A "hammer verdict": Thomson Reuters wins a copyright lawsuit against an AI company. A turning point for the AI ​​industry

The media group Thomson Reuters sued the American AI company Ross Intelligence in 2020. Thomson Reuters accused Ross of improperly using the content of its Westlaw platform to train an AI. Now a judge has ruled in favor of Thomson Reuters.

Media lawyer Ralph Oliver Graef from the Graef law firm in Hamburg wrote in response to a request from the NZZ: "This is a bombshell ruling that has caused quite a stir in the legal AI bubble. The makers of Open AI and other AI companies will now be sweating and cursing."

Thomson Reuters, to which the news agency belongs, is just one of many rights holders who are fighting back against the practices of AI companies. In London, the photo agency Getty Images filed a lawsuit against the English AI company Stability AI in 2023. Stability AI is said to have violated Getty's intellectual property rights by using copyrighted images without permission to train its AI model.

In a 69-page indictment, the New York Times accuses Open AI and Microsoft of using articles and content from the newspaper to train the AI. If the decision is made in favor of the New York Times, the media company could demand millions in damages or even the destruction of the data set on which Chat-GPT was trained.

Westlaw, operated by Thomson Reuters, is the largest platform for researching legal cases. Paying users can access contracts, court rulings and laws there. Much of this content is publicly accessible elsewhere and therefore not protected by copyright. But Westlaw offers an additional benefit. The so-called headnotes are editorially created content, summaries, explanations and further information on the individual cases.

The court argued that although the headnotes did not have the creative value of a work of art, they were still original enough to be considered a copyrighted work in their own right - an assessment that another judge may not share.

The startup Ross Intelligence wanted to build a competing product to Westlaw. Ross trained a legal case search engine that uses AI to filter out direct quotes from court cases. It is not generative AI because no new texts are created. However, copyrighted training data was used.

To train his AI, Ross needed legal data. Ross originally wanted to license the content from Westlaw, but Westlaw refused. Ross then turned to Legal Ease and purchased 25,000 digital file cards with answers to legal questions as training material. The answers are based on the headnotes.

In his defense, Ross relied on the fair use doctrine. This regulation of American law allows the use of copyrighted works in exceptional cases. These cases include educational or research purposes. Open AI also argues this.

The judge in Delaware rejected Ross's argument on this basis. "When applying the fair use doctrine, there is always a balancing of interests regarding the specific use of copyrighted works," says Fabian Reinholz, partner at Härting Rechtsanwälte in Berlin. Another judge could weigh the arguments differently. In the USA, it is therefore decided on a case-by-case basis whether the fair use doctrine is applicable.

"It can be assumed, however, that other US federal courts will be guided by the Delaware decision and we could soon see further rulings against AI companies," writes Graef.

If the courts hearing similar cases, such as that of the New York Times against Open AI, agree with the ruling, this will call into question the business model of Open AI and other AI companies.

Graef uses a comparison to show how difficult it could be in practice to delete individual data from the models: "It may be that the AI ​​training cannot be undone, just as toothpaste cannot be squeezed back into the tube, but at least there is an obligation to pay compensation for the rights holders and creatives." The AI ​​companies would then have to compensate the rights holders financially.

"It wouldn't turn the AI ​​world upside down, and it wouldn't stop AI per se, but it would be a turning point," says Reinholz. Ross himself ceased business operations in 2021. The legal costs were eating into the startup's financial resources.

Christian Kramarz is a specialist lawyer for copyright and media law. He doubts that the US jurisprudence will have a major impact in the EU or Switzerland: "In America, we have a completely different legal tradition."

EU law also allows the use of copyrighted works within certain limits. These include caricatures, parodies or private use. There is also a so-called text and data mining limit that allows use for scientific purposes. The judges are divided on whether this regulation is also applicable to AI.

This is particularly relevant for the case of Laion e. V. against the photographer Robert Kneschke at the Hamburg Regional Court. The photographer found one of his images in a dataset for training AI, although he had objected to its use. The non-profit research network Laion had created the dataset and checked whether photos and image descriptions actually matched. The dataset was then made available to AI providers for training their models.

Since Laion is a non-profit organization that does not pursue any commercial intentions itself and it created the data set for scientific purposes, its use by Laion was permissible, the Hamburg judges decided. Whether a commercially oriented AI company that trains its AI on the basis of the Laion data set would have broken applicable law was not up for debate in court.

"The ruling in the USA is also being observed in Switzerland," says Nicole Beranek Zanon, partner at Härting Rechtsanwälte in Zug. Unlike in Germany, however, there is still no case law in Switzerland on copyright in the AI ​​age.

Lawyers like to talk about individual cases that need to be assessed on a case-by-case basis. But in the end, an AI company trains its model based on someone else's intellectual property, earns millions from it and ignores the authors' remuneration, arguing that they are doing this for the good of society.

In the USA, the EU and Switzerland, courts will have to deal with this issue in the coming years. Beranek Zanon says: "Personally, I would very much welcome it if the legislature or the courts continued to protect female authors. Otherwise, that means that we will completely degrade knowledge work in the future."