What is the equivalent of streaming for generative AI?

In the 1990s, the Internet was quite different from what we know today. It was the era when the web was just getting established, and we were gradually discovering the wonders of this new communication tool.

It was also the time when, suddenly, we discovered that we could digitize content (such as CDs or DVDs) and share it on the Internet with friends, but also obviously with strangers. And then, we discovered that all this content could not only be reproduced and transferred to the other side of the world in seconds but could also be modified and remixed in just a few clicks.

This was excellent news for those who wanted to maximize the visibility of this content and encourage greater consumption and production of cultural content, but it was bad news for copyright holders, who wanted to prevent unauthorized reproduction and distribution of their content. With the arrival of peer-to-peer file-sharing networks (such as BitTorrent, Gnutella, or Kazaa, for those who remember), the Internet began to be perceived, particularly by rights holders, as a threat to their business model.

With digital file piracy becoming accessible to everyone, no one would buy CDs or DVDs anymore, and the cultural industry would then collapse, as artists and content publishers would no longer be able to collect enough revenue to compensate for their efforts.

The first reaction was to strengthen copyright law by criminalizing the unauthorized sharing and downloading of protected digital content. In France, the HADOPI law introduced a “graduated response” system going as far as suspending internet connection for those who would not respect copyright law. Although this may have discouraged few people, these legal reforms did not succeed in containing the negative effects of online piracy on creative industries’ revenues.

The private sector then took up the issue, developing technological solutions aimed at restricting the possibilities of reproduction and redistribution of digital content. This is the case with all anti-copy systems to prevent the digitization of CDs or DVDs. However, these technological solutions extended well beyond the scope of copyright, to encroach on the access and consumption of digital content itself: When you bought an MP3 on iTunes, you could only listen to it on Apple devices, and no more than 5 devices at once. And if you bought a DVD in the United States, you might not be able to play it on your computer in France. These usage restrictions are such that content “protected” by these technological solutions is of inferior quality to pirated content, which is freely accessible on the Internet. And this naturally contributes to reinforcing the advantages of online piracy.

The solution to piracy was not a legal reform nor a technological solution, but a commercial innovation. Most people who downloaded digital content in violation of copyright did not do so because they didn’t want to pay for the content, but because it allowed them to access a practically unlimited collection of content, accessible in a few clicks, and without any usage restrictions. This was more interesting than most commercial offers that existed at the time.

Eventually, a new commercial offer arrived with “streaming” — with platforms such as Deezer or Spotify providing internet users with an almost unlimited amount of content, accessible instantly, without having to download files to their devices. The film industry followed suit, with platforms like Netflix, which make available a large collection of movies or series, accessible at any time.

And suddenly, the Internet, which appeared as a threat to creative industries, translated into a new commercial opportunity, allowing passive income generation for rights holders. So, unfortunately, many artists are exploited by streaming platforms that hold a monopoly position, but the emergence of these platforms has nevertheless demonstrated that it is possible to monetize digital content, even that which is freely accessible on peer-to-peer networks, as long as you provide a superior quality service for which people are willing to pay.

And then, as usual, history repeats itself. Today, with Artificial Intelligence, we are experiencing a new period of disruption, with a new technology that — like the Internet — presents both opportunities and threats to creative industries. On one hand, Artificial Intelligence can generate thousands of personalized content pieces in seconds, and almost automatically. With artificial creativity increasingly rivaling human creativity. On the other hand, this automation of creativity questions the very foundations of copyright law, since the law only recognizes humans, not machines, as having the capacity to be authors. Any work generated by AI will therefore not be protected by copyright and will automatically belong to the public domain.

However, the real problem raised by rights holders is that AI threatens their business models, both through its efficiency (since AI can generate content much faster than a human) and its economy (since AI work costs much less than human work). Yet these artificial intelligences have been trained on huge databases, fed by works produced by artists worldwide. All cultural content accessible on the Internet has been used as training data to create increasingly powerful generative AI models, capable of producing an infinity of content, inspired by and derived from training data, which is often protected by copyright. But if art feeds the machines, who feeds the artists?

Last month, Meta admitted to downloading more than 100 terabytes of pirated books from BitTorrent (equivalent to half a million digital books) to train its artificial intelligence models. In the HADOPI era, this would have been considered a criminal activity punishable by 3 years in prison.

