Andrea Stuart

Les informations de ce site Web à été fournie par des sources externes. Le gouvernement du Canada n'assume aucune responsabilité concernant la précision, l'actualité ou la fiabilité des informations fournies par les sources externes. Les utilisateurs qui désirent employer cette information devraient consulter directement la source des informations. Le contenu fournit par les sources externes n'est pas assujetti aux exigences sur les langues officielles et la protection des renseignements personnels.

AI Authorship and Ownership Issues Resolved Using Neighboring Rights Framework

By Andrea Stuart, LLM student, Osgoode PD intellectual property program

In Canadian copyright law, an author is usually automatically vested with copyright ownership.Footnote 1 This becomes problematic when a work is created by artificial intelligence. The following considers the potential candidates for authorship of AI-generated works with minimal or no human contribution (“AI works”); compares copyright ownership in AI works to photographs when the photographer can no longer be recognized as an author; and finally, suggests a way forward for copyright authorship and ownership issues in AI works using the existing legal framework for dealing with copyright and neighbouring rights in performances.

Assuming AI works are copyrightable,Footnote 2 who should own the copyright? Possible owners include: the AI itself; the copyright owners of the copyright in all of the input data used to inspire and inform the AI (“training data”); the computer programmers and developers that created the AI machine’s code (“AI creators”); or the person who made arrangements for the AI machine to generate the work (“AI user”).

Non-human AI cannot own copyright

The AI itself is unlikely to be recognized as a copyright owner. Canadian and American copyright laws agree that copyright can only be granted to a human.Footnote 3 Practical issues arise should these statutes be changed to provide non-humans with copyright ownership. For example, who would represent the non-human in licensing negotiations? What happens to the non-human’s royalties? At present, granting copyright ownership to artificial intelligence for AI works would be at best impractical and at worst a complete market failure.

Training data from different copyright owners: problems of joint-authorship

Granting any portion of copyright authorship and ownership of AI works to those who own copyright in the AI’s training data is problematic. If a large amount of content is inputted into the AI machine, (perhaps every single painting ever made, including everything in the public domain and everything still under copyright,) it is beyond impractical to give so many contributors joint-authorship status. Indeed, joint authorship among just two joint-authors can be problematic.

Further, uncertainty in Canada persists around whether joint-authorship ought to be treated like the British joint tenancy or the American tenants in common – either way, joint-authors could be subject to a tyranny of a minority of owners e.g. one owner could refuse the wishes of the others or one owner could unilaterally make a work available free of charge. Giving joint-authorship of AI works to the owners of the training data is ripe for confusion.

Another issue in regards to granting joint-authorship to the training data owners is determining the impact on the AI work of any single work within the larger training dataset. The more negligible the contribution of a single data point, the weaker the claim for ownership over the final product.

Further, if all the owners of the works included in training data are granted joint-authorship with little regard for how significant their work contributed to the final AI work, then surely joint-authorship should at the same time also be extended to the AI’s creator and the AI’s user whose impacts on the AI output are more direct. It would be a quagmire to vest joint-authorship in the contributors of the training data—complicated even further if the training data includes a combination of public domain and copyrighted works.Footnote 4

Training data from only one copyright owner

Even in the case where all input is sourced from a single author, it is doubtful this would be sufficient for the owner of the training data to claim copyright in an AI’s output.

Developers recently created just this scenario, sourcing all training data from a single author, albeit one that is long dead, in order to build the Rembrandt AI painting machine that produced The Next Rembrandt.Footnote 5 In this situation, if Rembrandt were still alive and his paintings still under copyright, he probably would have no copyright claim in the new work despite the fact that it is entirely inspired by his oeuvre. The new work is not a copy of a Rembrandt painting; the new work is an algorithmic calculation of Rembrandt’s style. It is likely a court would view the new painting as not infringing on Rembrandt’s copyright because copyright does not protect style.

Copyright does, however, protect derivative works. In the hypothetical case of a living Rembrandt suing for copyright infringement in The Next Rembrandt, it could be found that the new work is a derivative work that infringes on Rembrandt’s copyright. Distinguishing between an AI that copies a style while creating an original work and an AI that creates a derivative work would likely depend on specific facts—regardless of whether the AI has simply copied Rembrandt’s style or created an infringing derivative work, if there is no wholesale copying of Rembrandt’s works, it seems impossible to recognize Rembrandt as an author without also extending joint-authorship to the AI’s creator or the AI’s user—which would bring us back to the difficulties of joint-authorship, a solution that would create many more problems.

As a separate issue, Rembrandt might be able to claim copyright infringement, if at any point a copy of his work was put into the AI machine without sufficient prior clearance. In this instance, a fair use defense in the USA might be successful. In Canada, a fair dealing defense would be less likely to succeed until such time that an exception is included in section 29 of the Copyright Act to explicitly allow for AI training data input or section 29’s list of permitted purposes is made to be illustrative, like fair use, with the addition of “such as” language.

Comparing AI to a camera and AI works to photographs

The AI’s creator and the AI’s user are the last two candidates for copyright ownership in AI works. A comparison between the creative process that results in an AI work and the use of cameras to make photographs may be helpful.Footnote 6 An AI’s developer owns the copyright of their code, just like how the Polaroid company owns the intellectual property in their cameras. The camera produces photographs; the AI produces AI works. In the case of cameras, the Polaroid company has no ownership over images created using their machines. Even though a Polaroid camera was used to make the Polaroid photo, “the composition originated with the person behind the camera.”Footnote 7 Like the camera’s creator, the AI creator likely has no claim over the AI work.

