Introduction

In recent years, artificial intelligence (AI) and more specifically generative AI (GenAI) has developed at an exponential rate. GenAI is AI that is capable of generating content, such as text, images, videos or audio when prompted by a user. GenAI models are often trained through a process called “web scraping”, which refers to the process of extracting data from websites. They learn patterns, relationships and knowledge from this training data and when given a prompt, they can reference tens of millions, if not billions, of housed data sources in seconds to provide a targeted output. In doing so, GenAI models have the ability to generate unique results.

The use of GenAI has raised novel issues with respect to copyright ownership, including use of unlicensed copyrighted works in training GenAI models and in the outputs created by GenAI. Canadian courts have not held that web scraping for the purpose of training AI is fair dealing.^[1] With respect to fair use^[2] in the United States, the United States District Court, Northern District of California recently considered a case^[3] where a social media company (X Corp.) attempted to bar a data collection company from extracting and copying public data from its platform since its terms of use prohibit (among other things) web scraping without its prior written consent. ^[4] The United States District Court stated that imposing strict contractual web scraping restrictions would frustrate the operation of the fair use doctrine.^[5]

It can be expected that these types of fair dealing or fair use issues will continue to develop in case law, given that courts in Canada and the United States have considered their respective frameworks in the context of other types of technology. For example, in 2011, the BC Supreme Court ^[6] applied the fair dealing framework in the context of a text and data mining activity (being a process through which large amounts of information can be analysed electronically). In that case, a web-crawler (i.e., an automated program or bot that systematically searches websites and indexes the content on them) gathered text and photos from websites to populate the defendant's website. The court ultimately found the activities were infringing. In a more recent decision from the United States,^[7] the US Supreme Court considered a case regarding the appellant’s use of certain application programming interfaces (APIs) and lines of source code, owned by the respondent, within early versions of the appellant’s operating system. The US Supreme Court ultimately ruled that the appellant’s copying of the APIs was a fair use of that material as a matter of law.

Use of unlicensed copyrighted works as training data

There has been significant discourse in recent years regarding the unlicensed use of copyrighted material in training AI models. Many creators are concerned that AI web scrapers are training on and absorbing their original works and consequently violating their copyright. This issue recently came to the legal forefront in a pre-trial motion to dismiss a class action in Sarah Andersen, et al. v. Stability AI Ltd., et al. (Andersen), dated August 12, 2024.^[8]

In Andersen, the plaintiffs alleged that the defendants infringed their copyrights by using their works without their permission as “training images” for the defendants’ GenAI model named Stable Diffusion.^[9] Five billion images were allegedly scraped into datasets to train Stable Diffusion.^[10] The plaintiffs contended that, as a result of web scraping and training on data that included their copyrighted works, Stable Diffusion can produce images replicating their unique artistic styles.^[11]

On a prior pre-trial motion to dismiss,[^12] the United States District Court had found that the allegations for direct copyright infringement (i.e., using copyrighted works without permission in training Stable Diffusion) were sufficient to allow the direct infringement claims to survive dismissal since Stability AI Ltd. allegedly scraped copies of the plaintiffs’ copyrighted works from the internet and used those works to train Stable Diffusion.^[13]

Among other claims, in the most recent Andersen motion, the plaintiffs asserted induced copyright infringement (i.e., intentionally encouraging or causing another person to directly infringe on a copyrighted work by distributing Stable Diffusion to any third party that downloads, uses or deploys it). The United States District Court noted that Stability AI Ltd.’s CEO had stated that Stability AI Ltd. had compressed thousands of images into a small file and can “recreate” any images from that file.^[14] The United States District Court further stated that it is plausible to infer that Stable Diffusion, when operated by end users, “creates copyright infringement and was created to facilitate that infringement”^[15] and so the allegations of induced infringement were also sufficient to survive dismissal.

Now that the direct and induced infringement claims have survived their respective dismissal motions and been given the green light to proceed towards discovery and eventually trial, the impending judgment could set a crucial precedent for similar cases in the United States. It could also be considered by courts globally.

On one hand, if the plaintiffs’ arguments are accepted, web scraping for GenAI purposes will likely become illegal. On the other hand, if the infringement claims are ultimately dismissed at trial and Stable Diffusion’s web scraping of copyrighted works is validated, then this will have a substantial impact on copyright, its protection and its value to the copyright owner.

Notably, if the defendants in the Andersen case are successful at trial, copyright owners may soon more widely use digital watermarking to protect their copyrighted works. Digital watermarking is the process of embedding code (i.e., the watermark) into a digital asset (e.g., a graphic, audio or video) in order to provide copyright information. The watermark is typically undetectable during normal use of the file, but it can be detected by a computer algorithm to prove the authenticity and integrity of the digital asset.

AI creating potentially copyrightable works

In December 2021, the Canadian Intellectual Property Office (CIPO) granted a copyright for a “Starry Night-inspired” painting titled “Suryast” through its automated copyright registration process. To create Suryast, an individual named Ankit Sahni (Sahni) prompted a GenAI painting tool called RAGHAV Artificial Intelligence Painting App (RAGHAV) to combine Vincent Van Gogh’s painting, “The Starry Night,” with a photograph that Sahni took of a sunset. When submitting the Canadian copyright application, Sahni listed himself and RAGHAV as co-authors.^[16] Accordingly, since CIPO granted the application, GenAI is currently listed as an author of a Canadian copyrighted work.

