Inside Deezer: how AI tagging supports artist-centric economics

Since Autumn 2023 and the launch of its artist-centric model, Deezer has seen a major risk emerge with the massive arrival of AI-generated content on streaming platforms.

In anticipation of the general public’s wide adoption of generative music models, the platform developed an in-house detection and tagging system that’s central in its efforts to combat revenue dilution and streaming fraud.

In this interview, Ludovic Pouilly, SVP Institutional & Music Industry Relations at Deezer, discusses with IDOL the motivations behind this work, how it fits into the continuity of the artist-centric model, and the issues of transparency and equitable remuneration at a time when 44% of new tracks uploaded are AI-generated.

We started working on detecting AI-generated content when we launched the artist-centric model. That was in October 2023; we were already anticipating the possibility of a massive influx of AI-generated content as generative models were becoming accessible to the general public, which facilitates fraud.

Our Research & Development teams developed detection tools that became operational at the end of 2024. Their implementation allowed us to share the first figures with the industry starting in January 2025, as there was complete opacity around this phenomenon. We took the lead, as no one else had this type of tool available.

To make informed decisions, we needed concrete data fast. Therefore, we started communicating the figures in January 2024. The phenomenon is massive: initially, about 10% of the content delivered daily was 100% AI-generated. This percentage is constantly increasing, recently reaching 44%.

Our philosophy at Deezer is to promote musical diversity and local production. This guides our thinking on remuneration models as much as our work on streaming fraud and, more recently, on the detection of AI-generated content.

When we adopted the artist-centric model, one of the measures was to remove non-musical content, which we call “noise.” In our view, this content should not be treated as musical works. Deezer is first and foremost a music platform; that is a choice we’ve taken responsibility for.

We therefore removed this content from the remunerated catalog, then replaced it with our own productions. The idea was not to degrade the user experience, as this content meets real uses, such as white noise frequently used to help infants fall asleep. This in-house content is not taken into account in the calculation of market shares and therefore does not dilute the remuneration of rights holders.

Also, in this logic of combating everything that dilutes the remuneration of rights holders, we realized that fraud was proportionally high on this type of “noise” content. We deduced that they are easy to produce and therefore particularly tempting for malicious actors.

Paradoxically, while we often talk about fraud in rap, there is very little proportionally; however, on this non-musical content, it is massive. Today, we estimate that about 70% of streams generated by AI are fraudulent. They are therefore excluded from the royalty pool.

Our position is that if a user chooses to listen to an AI-generated track, that is their choice. For the sake of transparency, we inform them that these tracks were generated by artificial intelligence, but we do not recommend them. The user must therefore take the initiative to search for them.

A survey we conducted at the end of 2025 with Ipsos showed that 97% of listeners do not differentiate between human and AI creation in terms of quality. However, they demand transparency about this content and believe that it should not be remunerated in the same way as human creations.

Our first measures were to stop recommending this content and to be transparent, including towards the industry, by sharing our figures. We remain open to other actions, but they will always be based on solid data.

Specifically, the problem is this mass delivery. It generates significant storage and processing costs for content, 80% of which will never be listened to and which very often serves to commit fraud. To address this, we very early on asked our suppliers not to deliver 100% AI content to us.

We are currently modifying our policy to strengthen this position and are now asking our suppliers to flag AI-generated content, a process that takes time to implement.

If this content is not recommended, as we have chosen to do, it does not generate streams, and dilution remains limited. Conversely, a platform that pushes it through its algorithms mechanically creates dilution, especially if its remuneration model is based on the market-centric model.

The fight today is against misuse. Copyright fraud, for example, when the voice of a well-known artist is used to capture listens, which falls under copyright law and requires deleting the content. Or tracks that achieve a certain success but remain unauthorized remixes of pre-existing works, such as what we recently saw around Stromae’s ‘Papaoutai’. There is a wide diversity of situations, which we must analyze very precisely.

The sharing of our figures with the industry generated a lot of discussion to the point where several players asked us to make our detection tools available. This is not our core business; we remain primarily a streaming service, but the requests were such that we decided to develop solutions to make this technology accessible outside of Deezer, via technical arrangements adapted to specific needs.

We have therefore initiated discussions with other DSPs, because for us, this is not really a subject of competition. At Deezer, we benefit from a great Research and Development team, which means we have solid technology, tested by numerous professional organizations and by the majors. We have full confidence in the quality of these tools, as well as their capacity to evolve, because we are continuously improving them.

Even though we do not hold copyrights, we consider it unfair that generative AI models are trained on protected content. That is why we were the first to sign the international declaration on AI training in October 2024, in line with our artist-centric philosophy.

We want rights holders to be properly remunerated and to have a say in the use of their content. Using human creations to generate artificial music without the agreement or remuneration of rights holders does not seem acceptable to us.

Deezer has always been a pioneer, and continues to innovate on issues of equity in the music industry. Since we started communicating and sharing our work on AI, our image as a tech-oriented platform has been greatly strengthened and is now widely recognized and valued by rights holders.

Faced with attempts to bypass detection, one of the strengths of Deezer’s research tools is that they are both interpretable and generalizable, which makes it possible to spot the emergence of new AI models.

The teams are working in particular on the granularity of analyses, taking into account modifications made to AI-generated content to trick detection, on the ability to quickly provide proof in case of dispute, as well as on ongoing projects such as detection at the stem level. In an environment that is evolving so quickly, this anticipation is crucial for the future.

For rights holders, it remains very difficult to prove the illegal use of their works by generative AI models, hence the ongoing lawsuits, particularly against Suno. The lack of transparency of certain platforms complicates the demonstration that a model was trained on protected material without authorization, which leads to unfair and unremunerated competition for human creation.

Regulation is therefore desired, and transparency is essential. It is legitimate for a rights holder to be able to know if their works were used to train an AI model, and under what conditions. In France, the Darcos bill, approved by the Senate, proposes a reversal of the burden of proof: in the presence of evidence of use, it is up to the model to prove that it did not do it. This facilitates the protection of rights holders.

Things are evolving quickly. Some agreements have been found; we know that Universal, Warner, and Merlin have signed with Udio, for example. We do not have the details of all the agreements, but we can clearly see that a negotiation movement is underway with these generative models. They are clearly looking to find proposals that are acceptable to the industry.

I am not privy to the discussions, but I hope that in the medium term, rights holders will find common ground with these models. We do not yet have a definitive decision to say what is legal or illegal, but it is moving forward. That is encouraging.