AI-generated songs are flowing into streaming services faster than people can listen to them. A research team at the University of Chicago conducted a large-scale measurement of Spotify and found that AI-generated tracks made up over 40% of newly published songs in early November 2025. At the same time, 93% failed to reach the minimum threshold of 1,000 plays required to generate revenue. While the number of tracks has surged, demand hasn't kept pace, and a business model built on low-cost mass posting in pursuit of a rare hit is taking root ahead of actual listener demand.
These findings come from an unreviewed preprint titled "An Empirical Analysis of AI Slop in Music Streaming," published on June 16. In addition to catalog analysis, the research team submitted AI-generated tracks they created themselves to 11 distribution services and conducted experiments to evade AI detectors. What emerged was a picture in which the mechanisms for limiting volume at the distribution entry point are lagging further behind than the technology used to create the audio itself.
Over 40% of New Tracks, Yet Plays Don't Follow
The research team used metadata for 256 million tracks collected by Anna's Archive. The publisher claims this covers 99% of Spotify's catalog as of July 2025, and the research team verified consistency using multiple metrics. However, this is not official data provided by Spotify. After removing duplicates using International Standard Recording Codes (ISRCs), 185.36 million tracks remained, of which 52.92 million were published in 2024 or later. The team also traced Spotify's public recommendation API to build a graph consisting of 33.51 million tracks and 3.5 billion recommendation relationships.
Whether each track was AI-generated or not was determined by cross-referencing it with AI labels assigned by Deezer. By linking Spotify and Deezer track information, the team classified 40.75 million tracks as either AI-generated or human-made. In a verification exercise using a sample of 1,000 albums, the matching was correct 99.6% of the time.
The proportion of AI-generated tracks among newly published songs rose sharply from under 1% in January 2024. In the final week for which data was available, the week of November 3, 2025, there were 202,046 AI-generated tracks, 238,646 human-made tracks, and 60,022 unclassified tracks. The share of AI-generated tracks exceeds 40%. However, the paper's statement that "AI tracks may have surpassed human-made tracks by 2026" is an extrapolation from this growth rate, not an actual measurement for 2026.
An increase in supply cannot simply be read as popularity. While AI-generated tracks accounted for 5.1% of the entire analyzed catalog, they made up only 1.2% of the recommendation graph. Looking at cumulative plays, 93% of AI tracks had fewer than 1,000 plays. Even when comparing tracks over the same playback period from January to May 2026, 92.7% of AI-generated tracks and 67.5% of human-made tracks fell short of 1,000 plays.
Since April 2024, Spotify has only counted tracks toward recorded music royalty calculations if they received at least 1,000 plays over the trailing 12 months and met an undisclosed minimum unique listener threshold. In other words, the vast majority of AI tracks generate no revenue even after being published.
The 2.7% Posting Over 30 Tracks a Month Reshaped AI Track Supply
The reason posting continues unabated despite weak demand is that the cost of producing a single track is extremely low. According to the paper's estimates, a $100 annual subscription to Suno or Udio can generate roughly 500 tracks a month, bringing the generation cost per track to under $0.02. Rather than screening out failures, it can be more rational to churn out volume and wait for an occasional hit.
The research team operationally defined "slop producers" as AI artists who publish 30 or more tracks per month. Such artists make up only 2.7% of all AI artists. Even so, this group's track output surpassed that of all other AI artists combined by October 2025. A small number of high-volume posters are driving the apparent growth of AI music.
That said, the 30-tracks-per-month threshold is not a measure of musical quality. It is a research classification based on posting frequency, and it cannot fully separate musicians who use AI as a production aid from low-quality mass posters. Even accounting for this limitation, the gap in production speed is substantial. Since 2024, AI artists have published an average of 27 tracks compared to 13 for human artists. On a monthly basis, that's 5 tracks versus 1, with release intervals of 16 days versus over 50 days.
The royalties earned by AI tracks from January to May 2026 were estimated at roughly $13.4 million, about 0.3% of Spotify's monthly payouts. Even narrowing the focus to the top 7.15% of AI tracks that earned any revenue, 76.5% earned less than $10. Only 1,632 tracks—0.27% of all AI tracks—earned more than $1,000. While this estimate treats per-play rates uniformly across regions, it captures well the coexistence of low success rates and mass production that characterizes this economy.
