, musical, artistic) and giving creators exclusive rights over the copying and distribution of that material. Two concepts do most of the work in the statute: originality and authorship. Original work is protected, for the author. Both ideas are now confused, with the rise of AI.
Getty Images v Stability AI: Holdings
The following is a very simplified version of the Getty Images judgement, largely ignoring the holdings on trade mark infringement, which will be explored in more depth, in future articles.
The players: Getty Images is a visual media company, supplying many of the stock photos you will have seen online, editorial news pictures, and even music and videos. Stability AI is an artificial intelligence company that develops generative AI models. They are best known for creating ‘Stable Diffusion’, a text-to-image generator, which, by its own account, produces “photo-realistic images”, and “cultivates autonomous freedom to produce incredible imagery”. Users input their requests, and Stable Diffusion draws the picture for them.
The issues: Getty Images alleged that Stability AI had used their images to train its Stable Diffusion model. There were two claims.
The first was of primary infringement. Primary infringement targets direct acts of reproduction–Getty Images alleged that Stability AI used their images, without permission, to train and develop their photo generator model. That initial primary copyright infringement claim was dropped for evidential reasons, as Getty Images could not prove that Stability AI had been training its models in the UK.
The second claim was of secondary infringement, which targets the distribution of articles whose creation involved infringement. Getty Images argued that their material was used to create the images Stable Diffusion fed back to its users. Under sections 22 and 23 of the CDPA, secondary infringement is described as the unlicensed importation of “an article which is, and which he knows or has reason to believe is, an infringing copy of the work”. So, Getty Images had to prove that the AI model Stability AI had trained was both 1) an ‘article’ and 2) an ‘infringing copy’, to fall under this provision.
The ‘article’ hurdle proved low. The world of the CDPA is one of created, often intangible things–literary, dramatic, artistic and musical works. Extending ‘article’ to an AI model stored in the cloud was a manageable step, after some debate in the judgement.
The second hurdle was more difficult. The High Court held that an ‘infringing copy’ must not merely be a product of copying–it must itself be a copy of the work. To be a copy means to “reproduce the work in any material form”, which “includes storing the work in any medium by electronic means” (section 17 CDPA). The model retained no reproduction of the images, and something derived from a work is not, without reproduction, a copy of it. As such, Stable Diffusion was not an infringing copy, and Getty Images failed here.
The Human Response vs The Law:
The most interesting part of the Getty verdict is what it did not address, or could not reach: instinctively, it feels as though AI is copying human work. Even if dressed in hopeful language in some spaces, the instinct that something is being ‘taken’ from us when human work is used as an input is clearly widespread. Perhaps that is a result of misinformation, or equally a mere lack of education about AI. But that instinct may also just be human, attaching to the two core concepts CDPA protects, as mentioned earlier: authorship and originality.
70% of US adults think that artists should be compensated when generative AI uses their work to produce images. The suggestion here is that, without compensation, there is a feeling of ‘theft’. Similarly, artists are disheartened and losing potential income where AI models are processing (and in some ways reproducing) their work more than 400,000 times (in the experience of one interviewed artist, Greg Rutkowski) in under 5 years. Outrage poured from the discovery of Project Panama, an operation by Anthropic, where the company purchased, scanned and then destroyed millions of books, written through the skill and labour of humans, to train its AI chatbot, Claude. Indeed, when Ai-Da, the humanoid robot artist, unveiled her AI-painted portrait of King Charles III at a UK Mission event in Geneva, even the government conceded she raised “timely questions about the nature of creativity, authorship, and the future of art in the digital age.”
That instinct must be balanced with recognition of the legal judgement. The Court’s decision in Getty Images makes sense. The law of copyright protects expression rather than ideas–however original a thought is, it is not protected until put into something that enjoys protection through the CDPA. An art student who internalises a thousand paintings and then paints in their style infringes nothing, as style and technique are ‘ideas’. Generative AI models learn from large amounts of data to discern and replicate patterns and structures. An art student and a generative AI model (like Stable Diffusion) seem like easy parallels.
