Some days, AI-generated music represents more than half of the new tracks uploaded to Deezer. To prevent this surge from disrupting recommendations and penalizing both listeners and artists, the platform detects and labels tracks that are entirely generated by AI, then removes them from algorithmic recommendations and editorial playlists. However, these tracks remain available when a user searches for them directly, unless removed for fraud or prolonged lack of listening. Today, it takes only a few words in a prompt to generate a poster, a video, and even a complete song. With just a few clicks, it can be shared on streaming platforms. This simplicity is profoundly changing the way music is produced—and raises a new question: how to distinguish a human creation from a track entirely generated by artificial intelligence? Deezer has chosen not to wait for the "AI slop" to flood its recommendations. Since January 2025, the platform has had a tool capable of identifying tracks entirely generated by AI, then limiting their visibility and detecting potential fake streams. "AI slop" refers here to content generated in bulk by artificial intelligence, often of variable quality, which can clutter platforms or be used in fake stream strategies. The term does not refer to a musical genre or all creations using AI. The distinction is important. Deezer specifically targets tracks entirely generated by AI. An artist who occasionally uses AI technology in their creative process is therefore not automatically affected. In June 2026, Deezer received an average of nearly 90,000 tracks entirely generated by AI per day, and these exceeded, during peaks, half of the new uploads. In April, the platform mentioned about 75,000 titles daily, representing 44% of new tracks. Deezer analyzes the audio file directly to search for signatures left by generative AI models. These signatures are generally inaudible to humans but can be identified by the detection system. Deezer's technology is particularly capable of recognizing productions from models such as Suno and Udio, and can evolve to accommodate other models. For the system, it is not enough to look at the artist's name, the album cover, or the track's metadata. Deezer claims a detection accuracy of 99.8%: approximately two AI tracks out of 1,000 may not be detected, while less than one human-generated track out of 10,000 would be incorrectly flagged. To further limit the risk of error, the AI label is applied at the album level. The "AI" label indicates to the listener that content has been identified as entirely generated by artificial intelligence. It does not constitute a quality rating and does not mean that a track is bad. It simply provides additional information about its origin. In a study conducted by Ipsos for Deezer among 9,000 people in eight countries, 97% of participants did not correctly distinguish tracks entirely generated by AI from tracks created by humans during a blind test. Furthermore, 80% believed that music entirely generated by AI should be clearly marked. Are AI-generated tracks removed from Deezer? No, because identifying a track as generated by AI does not automatically mean its removal from the catalog. Deezer first applies a logic of transparency and limitation of visibility. Detected tracks are labeled and removed from algorithmic recommendations and editorial playlists. However, since July 2026, Deezer has announced its intention to remove AI tracks used to generate fraudulent streams, as well as certain tracks that have generated no listening for at least six months. Deezer removes detected AI-generated tracks from Flow, algorithmic recommendations, and editorial playlists to limit their artificial exposure. At Deezer, receiving thousands of automatically generated tracks is one thing. Automatically giving them access to discovery mechanisms is another. By preventing these tracks from entering recommendations and editorial playlists, Deezer significantly reduces their ability to gain organic audience. While AI-generated tracks can represent more than 50% of new uploads during peaks, they constitute only 1 to 3% of the total listening volume. Therefore, the number of tracks uploaded does not correspond at all to the number of tracks actually listened to. The risk appears when the mass production of AI-generated tracks is associated with artificial listens intended to generate revenue. Today, creating thousands of tracks has become relatively simple and inexpensive. Some can then be artificially pushed by fraudulent streaming networks. Deezer has identified a particularly marked phenomenon on AI-generated tracks, and up to 85% of their streams were considered fraudulent in 2025. These listens are then excluded from the royalty calculation. It is important to note that an AI-generated track is not necessarily fraudulent. It is the behavior of the listens that allows the identification of potential manipulation. Deezer removes identified fraudulent streams from the royalty calculation so that they do not dilute the revenue from legitimate listens. The platform is therefore not only trying to protect the user from an undesirable recommendation. It is also trying to prevent an artificial volume of listens from diverting part of the compensation intended for artists and rights holders. Deezer indicates that fraudulent streams are excluded from royalties and are therefore not paid. Since January 2026, Deezer's detection technology is also offered under license to other industry actors. It has notably been tested with actors such as Sacem and is used by Billboard to identify AI-generated tracks in the context of its charts. It has become very difficult to reliably detect AI-generated music solely by listening. Some voices may seem particularly smooth, some lyrics very generic or some repetitive structures. However, none of these indicators constitute proof. The most reliable method is therefore to look at the information provided by the platform. On Deezer, the "AI" label precisely identifies tracks entirely generated by artificial intelligence when the system detects them. And that is probably the whole point of the system: the listener no longer needs to play detective to know what they are listening to.