Image: Pixabay
Last month I highlighted the first AI/Copyright case in Canada to reach the courts, CanLII v CasewayAI. CanLII, (the Canadian Legal Information Institute), a non-profit established in 2001 by the Federation of Law Societies of Canada, sued Caseway AI, a self-described AI-driven legal research service, for copyright infringement and for violating CanLII’s Terms of Use through a massive downloading of 3.5 million files which Caseway allegedly used to populate its AI based services. Now the principal of CasewayAI, Alistair Vigier, through an article (Don’t Scare AI Companies Away, Canada – They’re Building the Future) published in Techcouver, has responded publicly by trotting out many of the tired and specious arguments put forward by the AI industry to justify the unauthorized “taking” of copyrighted content to use in or to train generative AI models. Let’s have a closer look at these arguments.
Vigier opens by referencing another AI/Copyright case in Canada where a consortium of Canadian media companies is suing OpenAI for copyright infringement. He claims this is all based on a misunderstanding of how AI training works, stating that “AI systems like OpenAI rely on publicly available data to learn and improve. This does not equate to stealing content.” Whether data is “publicly available” or not is irrelevant when it comes to determining whether copyright infringement (aka stealing content) is concerned. Books in libraries are publicly available, or so is a book that you purchase in a bookstore, or content on the internet that is not behind a paywall. (It is worth noting that the Canadian media companies also claim that OpenAI circumvented their paywalls to access their content when copying it). But in none of these cases is copying permitted unless the copying falls within a fair dealing exception, which is very precise in its definition. Labelling copied material as “publicly available” is a red herring.
Vigier’s next argument is to equate the ingestion of content by various AI development models with a human being reading a book. We know that humans enhance their knowledge through reading and are thus able, presumably, to better reason based on the content they have absorbed. Vigier says, “This is how AI works. The AI “reads” as much as it can, gets really “smart,” and then explains what it knows when you ask it a question. Like a human learns from reading the news, so does an AI.”
Really? A human does not make a copy, not even a temporary copy, of the content although some elements of the content are no doubt retained in the human brain. But AI operates differently. It makes a copy of the content. This should be beyond dispute although the AI industry continues to muddy the waters by claiming that when content is “ingested” it is converted to numeric data and is thus not actually copied. This is a fallacious argument. Just because the form changes, this does not mean there is no reproduction. When you make a digital copy of a book, there is still reproduction even though the digital form is different from the original hard copy version. When a work is converted to data, the content is still represented in the dataset.
Vigier dubiously states, with regard to OpenAI, “OpenAI’s models do not reproduce articles verbatim; they process vast datasets to identify patterns, enabling insights and efficiency.” Apart from the fact that the New York Times in its separate lawsuit in the US has been able to demonstrate that by typing in leads of articles, it can prompt OpenAI to reproduce verbatim the rest of the article (OpenAI claimed that the Times “tricked” the algorithm), copying is copying even if the result of the copying is somewhat different from the original. The Copyright Act is crystal clear on this point. Section 3 (1) of the Act states that, “For the purposes of this Act, copyright, in relation to a work, means the sole right to produce or reproduce the work or any substantial part thereof in any material form whatever…“. If copyright protected content is reproduced in its entirety without permission for a commercial purpose (eg for AI training), that is infringement, unless the use qualifies as a fair dealing under Canadian law or fair use in the US.
The issue of whether ingestion of content to train an AI application results in copying (reproduction) has been carefully studied and documented. One of the most thorough examples is a recent SSRN (Social Science Research Network) paper, entitled, “The Heart of the Matter: Copyright, AI Training, and LLMs” with noted scholar Daniel Gervais (a Canadian by the way) of Vanderbilt University as lead author. The article goes into a detailed discussion on how copying of content occurs during AI scraping to build a Large Language Model (LLM), including the stages of tokenization, embedding, leading to reward modelling and reinforcement learning. The section of the article explaining how copying occurs (pp. 1-6) is dense, technical text but the conclusion is clear, “LLMs make copies of the documents on which they are trained, and this copying takes various forms, and as a result, with appropriate prompting, applications that use the LLMs are able to reproduce original works.” A shorter (and earlier) version explaining how the LLM copyright process works can be found in this article (“Heart of the Matter: Demystifying Copying in the Training of LLMs“), produced by the Copyright Clearance Center in the US. It is also worth noting that these explanations refer only to ingestion of text. AI models that train on images and music are even more likely to produce exact or close-to-exact reproductions of some of the works they have been built and trained on.
So much for the misinformation in Vigier’s article. Now to the scare tactics. He says that the recent Canadian media lawsuit against OpenAI sends a negative message to innovators that Canada may not be open to AI development.
“If Canada wishes to remain relevant in this (AI) sector, it must balance protecting intellectual property and promoting technological progress.”
