where are we in the battle between creators and artists and AI developers?

Creators: 0. Artificial intelligence developers: 1. In the battle between authors and artists and AI developers, companies like OpenAI, Meta, Microsoft, Google and StabilityAI are still one step ahead. And it’s not copyright that’s going to slow them down. For months, writers, actors, voice actors, artists and content creators have continued to publish op-eds and appeal to legislators. Their goal: for AI systems to stop “ plunder their works » for training purposes, unless you have obtained authorization, for remuneration.

AI systems must ingest billions of data points to learn how to generate text, images, code or music. However, there is great opacity regarding the databases used, with these tools not making the difference between data protected by copyright and copyright-free data. Enough to make these AI systems capable of imitating – or reproducing – perfectly, a given artist, or of competing or even replacing creators, many of them fear. But despite their actions, things have not changed one iota.

Also read: A great “digital plunder”: when generative AI challenges copyright

Discussions on future regulations are still going well, without having yet come to fruition. And despite some concessions, AI developers repeat that they are not ready to put their hands in their pockets to remunerate authors and artists…

In the small victories of creative artists: Hollywood?

With a few exceptions, because artists have achieved small victories in recent weeks. Wednesday November 8, we learned that the strike of screenwriters and actors in Hollywood ended. The artists’ negotiating committee would have obtained, in the United States, ” unprecedented provisions to ensure consent and compensation (from actors), in order to protect our members from the threat of AI.”

According to Reuters who was able to interview one of the negotiators of this agreement – ​​the text has not yet been published at the time of writing – film studios will now have to obtain authorization from actors to use their image in sequences generated by artificial intelligence (AI). And they will have to pay performers each time their digital double appears on screen, under the labor agreement that ended a 118-day strike. Minimum remuneration has been set for any use of their digital double, with actors free to negotiate more. The latter feared being transformed into digital characters that studios could use “ for the rest of eternity “. The screenwriters, for their part, feared being replaced by AI or being relegated to proofreading or editing.

Also read: Cinema: how the world of dubbing is trying to stem its “looting” by AI

In France, filmmakers demand the maintenance of transparency obligations in the AI ​​Act

Concern is also alive in France. Last November 8, during the cinematographic meetings of the ARP (a group of filmmakers) which took place in Le Touquet, the director Radu Mihaileanu recalled, during a debate, the importance of two articles of the Regulation on AI (or AI Act). The latter, currently under discussion in trilogue, would provide greater transparency in this area. Section 28B would force AI developers to demonstrate transparency on the training data of the systems, since the latter would be required to publish a more or less detailed summary of the protected data used for training. Article 52 would, for its part, require that all content generated by AI be labeled as such. “ It is essential that these two articles remain » in the final version of the future law, argued MEP Laurence Farreng during the debate, referring to the negotiations currently taking place between the European Parliament, the Council and the European Commission.

While waiting for the “AI Act” to be negotiated and adopted – its application is planned in two and a half years at the earliest – there will be good a latency time. And, subsequently, a second level of regulation will be necessary, argued the European parliamentarian. A revision of copyright could well be put on the table, suggested one of the MEPs in charge of the text to our colleagues from Context, Friday November 10. If content protected by copyright is used when training the system, remuneration will be required, via a catalog repurchase, a tax, or another means that remains to be invented, argued MEP Laurence Farreng .

Negotiate directly with AI companies while awaiting regulation

In France, a Franco-French bill was tabled last September to further protect content creators from AI. Because ” why not demonstrate, here too, cultural exception », asks Mathilde Croze, interviewed by 01net.com. For the associate lawyer at the Lerins firm, specializing in digital law, “ in France, we have still been able, in the past, to make cultural exceptions to try to protect French heritage and cultural exception. Why wouldn’t we be in a field in which we would be able to do the same thing with AI? », she questions.

An interministerial committee on AI and a sectoral group specific to culture have been set up by the Government to study the impact of AI, particularly on the cultural sector. But they will only present their recommendations to the Government that within six months.