Why does that latter example feel so uncomfortable?
Authorship. In the CDPA, an ‘author’ is defined in section 9(1), very clearly as “the person who creates”, and cases where work has been computer generated have their own definition, with the author being “the person by whom the arrangements necessary for the creation of the work are undertaken” in section 9(3). Parliament anticipated that machines might produce work, but the machine it pictured was the automated output of the late twentieth century (think: a weather chart, or a formatted report), where a human’s arrangement and labour could still be traced through the finished product, even if computers handled what happened in between. Generative AI breaks that completely. While a prompt is arguably an authorial act, there will be several, unrelated authors, with no clear relationship to one another: the developer who built the model and the company that fine tuned it, as well as the potentially millions of artists used to train the model. Who are we protecting now? Whose work?
When one human learns from another, they do so as equals. Each is a ‘person who creates’, and each is protected in turn. Both are capable of giving back to society in terms of style and technique, because even if they are using the foundations of another’s work, humans have imagination, defined as the power to form a “mental image of something not present to the senses or never before wholly perceived in reality”. When generative AI models train on millions of images, they are mapping statistical relationships between pixels and words. That is not the kind of imagination the human race has grown to respect or admire, as it produces no new style, or idea, drawing from experience or emotion. That is also not the kind of imagination copyright law thought it was dealing with.
This is where the discomfort lies even once you accept the Court’s reasoning in Getty Images. Authorship in the CDPA has always presupposed a traceable chain from a human imagination to a human product. Getty Images had to apply old assumptions about ‘creation’, to a process that has clearly outgrown them.
Originality. The word “original” is inserted into section 1(1) of the CDPA, emphasising the importance of protecting the first, real work. There is perhaps an important distinction to be made between being ‘original’ and being ‘novel’: the work need not be an entirely new invented idea, or something never seen before; work need not be novel. The underlying subject matter may be common knowledge, but the product must have involved some degree of the author’s own intellectual contribution, choice and effort.
For this reason, the courts have traditionally interpreted ‘original’, for copyright purposes, to mean something that results from the skill, labour, judgement and effort of an author. That minimal requirement of skill and labour was clarified more recently: work must be the “author’s own intellectual creation”, suggesting the exercise of free and creative choices. In other words: a personal touch. It seems these tests were never built to apply to computers. On the one hand, the only ‘skill’ or ‘labour’ an AI model possesses is merely a copy of human skill and labour it was taught to do. That skill or labour, instinctively, feels inauthentic, someway different to the skill or labour of a human. At the same time, is not all ‘skill’ and ‘labour’, when exercised by humans, equally the result of being trained by or mirroring other humans?
There are arguably 3 things that separate machine learning from the human learning we have always tolerated in the context of copyright infringement. These explain why, still, the Getty judgement feels uncomfortable.
Scale and speed. Copyright’s tolerance for ‘learning from’ another’s work has always implicitly assumed human limits. Humans likely cannot copy an image 400,000 times in under 5 years.
Sacrifice. A human must also sacrifice something in order to learn. Whether that is time, peace, or money, a human absorbs influence over a period of time where they have actively chosen to learn and take inspiration from something, instead of doing something else. AI models don’t face that limitation, and there is no sacrifice.
Perfection. A human’s learning is filtered through fatigue, memory decay, interest, or even the simple impossibility of knowing a billion facts and images at any one time. An AI model is perfect, with retention far beyond our abilities, and the ability to spit out anything, at any time.
To summarise, the premise in Getty makes sense: an AI model is not a direct reproduction of an image and is therefore not an ‘infringing copy’. But that feels wrong. It feels wrong because, perhaps, humans are just struggling to come to terms with something that learns quickly, perfectly, and without obstacles. We want that new kind of ‘learning’ to constitute ‘copying’, perhaps out of fear. Whether the law needs to adapt to consideration of generative AI models, or whether humans need to accept that AI models are on the same playing field, is left to be seen.