In July 2024, the Samuelson-Glushko Canadian Internet Policy and Public Interest Clinic (CIPPC) submitted a notice of application under section 57(4) of Canada’s Copyright Act,^[17] seeking an order for the rectification of the Register of Copyrights.^[18] CIPPC claimed that Suryast should be struck from Canada’s copyright registry because there is no copyright in the image (or, alternatively, that the registration should be amended so Sahni is the sole owner).^[19]

CIPPC alleged that the image “lacks originality” and so it does not satisfy the relevant copyright requirements.^[20] Contending that Suryast was created with GenAI through “a purely mechanical exercise of data entry and algorithmic luck” with “no exercise of human skill or judgment,”^[21] CIPPC submitted that CIPO “[derogates] from its obligations to administer copyright in a fair and balanced manner” by instantaneously accepting copyright registration requests.^[22]

Canada’s Federal Court may follow the United States’ Copyright Review Board’s decision regarding Suryast, i.e., that AI systems cannot be authors for the purpose of copyright in the US and that Sahni’s input was insufficient to rise to the level of human authorship.^[23] However, if Canada decides that copyright protection does extend to AI-generated work, then a number of other intellectual property-related issues will need to be considered, such as who or what can own an AI-generated work and whether an AI system can license or assign its copyright or waive its moral rights to the copyrighted work.

Conclusion

These disputes illustrate the growing concerns that persist around ethical and legal boundaries in the development and use of GenAI. Despite ongoing attempts to appropriately regulate AI,^[24] concerns regarding copyright (and other intellectual property rights) remain prevalent. Many courts are on the cusp of rendering verdicts that could vastly impact the intellectual property landscape. Despite being useful in many capacities, GenAI’s exponential growth is currently surrounded by legal controversy.

If you have any questions about artificial intelligence or intellectual property and their intersection, please reach out to the authors, Tom Sides or Jaclin Cassios, or any member of Dentons’ Intellectual Property and Technology group. Thank you to our articling student Jake Katz for his significant contribution to this insight.

[1] Fair dealing is a provision in the Canadian Copyright Act, RSC 1985, c C-42, that allows the unlicensed use of copyrighted works in certain circumstances.

[2] Fair use is a legal doctrine in the United States that allows unlicensed use of copyrighted works in certain circumstances.

[3] X Corp. v Bright Data Ltd. 2024 WL 2113859 (ND Cal).

[4] Ibid at 1-3.

[5] Ibid at 23.

[6] Century 21 Canada Limited Partnership v Rogers Communications Inc., 2011 BCSC 1196.

[7] Google LLC v. Oracle America, Inc., 141 S. Ct. 1183 (2021).

[8] 2024 WL 3823234 (N.D. Cal.).

[9] Ibid at 1.

[10] Ibid at 2.

[11] Ibid at 1.

[12] Andersen et al. v Stability AI et al., 700 F Supp (3d) 853 (N.D. Cal. 2023).

[13] Supra note 8 at 3-5.

[14] Ibid at 8.

[15] Ibid at 9.

[16] Under Canada’s Copyright Act, RSC 1985, c C-42, all denoted co-authors share the rights and privileges which are associated with a validly held copyright. More specifically, Section 9 of the Copyright Act pertains to joint authorship and states that “copyright subsists during the life of the author who dies last, for the remainder of the calendar year in which that author dies, and for a period of 70 years following the end of that calendar year” (i.e., the “life plus 70” rule). In this case, RAGHAV is not a natural person and cannot die in the same way a natural person can. This raises the issue as to how the “life plus 70” rule would be determined, especially since CIPO did not make that determination when it initially registered the copyright.

[17] RSC 1985, c C-42.

[18] Samuelson-Glushko Canadian Internet Policy and Public Interest Clinic v Ankit Sahni, (8 July 2024), Ottawa, FCC T-1717-24 (notice of application).

[19] Ibid at para 1.

[20] Ibid at para 30-31.

[21] Ibid at para 32.

[22] Ibid at para 23.

[23] Copyright Review Board, “Second Request for Reconsideration for Refusal to Register SURYAST” (11 December 2023), online (pdf): <https://www.copyright.gov/rulings-filings/review-board/docs/SURYAST.pdf>.

[24] For example, Canada has implemented a “Voluntary Code of Conduct on the Responsible Development and Management of Advanced Generative AI Systems”, which provides Canadian companies with common standards and enables them to demonstrate (voluntarily) that they are developing and using generative AI systems responsibly until formal GenAI regulation is in effect: (8 November 2024), online: <https://ised-isde.canada.ca/site/ised/en/voluntary-code-conduct-responsible-development-and-management-advanced-generative-ai-systems>; and “Artificial Intelligence and Data Act” (29 September 2023), online: < https://ised-isde.canada.ca/site/innovation-better-canada/en/artificial-intelligence-and-data-act>.