While waiting for adequate regulations to be adopted, what should we do? For Mathilde Pavis, an expert in intellectual property law who was part of the debate at the ARP meetings, companies like artists must renegotiate collective agreements and contracts, she argued. Problem: the author or creator of content is often not in a strong position against the AI ​​giants. Hence the need for regulation on a European scale.

This renegotiation is precisely what the French press has been trying to do for several weeks. Last August, Reporters Without Borders pleaded for the media to prevent OpenAI, by configuring their site, “to collect their content for free”. “The media must be paid for their work in the general interest from which the tech giants would like to profit handsomely”added the NGO on.

As we explained to you, the 2019 Copyright Directive allows AI to collect data on the web (the so-called “ text and data mining “), but only for research purposes, not for commercial purposes, and only in the case where the author has not used ” opt-out “. Translation: only if he has expressly argued that his content should not be ingested by crawlers, these robots exploring the Web, the AI ​​does not have the right to ingest his works. How to express this refusal of access ? For now, this question is not resolved, and that is the whole problem.

But one way, recommended by the Geste, which brings together nearly 90 online publishers, would consist of integrating the “ TDM Reservation protocol » to the sites in question. This is'” a necessary prerequisite for the establishment of licenses within a negotiated framework », Wrote the organization in a September press release. And this instrument has been adopted by many press titles, Mind Media reported on October 18.

First court decisions and pleas from AI giants

But apart from these few advances, the balance of power is far from being reversed. The AI ​​giants do not seem ready to put their hands in their pockets and pay the rights holders or the creators of content protected by copyright. This is what shows comments recorded by OpenAI, Meta, Microsoft, Adobe, Anthropic, Hugging Face and StabilityAI last week in the United States.

The AI ​​giants took advantage of a public consultation period, which was organized by the US Copyright Office on possible copyright reform, to defend their point of view at length .

If the arguments differ, their conclusions are similar: according to the latter, no, they shouldn’t pay copyright – just like today. This will not fail to provoke a reaction from the many content creators, against what they consider to be a pillaging of their works. For Mark Zuckerberg’s group, all “ the American AI industry is based on the idea that copyright law does not prohibit the use of copyrighted content to train AI models “. Change this to “ imposing a licensing regime unprecedented now, well after the fact, will cause chaos. Because developers will seek to identify millions and millions of rights holders, for very little profit, given that any fair royalty due would be incredibly small in light of the insignificance of any one work among a training set Al », writes Meta.

The reason: forming a system would not, in itself, be a use amounting to economic exploitation of content protected by copyright. It would not be a question of reproducing the work, but “ identify patterns in a large corpus of content ”, under the “ fair use ”, an exception to copyright provided in American law – but which does not exist in Europe and in France.

Same story at Google, for whom training an AI system with protected works would be like reading a book. For Anthropic, current (American) law should not be changed. Moreover, some of the companies in the sector seem so sure of themselves that they have not hesitated to commit to reimbursing their users’ legal costs, if they are attacked on the copyright front. . After Adobe in June, Microsoft in September, Google, in October, promised in a blog post that it would compensate its users in two cases. “Sit is proven that the data used to train its AI violates copyright or, if a company finds itself convicted of copyright violation after having directly exploited content generated by one of its AI “.

Finally, the first two court decisions which pit artists and authors against the AI ​​giants in the field of copyright have been handed down. Although their scope must be put into perspective, since they were made based on American copyright law which is very different from European/French law, they both lean in favor of AI companies.

Thursday, November 19, an American judge, who was seized by several writers, ruled that the texts generated by Llama, Meta’s AI system, did not violate the authors’ copyrights, reported Reuters. At the end of October, another American court judged another case between artists and Stability AI, Midjourney and DeviantArt. The three companies were accused of having drawn from photos and works, without authorization, to train their AI. The judge ruled that the charges against Midjourney and DeviantArt were dismissed. The reason: no evidence was provided of any use of copyrighted works by these two companies – the latter had not published their training databases. Only Stability had shown more transparency. For the latter, the case is still ongoing.