The influence of the AI’s user varies considerably when compared to a photographer’s. Given this, it is interesting to consider the following thought experiment and how this would fit into current copyright law. A photographer sets up a camera to automatically produce photos at regular intervals of a changing landscape. Even though the photographs happen automatically and the photographer is no longer required after the initial set up, she is entitled to copyright in the images because she made the arrangements that enabled the creation of these photos, they are original, fixed, and demonstrate that the photographer used some skill and judgement in creating this set up. What happens if the photographer dies and the photos continue to be produced for decades? What happens to the copyright in these new photos as they continue to be produced long after the photographer’s copyright term (life +70 years) has expired? Section 7 (1) of the Canadian Copyright Act deals with posthumous works only where “copyright subsists at the date of the death of the author” – given that photos taken after a photographer’s death do not exist—let alone have copyright subsisting in them at the time of the photographer’s death—Section 7 might not apply.

These posthumous photos are almost “authorless” works. They could immediately enter the public domain. Copyright is not needed here to incentivize the dead photographer to continue to create the photos, on the other hand, without copyright, it is possible the photographer would not have been incentivize to create this set up in the first place. Further, copyright revenue from these authorless pictures might be necessary to enable distribution of the new photos and for the maintenance of the camera equipment in perpetuity.

Applying Work Made For Hire doctrine

Pamela Samuelson and Annemarie Bridy both discuss how the American work-made-for-hire doctrine is sufficient for settling the AI authorship dilemma by “directly vesting ownership of a copyright in a legal person who is not the author-in-fact of the work.”Footnote 8 For Bridy, (depending on the facts of the case)  the AI could be treated like the employee of either the programmer or the user with the net result that the output under a WMFH scheme would vest ownership of the AI work in one of these two as author-in-law.Footnote 9 For Samuelson, the user “possesses the outputs, discovered […] the potential commercial value of the outputs, and is generally best situated to assess and exploit that value”Footnote 10 of AI works and should therefore be granted copyright authorship-in-law and ownership.

Bridy expresses concern that “cut[ing] out the middle-machine […] misses something very important about the nature of these works and the process by which they are produced.”Footnote 11 I would tend to agree: the process by which the work is produced is important and viewing these works through Canada’s performer’s performance rights is a way forward for reconciling current copyright laws with the authorless AI work.

Fixed AI-generated works are like the fixed recording of a performance

For the traditional performing arts, the texts of a play or the musical composition (black ink on a white page) are like a substrate for performance. Though based on a fixed text, every performance comes together in a unique combination of people, circumstances and contexts and with it, in Canadian copyright law, comes performer’s performance rights. Performances can be fixed in recordings which are then a bundle of layered rights and neighboring rights.

Traditional performance has many parallels with the AI creation process. Theoretically and depending on how the AI machine is constructed, the AI’s code is like a substrate or script for the production of a unique work every time the AI is asked to execute its code. The AI code is fixed but everything after this can vary. The AI’s processing of its code is like the performer executing a play’s script during a performance. The language of computer programmers hint that this understanding isn’t entirely foreign to them – segments of computer code are called scripts and when computers work through these scripts, this is called execution --“execution” being a synonym for “performance” in at least one dictionary. In this paradigm, the AI’s creator is like the playwright and the AI’s user is like a producer.

The creative contribution of the AI’s user might be minimal or pivotal, like the creative contributions of a music producers, record labels or TV production companies. In the very least, the AI’s user, like the producer, record label and production company, provide the circumstances for the fixation of the performance. The AI work – the object with which we are concerned given questions about who should be granted copyright ownership over it – is like the recording of the performance. The following chart shows the roles, objects and components of traditional performance and proposed equivalents in AI-generated works.

traditional performance creation of AI-generated works
playwright / composer AI creator
script / composition AI code
theatrical producer / record label AI user
performer AI
live performance execution / processing of AI’s code
recording device AI
sound recording / fixation of the performance AI work

Viewing fixed AI-generated works as a fixation of a performance helps explain both the value and allure of AI artworks. Take for example the AI-created portrait of “Edmond de Belamy.” This work did not recently sell at auction for $432,500 simply because of how it looked but most of its value is likely because it is an AI work.Footnote 12 Similarly, Jackson Pollack’s paintings are considered valuable not just for their aesthetics but because of the artist’s novel process in creating them. Audiences appreciate both these works and perceive value in them in part because the audience knows and understands how these works were made even if they didn’t actually witness the process. Just as Bridy flagged in the aforementioned quote, the process behind the AI-generated work is important: both for audiences’ appreciation of the work and for reconciling these works with modern copyright laws. Dealing with AI-generated works through this performance paradigm allows for the layering of rights in works just as one finds in sound recordings.

Canadian courts could approach AI works through relevant statutes and case law surrounding established licensing norms in the theatrical worldFootnote 13 and the concept of neighboring rights and copyright in performances. Before using AI to create works, an AI user could be required to pre-arrange a production license with the AI creator—either at the time of purchase of the AI machine or at a later time—similar to how a theatre producer would acquire production rights from a playwright or publisher either at the time of purchasing a script or at a later point. Absent any contracts stipulating otherwise, the AI work would vest neighbouring rights in the AI user while the underlying copyright remains with the AI developer.

Current copyright law already contains frameworks that are sufficient, given some flexibility, to accommodate some of the challenges AI works are purported to present for legislators. In a framework where the AI is equivalent to a performer, the AI does not get vested with the rights a human performer would normally enjoy. As mentioned above, only humans can hold copyright. (As an aside, it might be worth further consideration elsewhere whether human performer’s performance rights are an aberration in Canadian copyright law given that they require neither originality nor fixation of the performance.) The layering of many rights in a single work has some challenges. Nonetheless, using this framework may be helpful for navigating this emergent landscape.