Thomson Reuters has won a legal battle against Ross Intelligence for the latter using material from Reuters’ Westlaw platform to train their AI model. On February 11, the 3rd U.S. Circuit Court of Appeals overturned a 2023 decision, issuing a summary judgment decision that U.S. copyright law did not permit Ross to use Reuters’ content to help them build what was essentially a competing platform.
Ross was arguing fair use, but the judge in the case – Stephanos Bibas – shot that down, finding that their use was commercial in nature. This decision would seem to have potentially far-reaching implications since numerous other cases involving AI models using copyrighted material without permission are also attempting to make a fair use argument.
However, Judge Bibas made a point of clarifying that his decision addressed non-generative AI specifically – a type of AI that just analyzes and categorizes data rather than generating something new from it. Why does this distinction matter?
Potentially Far-Reaching Implications… Or Maybe Not
Upon initial viewing, the Judge’s decision would seem to strongly favor all copyright holders in their battle against AI models using their work without permission. But there is a reason that Judge Bibas was so quick to make the distinction between generative and non-generative AI.
Almost all of the big cases we all keep hearing about in the news, where authors, musicians, and others are suing companies who have used their copyrighted work to train their AI models, involve generative AI. Again, non-generative AI simply analyzes and categorizes data. It is taking something existing and organizing it for its own purposes. In contrast, generative AI absorbs and compiles copyrighted works and then mixes all of that information together to generate something new based on user prompts.
This is where the fair use doctrine really comes into play – and where things start to get tricky. Section 107 of the Copyright Act details four factors that should be considered when determining whether or not fair use applies to a work. One of the key elements at play in the consideration of the works of generative AI? Whether or not they are “transformative” uses.
Under Section 107:
“…‘transformative’ uses are more likely to be considered fair. Transformative uses are those that add something new, with a further purpose or different character, and do not substitute for the original use of the work.”
In other words, one of the big questions at play is whether or not generative AI is actually creating something new since the sum total of information it has to draw from is the work it absorbs – much of it copyrighted. Obviously, that is not the only question, but it is of particular interest when comparing it to this case involving non-generative AI because that concern did not even exist.
So, What Is Going to Happen in Those Generative AI Cases?
The takeaway? While this decision may have chipped away at some of the arguments made by the owners of generative AI models, the ultimate outcome of those cases is still very much undecided.
There is no telling whether future judges will build off of Bibas’ ruling and apply it to generative AI… or if they will decide that the situation is unique and use a different set of criteria. The only thing certain is that the outcome of those cases will have an enormous impact on copyright and intellectual property in general.