Key copyright and AI ruling with broad implications for UK lawyers and beyond

The Thomson Reuters v ROSS Intelligence decision is worth a read for anyone interested in copyright and AI.

It’s a US decision from a Federal Judge, which is important because Federal Courts often set precedents that can influence subsequent legal interpretations and decisions across the whole US. We are not US lawyers, but our thoughts from a UK perspective are below.

To start with some background, there are currently over 40 copyright claims against AI companies across the globe. Most of these are in the US. Most of the training happened in the US, and one of the key questions is: does training AI on copyrighted content fall under the US doctrine of fair use? It’s a hugely emotive issue.

Thomson Reuters v ROSS Intelligence

In this case, Thomson Reuters sued a start-up AI competitor, ROSS Intelligence.

ROSS had been developing a non-generative legal research AI tool. It sought a licence from Thomson Reuters to use Westlaw, a legal research tool developed by Thomson Reuters. Thomson Reuters refused, and ROSS instead engaged LegalEase, who provided “Bulk Memos”. These memos had been created using headnotes from Westlaw. Thomson Reuters thought that these memos infringed and commenced proceedings against ROSS for copyright infringement of Westlaw’s 21,787 headnotes, 500 editorial decisions in various judicial opinions, and its Key Numbering System.  

Both parties originally sought summary judgment in 2023. At that time, Judge Bibas refused it; he held that many critical facts still remained in dispute.

However, this year, he revised that decision and held that ROSS was liable for direct copyright infringement in respect of 2,243 of those headnotes, and that its “fair use” defence failed. In his revised opinion, he held that the headnotes were protected by copyright and noted that the originality test for copyright subsistence in the US is an “extremely low” threshold.

Copyright subsistence in the UK and EU

In the UK and EU, copyright subsistence is also a low hurdle. The test, from Infopaq, is whether the work is a product of the author’s own intellectual creation. If that test is passed, the work is protected by copyright.

In the UK, Lord Justice Arnold summarised the test in THJ Systems Limited: it is a question of whether the author was able to make free and creative choices in the production of the work, “so as to stamp the work with their personal touch”. 

But in practice, this test is applied very generously. It was recently discussed by the UK Court of Appeal in Lidl v Tesco [2023] EWHC 873 (Ch), in which it was found that Lidl’s logo, comprising of the yellow circle on the blue background and the word “Lidl” was protected by copyright. Similarly, in Technomed v Bluecrest Health Screening Ltd & Anor [2017] EWHC 2142 (Ch), copyright was found to subsist in short sentences like "The 'normal' resting heart rate sits in a range of 60 – 99 beats per minute. Your heart rate is within normal limits."

Thomson Reuters v ROSS Intelligence judgement

The Federal Judge in the Reuters case reserved judgment on whether there had been direct copyright infringement of the Key Numbering System, the editorial decisions, and those headnotes that quoted judicial opinion verbatim (or near verbatim) for reasons of factual dispute. As for the remaining headnotes, Judge Bibas examined more than 3,000 of them and granted a partial summary judgment in favour of Reuters in respect of 2,243. He found that ROSS had actually copied those headnotes to such an extent that no reasonable jury could find otherwise. 

It is striking to see that Judge Bibas pored over more than 3000 head notes – in UK litigation, a judge will typically ask a Claimant to pick their best examples of alleged infringement which arguably put the rightsholder in a better position in terms of liability and the scope for claiming damages. This approach can be cheaper and simpler than arguing every point. 

Fair use defence

The most interesting part of the case from a UK perspective is the fair use defence. The UK does not have a fair use defence that would allow commercial text and data mining (Although in Getty Images, the Defendants are arguing that their use is “pastiche”. This reading of “pastiche” appears to us to stretch the normal definition.). It is, however, a key part of the defence to a lot of US actions.

In a California case, Kadrey & ors v Meta Platforms, Inc., 23-cv-03417-VC (N.D. Cal. Nov. 20, 2023), for example, Meta argues that its use of copyrighted text is analogous to Google’s use of copyrighted books to create an internet search tool, which was considered fair use. A hearing on Meta’s motion to dismiss in Kadrey & ors v Meta is expected to take place at the end of this month. Similar arguments are being run in many other cases. 

In the Reuters case, Judge Bibas considered four factors relevant to whether the fair use defence applies:

1. The use’s purpose and character 

ROSS intended to use the copyrighted material to compete with Westlaw. The use was not transformative. 

ROSS’s argument that copying only occurred at an intermediate step, after which the content was transformed into numerical data, and that such use is permitted under fair use, was rejected. 

Judge Bibas said that the non-generative nature of the AI model was important. We do not quite entirely follow why the use was not generative, but it is important to say that the judge saw a potential distinction. There is going to be further debate on this topic because there is not a bright line between this type of use and use which has been found to be fair use.   

In the UK, Stability AI is arguing (amongst other things) that its generative AI model, Stable Diffusion, does not store or recreate any of the copyrighted content. This seems a difficult argument to us, as it is still trained on the content, and so still needs to copy the content in order to be trained on it, but the law will continue to adapt and change with AI. 

2. The copyrighted work’s nature

Westlaw’s content was considered to have limited creativity, so this factor was decided in favour of ROSS. That said, Judge Bibas emphasised that factors (2) and (3) carry limited weight in comparison to factors (1) and (4) (and whether this is right is another topic that may be subject to debate).   

3. How much of the work was used and how substantial a part it was relative to the copyrighted work’s whole

This factor was decided in favour of ROSS; its outputs to end users did not include a Westlaw headnote. We think that this could be viewed in another way in other cases, as the important thing with AI is not just the output to users; it is often the training data. 

4. How Ross’s use affected the copyrighted work’s value or potential market 

Judge Bibas held that this factor was the most important. One has to consider the impact the use has on the market for the original work. ROSS was developing a competitor. This is an argument regularly made by the creative industries. Harvard’s Business Review’s 2024 statistics show a decline in writing, coding and image creation jobs since 2021 . 

Not all use is competitive, and there are plenty of rightsholder who are willing to licence data. Reddit, for example, licensed its user generated content to Google for approximately $60m per year. Open AI also has licensing arrangements in place with the Financial Times, Associated Press, Le Monde and Prisa Media, as well as recently landing a deal with News Corp estimated to be worth $250m over 5 years . 

The law in this area is interesting because the policy considerations are difficult. Of course, rightsholders should be compensated appropriately; but it is also clear that everyone wants to be able to reap the benefits of generative AI. The emotive responses to the UK IPO’s recent AI and copyright consultation show how complex and difficult this area is. Finding the right balance is an ongoing challenge, but this case is interesting in articulating some key considerations that the US courts are examining.