But the fundamental question remains whether training an AI on copyright-protected content actually constitutes copyright infringement? While some claim that any unauthorized processing of protected data should be considered “theft” of data, especially when this processing is done for commercial purposes, others claim that using data as a learning base for AI models only stores this data in their knowledge system, so they can later draw inspiration from it to create new works — Which should not, in theory, be protected by copyright.

Yet, while the legality of the training phase remains uncertain, the legality of the generation phase is much clearer: if the content generated by an AI model is too similar to the content the model has ingested, there will probably be an infringement. This poses a legal risk both to model producers and to users of these models, who risk generating illicit images unknowingly. Once again, the solution to this problem cannot be limited to legal reform, nor to an exclusively technological device, we need to seek the commercial innovation that will make the problem obsolete. In other words, we need to identify what is the equivalent of “streaming” in the AI era.

As a researcher at CNRS and Harvard University, I have dedicated my academic career to researching how copyright can coexist with digital technology, and how this medium can bring opportunities to rights holders, instead of threatening them.

Today, I can’t help but try to find a solution regarding copyright and AI. And here is my conclusion: Generative AI systems suffer from two main problems: hallucinations and homogenization.

Hallucinations are such that whatever we ask an AI, it will generate it for us, even if it doesn’t know what it’s doing. This is great in an artistic context, where hallucination translates into imagination; but it is problematic in any context where the veracity of information is important, particularly in legal or journalistic contexts. Homogenization is due to the fact that these models, which have been trained on an enormous variety of data, tend to reproduce what is statistically significant disproportionately, thus reinforcing biases and uniformity in generated results.

On the other hand, rights holders suffer from the fact that when their content is used as training data for AI models, they lose control over how this content will be used. Once ingested by AIs, this content will become part of the models’ imagination and can be regurgitated in different forms, more or less similar to their original form, without rights holders being able to either validate or prevent these uses. And obviously, without being able to obtain adequate remuneration for these uses, it is practically impossible to calculate precisely which training data contributed to enabling the creation of new content by AI.

The good news is that there exists a common solution element to all these problems: the solution consists of using complementary modules that allow enriching artificial intelligence models with specialized data (an artist’s works, or a singer’s voice), without directly integrating them into the base model. Instead of training the model with protected data, specialized circuits are put in place that plug into these models to provide them with more specific information in real-time. This allows AI solution developers to offer better performing services while ensuring that protected data remains isolated and separate from the model.

And these solutions already exist. This is what Mistral did with its partnership with AFP — enriching its language models with AFP dispatches to provide correct and verified information about current events. While offering remuneration to AFP.

Alias.studio is a platform that allows artists to train models from their own creations and then make them available to the public under specific conditions. If I want to enrich my AI with ORLAN’s works, for example, I can now do so, but only if I respect the conditions she has defined, including pricing conditions for the use of these works. This means that on one side, ORLAN maintains control over her creations but also benefits from a passive income stream generated by her own AI model. And, on the other side, I benefit from a higher-end solution, with an AI enriched by ORLAN’s works, and with legal certainty that everything generated by this AI has been not only certified (in terms of quality) but also authorized by ORLAN (with a copyright license) — and I will therefore be very happy to pay for this higher quality service, knowing that ORLAN will be the main beneficiary of my royalties.

Beyond images and artists, the same principle applies to texts, music, or videos, as well as the voice or face of personalities. This protected data can be used to enrich AIs without losing control over how it is used.

Thus, as the Internet did in the past, AI raises many questions about copyright and somewhat forces the reinvention of cultural industries’ business models. But as with the Internet, it would be a mistake to consider AI only as a threat, forgetting all the opportunities it also brings. If streaming allowed the creation of new sources of passive income — both for creators and content publishers — generative AI is also a potential source of passive income, both for artists and cultural industries.

The important thing is to seize the opportunities of generative AI, but to seize them in a thoughtful and deliberate way. Artists, creators, and content publishers now have the opportunity to create their own models, with their own data, and make them available to all those who want to develop high-end AIs and provide superior quality services to traditional AIs, which suffer from hallucinations and homogenization. And this, while remaining in respect of copyright, ensuring control and fair remuneration for rights holders.

Because it is by enriching AIs (with quality data) that we can effectively enrich humanity (in both senses of the word).