For more information on this topic or IP-related matters, contact the authors Eliane Ellbogen or Patricia Hénault, members of the Intellectual Property group of Fasken's Montreal office. [1]
This is the second in a series of articles that looks at technology-based art and its intersection with the law, particularly in the fields of intellectual property and privacy law.
The second work discussed is by French artist/curator duo Disnovation (Nicolas Maigret and Maria Roszkowska), whose practice blends contemporary art, research practices, and hacking. Their work generally questions, and even disrupts, what they describe as the “dominant techno-positivist ideologies” that govern the 21st century.
Their recent project, Predictive Art Bot, consists of a machine learning algorithm that turns media headlines into concepts for contemporary artworks, which are displayed, in continuous rotation, on large screens. The algorithm monitors emerging trends among the most prominent and popular news sources, in fields as diverse as politics, environment, innovation, culture, activism, and health. It then identifies and combines keywords to generate concepts for artworks in a fully automated way.
In automatizing the human creative process, Predictive Art Bot caricatures the predictability of Contemporary Art – with a capital C – which is largely influenced by the media; but also, it points to our extreme hypoconnectivity and the echo chambers this creates. Indeed, Predictive Art Bot functions such that only those headlines and hashtags that are massively reposted fuel the algorithm it relies on. This produces, what can only be described as, dull artistic concepts, which ostensibly are the result of an increasingly homogeneous collective imaginary.
Predictive Art Bot raises legal issues about copyright and artificial intelligence, because the originating headlines appear on one of the screens in the installation. They also fuel the “art bot” itself. This poses the question whether a machine learning algorithm infringes copyright in the process of mining and analyzing data, like newspaper headlines.
Predictive Art Bot, Copyright, And Machine Learning
In its most basic definition, copyright infringement occurs when a person, other than the copyright owner and without their authorization (or without the benefit of a statutory infringement exception) reproduces in material form all or a substantial part of a work in which copyright exists. However, there is no infringement if the reproduction is in quality and quantity unsubstantial.
In the case of Predictive Art Bot, the first question is whether there even is copyright in a newspaper headline. This is a relevant consideration because for copyright to exist in a “work,” it must be minimally original (requiring skill and judgment) and not merely utilitarian.
However, in the case of a newspaper headline, a blog post title, or a book title, the answer may not be that obvious. The Copyright Act states that a “work” includes its title, but only when such title is original and distinctive. However, even if a “work” may include the title, the opposite is not necessarily true. In other words, is a mere title, independent of the literary work to which it relates, copyrightable?
To date in Canada, no decision on the merits has found copyright to exist solely in a title. Although it is generally held that a title is not, by itself, proper subject matter of copyright, the courts have nonetheless recognized the possibility that copyright may exist in a title, provided it is sufficiently original and distinctive. For example, in an interlocutory decision, the Quebec Superior Court found that newspaper headlines and lead paragraphs were the fruit of sufficiently creative work to likely warrant copyright protection.
Given this, if we are to assume that copyright may exist in at least some of the headlines fueling the Predictive Art Bot, and assuming that a digital copy of the original headlines is made so that the algorithm may analyze them, then, for the purposes of this analysis, an infringing copy of those headlines in which copyright exists is made.
This is true of pretty much all machine learning processes that use datasets consisting of documents or files, such as literary texts, images, videos, or audio. However, as currently enacted, Canada’s Copyright Act creates uncertainty about the legal implications of key analytical techniques, such as machine learning. There are few statutory exceptions that could apply to AI processes and Canada’s fair dealing regime is limitative as opposed to open-ended, like in the United States.
One such statutory exception that may be used as a defense to copyright infringement in the context of a machine learning process is that temporary reproductions for technological processes do not constitute copyright infringement if:
- The reproduction forms an essential part of a technological process;
- The reproduction’s only purpose is to facilitate a use that is not otherwise an infringement of copyright; and
- The reproduction exists only for the duration of the technological process.
The “technological process” contemplated here is one that would not be able to function, or is otherwise useless, without the reproduction. In other words, the act of reproducing the original work must have no other independent purpose other than to allow the technological process (here, the art-concept generating process) to run its course. And, as the third condition suggests, the copy should cease to functionally exist when the technological process is complete.
However, it is the second condition which may be problematic in the context of machine learning, because the reproduction that is created during the mining and analytical process must not facilitate a subsequent use that may be infringing. For example, if the data that is extracted through a machine learning process is used to create a new work, derived from that data, the derived work may be infringing. In the case of Predictive Art Bot, it could be argued that the data extracted from the original headlines is not used to create a new work, since the constituent parts of the headlines do not find their way into the art concept that is generated by the bot.
As for fair dealing, the outcome of such a defense is uncertain. As of yet, Canadian courts have not taken a clear position on fair dealing in the context of data mining or machine learning. Nonetheless, a work such as Predictive Art Bot may fall under fair dealing, arguably as a form of research.
For the dealing to be fair, it must be done for one of the allowable statutory purposes, for example research, satire, parody, or private study (it is worth noting that there is no fair dealing purpose for artistic expression). Canadian courts have adopted a wide definition of “research,” such that it is not limited to strictly non-commercial or private purposes. For example, it has been held that a computerized form of market research that measured consumer interactions and preferences for the purpose of generating data for clients fell under “research purposes.” Disnovation’s work could possibly be considered research, given the analytical nature of the art-concept generating algorithmic process.
The heart of the fair dealing analysis, however, turns on whether the dealing itself is actually fair. Here, six factors are considered:
- The purpose of the dealing: if the user’s ultimate goal is commercial in nature and there is little to no benefit to the copyright owner or the broader public, the purpose is considered less fair.
- The character of the dealing: whether multiples copies were made and widely distributed.
- The amount of the dealing: the quantitative amount and the qualitative importance of the part copied.
- Whether there are alternatives to the dealing: if so, the extent to which they are expensive, technologically complicated, and market inhibiting.
- The nature of the work: whether the copied work was published prior to the reproduction.
- Finally, the effect of the dealing on the copied work: the market impact the reproduction has on the copyrighted work.
Overall, it could be argued that these factors militate in Disnovation’s favour, given the non-commercial nature of the work, the arguably insignificant impact the reproduction may have on the original headlines, and the minimal amount and importance of the copying.
Conclusion
Given the omnipresence of machine learning algorithms in our daily lives, whether to predict purchasing behaviours, global market trends, or traffic patterns, a work like Predictive Art Bot, which instead makes absurd art forecasts, is an exercise in testing the limits of non-human expression. Here, it is interesting to look at how art, liberated form the constraints of human creativity, will evolve vis-à-vis the relevant legal frameworks.
