Deezer plans an AI music filter for search
Deezer announced on October 13 that it plans to let listeners hide fully AI generated tracks, albums and related artists from search results. The reversible setting is due globally before the end of 2026.
The announcement distinguishes entirely generated music from human work with limited AI assistance. Deezer already labels detected synthetic content and excludes it from recommendations. The new control concerns search. It does not establish that tracks are being deleted from the catalog.
The demand comes from an older survey
Deezer points to a November 2025 survey conducted by Ipsos Digital. It found that 45 percent of music streaming users wanted a filter for fully generated music. That result describes preferences measured last year, rather than a fresh poll accompanying this announcement.
The online survey covered 9,000 adults aged 18 to 65 in eight countries, with quotas for nationally proportional representation. Its streaming user results came from a subgroup of 6,791 people. Within that group, 66 percent said they would try fully generated music at least once out of curiosity.
Another widely repeated finding needs its test conditions attached. The reported 97 percent failure rate came from asking participants to classify three tracks, two generated and one human made. It does not show that listeners fail on every song or cannot learn to recognize particular clues.
A willingness to sample something and a desire to control its visibility can coexist. The results support offering a choice more clearly than they support declaring one universal listener preference.
The filter depends on what the detector recognizes
Deezer researchers have published a technical explanation of one approach. Their 2025 paper on audio artifacts describes small patterns in the frequency content of generated sound. Certain components that turn a model’s internal representation into audio leave these patterns as a consequence of their design.
In that account, the signal comes from the generation architecture rather than the particular music used for training. The researchers tested the method on open models and commercial generators including Suno and Udio. This offers a concrete explanation of how detection can work without asking a person whether a song sounds artificial.
The accompanying official repository makes an important qualification. The research represents one component of Deezer’s production detector. Other layers and later improvements are absent, so running the published experiment would not reproduce the full service or establish its current accuracy.
New generators keep the problem moving
A July 2026 preprint by Deezer researchers studies detection when the music generator is unfamiliar. It examines both separating real from synthetic audio and grouping samples from different generators without an existing label for each one. The authors combine extracted audio artifacts with statistical decomposition and classification methods.
That work is research context. It is not confirmation that this exact system powers the upcoming setting. It does illustrate why a successful test on known generators is only part of the evaluation needed for a changing catalog.
What listeners and artists should watch
For listeners, a useful control should make its scope understandable and let them inspect the effect. If a search looks different, they should be able to tell whether a preference changed the results or whether the music was unavailable for another reason.
Artists face the other side of the same decision. A detection mistake could change whether a listener encounters their work. Clear explanations and a way to challenge an incorrect classification matter alongside the convenience of a single switch. These are evaluation questions, not claims about a verified appeals process.
Detection also leaves questions about payment and permission. ByteForward’s report on music publishers’ AI licensing principles explains the industry proposals on attribution and royalties.
