The flipside of human-machine co-creation
Thanks to spectacular developments in machine-learning technology, machines are now able to “create.” They are now able to extract existing content and translate it in their own language, make (relative) decisions, adapt to their environment and generate new information.
It’s often said that such quasi-autonomous computational systems can be distressing and concerning. Some initiatives, such as those of UNESCO and Mila, warn against potential risks and remind us of related issues pertaining to ethics, equity, and societal relevance.
In the arts world, when used on a mind-boggling number of works that are often copyright protected, generative AI creative tools (such as ChatGPT, Midjourney, Blender…) raise intellectual property questions that often fall into a gray zone. Faced with potential threats to the recognition of artists’ work, this milieu is reflecting on the subject, and reacting. The WIPO expressed its questioning in an online exhibition about intellectual property and artificial intelligence. Voices are making themselves heard, such as that of the European Writers’ Council. In the United States, a class-action suit was filed against the creators of the Stable Diffusion tool for copyright infringement. Artists suing other artists, claiming to to have suffered prejudice by the integration of machine learning into artworks they created.
Getting into the heart of the controversy, legal news sheds light on some issues, such as this recent decision by the U.S. Copyright Office regarding works generated by machines. With some nuances, it was determined that images produced mechanically with generative AI tools could not be subject to author’s rights, as they are not products of human creativity.
Meanwhile, Canada refers to a certain status quo, reporting that more data is required to properly evaluate this new phenomenon, as demonstrated in this consultation.
To help us better understand the challenges related to this new technology that confronts the oncepts of copyright and authors' rights, some legal and research experts generously answered our questions.
Vast, complex challenges
Due to their systemic and political scope, legal issues related to new artistic practices based on machine-learning are indeed complex.
Éliane Ellbogen, an intellectual property lawyer at Fasken, explains that authors’ rights aim to protect and promote creations of the mind. This embodies a fine balance between fair compensation for artists and accessibility of creative works in the name of research, public interest, and competitiveness on the global market in terms of innovation. Eventual judicial decisions will have implications on the balance of preserving the economic interests of artists, research and industries.
One thing is certain, as leaders in technological machine-learning innovation, industries are in a powerful position to ensure their interests are valued as owners and users of authors’ rights.
Legal challenge 1 - Text mining, an inherent machine-learning practice, involves copying data and therefore, potentially risks violating copyright and authors' rights. Should it be covered under fair use?
According to the Canadian Intellectual Property Office, it is the “sole right” of the copyright owner to “produce or reproduce a work or a substantial part of it in any form,” unless this act can be justified under fair use, the exception to copyright protection.
Collectively, our experts agree that under the current stipulations of Canada’s Copyright Act, text and data mining constitutes a violation when it involves creating a database based on protected works, without the authorization required for their use.
Other jurisdictions, such as Japan, the United Kingdom, France, Germany, and the European Union, anticipated a specific exception for text and data mining. In Canada, the issue is being evaluated to consider the impact of a judiciary decision on artists, research and industries.
“The question hasn’t yet been brought to the courts, to know whether this use would fall under the fair use doctrine. At the present time, the AI industry is lobbying to create an exception to even the law that would permit text and data mining, which would remove the inherent risk of this practice.”- Éliane Ellbogen
Without such an exception, the industry must acquire authorizations to reproduce works for machine-learning, unless these works belong to the public domain. And as noted by Caroline Jonnaert, Ph.D., lawyer and Trademark Agent, Partner at Robic, the use of outdated data could prove problematic “by leading to a model based on content belonging to the public domain, we risk developing models that are seldom updated, as well as perpetuating some biases and prejudices from the past.”
Legal challenge 2 - As modelers of style, generative AI tools appropriate artworks. What happens when machine-generated creations are too similar to the works by which they are inspired?
When artists find their own works within the spaces of these generative AI tools, without required authorization, the question becomes whether these machine-generated artworks are litigious.
On one hand, neither style nor ideas are protected by authors’ rights.
On the other, as demonstrated in a thesis by Tom Lebrun, a lawyer specializing in digital and copyright law, “a work that reproduces, in a substantial manner, the talent and judgment of an initial artwork, that is subject to copyright, that doesn’t belong to the public domain and without authorization, is a litigious work if not covered by fair use.” Furthermore, Caroline Jonnaert notes that using one distinctive element, such as 3 seconds of a song, for example, can be enough to qualify as plagiarism.
There’s a fine line between the legal act of reproducing a style belonging to the public domain, and what might constitute copyright infringement. Therefore, the Canadian jurisdiction has not yet clarified the status of these “stylized” artworks.
Some commercial practices are aware of this, and aim to prevent the overly-similar replication of images, as Open AI did with Dall-E.
Also, the stylized images found in these spaces have a special feature. They reflect the manner in which the algorithm built an abstract representation of data behaviour. Sofian Audry, researching artist and professor at l’École des Médias à l’Université du Québec à Montréal, and author of the book Art in the Age of Machine Learning, explains. “The images of these artists aren’t exactly their images, they embody a representation distributed in space that allows for the generation of other images, which never existed previously. These spaces therefore represent a kind of data compression of two billion images, for example, which instead of being directly stored as digital files, are represented in a network of artificial neurons, using much less memory space, but also with the capacity to generate an infinite quantity of new images.”
How does this emerging trend differ from the past?
The creative act that transforms an existing material is not new. Sofian Audry emphasizes that in the 20th century, the emergence of image- and soundtrack-reproduction technologies facilitated new artistic practices based upon collage and varieties of creations, such as Dadaism.
Yet machine learning techniques are revolutionizing remix technologies. As Sofian Audry explains, “before, we would remix content that already existed, whereas today, with algorithms that transfer style, we remix generative processes.”
Never before seen in the history of humanity, the machine can now automate creation processes with a model that itself becomes transformed, able to be modified and mixed. This new ability brings with it both threats and opportunities, such as the worrying deepfake phenomenon.
As Caroline Jonnaert observes, a new human-machine collaboration is emerging in which the creator can now create the creation; the rules of the creative process from which the work will be created.
Is the legal status quo erasing the artist in favour of the machine, of technology and programmers? Or conversely, is there not a risk in imposing too heavy a burden on the research and industry milieus? Is there a need to modify the law or to create parallel legislation, as other jurisdictions have done?
The law, established in the 19th century and with a certain universal timelessness, has gone through many technological changes without failing. Let us see how it adapts to this new era.