What 11 Submission Experiments Revealed About the Distribution Entry Point
Individuals cannot deliver tracks directly to Spotify or Apple Music. Independent musicians typically use distribution services to prepare audio formats and metadata before sending them to each platform. This intermediary layer, which checks for copyright infringement and duplication, is one of the few entry points where AI slop could be stopped.
The research team created AI tracks using the commercial services Suno and Udio, as well as the open-source tools DiffRhythm and ACE-Step, and submitted 8 tracks each to 11 distribution services. They also generated lyrics, track titles, and cover art using generative AI, minimizing human effort. Nine of the 11 services approved at least one track, resulting in a total of 73 tracks proceeding to distribution.
The study included 5 services whose stated policies prohibit AI tracks. Yet 4 of these approved all 8 submitted tracks, and the remaining one approved at least 1 track. In cases where rejection did occur, the reasons given were unrelated to AI usage. There is a wide gap between stated policy and actual screening practice.
Once a track passed through a distribution service, the streaming platforms largely accepted that judgment without further scrutiny. The research team deleted the published AI tracks after concluding the experiment. Although this was a small-scale study, it demonstrates—from the perspective of those seeking to flood the system with AI tracks—that there is little friction in the pipeline from generation to publication.
Even a 99.6%-Accurate Detector Was Evaded Using Compression and Reverb
If audio could be automatically identified, entry-point screening could be strengthened. The research team compared four detection methods and found that the Fourier method, which analyzes the frequency components of spectrograms, achieved the highest accuracy. Under standard conditions—using 1,000 AI-generated tracks extracted from Suno and Udio, and 1,000 human-made tracks from before 2021—it reached 99.6% accuracy.
However, there is no guarantee that a track's audio will reach the public in the same state it was in during training. The research team created 20 additional tracks using Suno, applying MP3 compression to 5 and pitch alteration to another 5. To the remaining 10, they applied either added reverb or autoencoder-based reconstruction. When these were submitted to Deezer via DistroKid, all 20 tracks evaded the company's commercial AI detection system. While the sample size is small, this shows that accuracy under standard conditions alone cannot determine real-world robustness.
This weakness also affects the research team's classification of Spotify tracks. The analysis relies on Deezer's AI labels, meaning that generators or processed audio that Deezer fails to detect are not counted as AI tracks. Combined with the fact that album-level labels were applied to all tracks within the same album, the figure of over 40% should be understood as a large-scale estimate rather than a rigorous, exhaustive count.
Meanwhile, figures published by Deezer itself in April 2026 closely align with the research findings. The company reports receiving roughly 75,000 fully AI-generated tracks per day, accounting for 44% of daily submissions. Yet these account for only 1-3% of total plays, and the company excludes up to 85% of plays on AI tracks from monetization after flagging them as fraudulent. Mass supply, low genuine demand, and fraudulent plays—these three characteristics overlapped on another platform as well.
Beyond Voluntary Labeling, Posting Costs and Identity Verification Are Needed
The music industry has begun addressing this issue through transparency measures. On July 10, eight organizations including IFPI and RIAA announced a voluntary labeling system distinguishing "AI-Generated" from "AI-Assisted" at the track level. This separates recordings whose principal creative elements were made by AI from recordings led by humans that incorporate AI in some capacity. Participating organizations also included WIN, an international body representing independent music companies, and SAG-AFTRA, the union representing actors and musicians.
This distinction is useful for legitimate artists explaining their production process. Spotify, too, has been piloting AI credit display via DDEX since April, receiving tens of thousands of disclosures daily. However, the company acknowledges that the absence of a credit does not mean AI was not used. The July 10 industry proposal is also voluntary and applies only to sound recordings. Lyrics, composition, music videos, and cover art currently fall outside its scope.
Expecting self-disclosure from those aiming for mass posting and fraudulent plays will not close the gap at the entry point. The paper suggests candidates such as per-track distribution fees, fee increases scaled to posting frequency, and caps on the number of tracks published within a given period. Of 18 distribution services examined, only 3 required identity verification. Rigorous identity verification would raise the cost of maintaining large numbers of accounts.
The success of the industry's labeling implementation cannot be measured by the number of labels displayed. What matters is how much mass-posted content lacking AI credits gets captured, how effectively continuous publishing by the same actor is curtailed, and whether fraudulent plays can be excluded from royalty calculations. As production costs continue to fall, businesses that push volume despite low per-track success rates will not disappear. Whether real costs can be restored at the distribution entry point will be the dividing line between AI music and AI slop.