The fact that there are currently more than 30 lawsuits in the US, including the seminal New York Times v OpenAI case, does not seem to have slowed down the AI companies in the US. In the UK, legislation has been introduced that would, according to British media reports, “ensure that operators of web crawlers (internet bots that copy content to train GAI, generative AI) and GAI firms themselves comply with existing UK copyright law. These amendments would provide creators with crucial transparency regarding how their content is copied and used, ensuring tech firms are held to account in cases of copyright infringement.” There is lots of AI innovation ongoing in Britain.
The Australian Senate Select Committee Report on Adopting AI has recommended, among other findings, that there be mandatory transparency requirements and compensation mechanisms for rightsholders. The EU is already way out in front on this issue. Its new AI Act stipulates that providers of AI generative models will be required to provide a detailed summary of content used for training in a way that allows rightsholders to exercise and enforce their rights under EU law. Even India now has its own version of the US and Canadian media cases against OpenAI. (OpenAI’s defence in part is based on the argument that no copying took place in India because no OpenAI servers are located there!)
If that is what the “competition” is doing, who does Vigier cite as being the jurisdictions most likely to attract innovators away from Canada? Why, it is those AI powerhouses of Switzerland, Dubai—and the Bahamas!
The argument that if legislators and the courts don’t give AI innovators a free pass on helping themselves to copyrighted content for AI training purposes, this will either slow down innovation or chase it elsewhere is a common fearmongering strategy of the AI industry. This is a race-to-the-bottom mentality whereby content industries are thrown under the AI bus. Vigier, having been the subject of his own lawsuit, argues that instead of resorting to litigation, the Canadian media companies should have sought a licensing solution. But the fact that no licensing agreement was reached with OpenAI is undoubtedly the reason for the lawsuit in the first place. That is certainly the reason behind the NYT v OpenAI lawsuit in the US; licensing negotiations broke down. If someone has taken your content without authorization, and then offers you pennies on the dollar in comparison to what that content is actually worth, then the stage for a lawsuit is set.
In explaining CasewayAI’s position in the litigation brought by CanLII, Vigier says that Caseway approached CanLII with an offer to collaborate but was rebuffed. As a result they developed other extensive web crawling technology that pulled the needed material from elsewhere. (Where exactly the material was downloaded from is the crux of the matter). Regardless, this makes it sound as if it was CanLII’s fault for refusing to share their content. Surely a rightsholder has the right to determine the terms on which their content is to be shared with others, if at all.
The fact that Caseway went to CanLII in the first place suggests that CanLII had developed the content that Caseway wanted. Caseway claims the material it accessed was on the public record, such as court documents and decisions. CanLII, on the other hand, claims that it had reviewed, indexed, analyzed, curated and otherwise enhanced the content in question, thus adding a wrapping of copyright protection to what otherwise would be public documents. Who is right, and whether the material was scraped from CanLII’s website without authorization, will be determined by the BC Supreme Court.
If the material taken by CasewayAI was not copyright protected, they are in the clear, at least with respect to copyright infringement. That is quite different, however, from arguing that no copying takes place during AI training or that if rightsholders use the courts to protect their rights, Canada will be a laggard when it comes to AI development. Robust AI development needs to go hand in hand with robust copyright protection for creators, with an appropriate sharing of the spoils of the new wealth generated from the creative work of authors, artists, musicians and other rightsholders. To say, as Vigier does in his concluding paragraph that;
“Canada has a choice to make. Will we embrace AI as the transformative force it is, or will we let fear and litigation stifle innovation? The lawsuits against Caseway and OpenAI message tech companies: you’re not welcome here. If this continues, Canada won’t just lose its AI startups; it will lose the future of job creation.”
What sheer self-interested nonsense!. This is fearmongering of the worst kind, based on an inaccurate and misinformed knowledge of how AI is developed and trained, that moreover impugns the legitimate right of a rightsholder to seek the protection of the law to protect their creativity and investment in content. Vigier might be correct when he says that licensing of content is a win/win for both parties. I agree with that. But licensing negotiations are about money and conditions of use and require willing parties on both sides. When licensing discussions break down, or when one party decides to do an end run on licensing because they have been rebuffed, then the way to gain clarity is through the courts whose job it is to interpret what the legislation means.
Canada still needs to come to grips with the question of how copyrighted content will interface with AI development. As I noted earlier, both sides in the debate made their cases in the public consultation launched a year ago, but since then there has been no movement in Ottawa. The law could be strengthened to ensure adequate protection of rightsholder interests in an age of AI, resulting in facilitating licensing solutions. In the meantime, misinformation and scare tactics need to be called out for what they are.
Adequate protection for rightsholders does not mean the end of AI innovation or investment in Canada. There is no need for panic. We can walk and chew gum at the same time.
© Hugh Stephens, 2024. All Rights Reserved.
Nice follow up piece, Hugh!
Randolph