Silicon Valley is purging AI slop from music, video and feeds. Every platform running a purge pays its suppliers out of a pool it already fixed, which means the slop cost the platform nothing and the cleanup returns it nothing.
Spotify removed more than 75 million spammy tracks in twelve months. LinkedIn shipped a button. Google researchers described a system that terminated 50,000 account clusters covering 130,000 channels in six months. All of it targets supply. None of it changes the payout formula that made the supply worth producing, and none of it costs the platforms money either way, because Spotify, YouTube Shorts, TikTok and X all pay creators a fixed share of a pool rather than a price per unit. Under that structure, extra content dilutes the humans in the pool and leaves the platform’s economics untouched. Deezer, the one company publishing consumption data alongside upload data, reports that AI tracks passed half of daily uploads in June 2026 while accounting for one to three percent of streams, and that 85% of those streams were fraudulent. Strip the fraud and genuine listening to AI music sits under half a percent.
You would expect a flood this large to show up in a price somewhere. Advertisers pay for attention, and there is more inventory than ever, so the auction should be clearing lower. Meta reported the opposite on July 29. Ad impressions across its apps rose 14% year over year in Q2 2026 and the average price per ad rose 12%, the second straight quarter at that rate. More supply, higher prices, ad revenue up 27% to $59.36 billion. Whatever slop has done to the internet, it has not done it to the money.
What Happened
Tiffany Hsu reported in The New York Times on August 17 that tech companies have started what she called a generative waste detox. Her roll call: Spotify removing 75 million bulk uploads and duplicates, LinkedIn calling AI slop a top priority and building a report button, Google researchers describing a video service that wiped 130,000 channels over six months. Pinterest added a control letting users limit AI posts. TikTok is testing a US feature that lets people dial synthetic content up or down, has auto-labeled more than three billion AI videos, and removed over 377,000 videos in the first quarter for breaking its AI rules. Substack wired in detection from Pangram and got complaints from writers whose human posts came back flagged.
Two of Hsu’s numbers need a date stamp. Spotify first disclosed the 75 million figure on September 25, 2025, in a newsroom post announcing its AI policies. Sam Duboff, Spotify’s global head of artists, marketing and policy, repeated it during a visit to Australia in July 2026, adding that the platform now takes in roughly 100,000 new song uploads a day. At that intake rate, one year brings in about 36.5 million tracks, so the 75 million purge cleared close to two years of total supply.
The Deezer figures Hsu cites as “nearly half” have already moved. Deezer put AI-generated uploads at 44% of daily deliveries on April 20, then reported on July 21 that June had crossed 50% for the first time, at a monthly average of 90,000 AI tracks a day. The climb from January 2025 runs 10,000, then 30,000 in September, 50,000 in November, 60,000 in January 2026, 75,000 in April, 90,000 in June.
The Backstory
Deezer is the only platform in this story publishing both sides of the ledger, and its numbers rearrange the argument. Consumption of AI-generated music on Deezer runs between 1% and 3% of total streams. Of those streams, Deezer says 85% are fraudulent and get demonetized. Multiply it out and genuine human listening to AI music lands somewhere around 0.45% of streams at the top of the range, from a category supplying more than half of everything that arrives each day.
Deezer says so itself: fraudulent streams come out of the royalty pool before it pays anyone, so the effect on human artists’ royalties is minimal. CEO Alexis Lanternier frames the July announcement as safeguarding rights, and Deezer will now take down AI tracks tied to fraud and AI tracks that nobody has streamed in six months. The company also excludes every AI-tagged track from algorithmic recommendations and editorial playlists, which is the intervention that actually did the work.
Video runs the other way. Kapwing analyzed 15,000 popular YouTube channels, the top 100 in every country, and found 278 built entirely from AI slop, holding 63 billion views and 221 million subscribers, with estimated annual revenue near $117 million. Kapwing also opened a fresh account and counted 104 AI slop videos in the first 500 Shorts it was served. India’s Bandar Apna Dost, an animated rhesus monkey fighting demons alongside a Hulk lookalike, carries an estimated $4,251,500 a year.
Same content economics, opposite outcomes. On Deezer a listener picks the next song. On Shorts the ranker picks the next video. Slop finds an audience wherever the platform does the choosing.
The Plan
Google’s research paper, “Scalable Detection of Adversarial Synthetic Slop and Coordinated Media Abuse: A LoRA-Enabled Multimodal Defense System,” describes the Scalable Cluster Termination System. Rather than grade videos one at a time, S-CTS looks at upload pace, repeated templates and shared infrastructure, then kills whole networks. Over six months it terminated 50,000 clusters covering 130,000 channels and cut human review load by roughly half. The reported overturn rate sits below 1%, which at 50,000 clusters still leaves around 500 wrongly killed networks. YouTube told the Times that academic research does not reflect its full defense stack and that it protects well-intentioned creators using AI tools. Kurzgesagt, an animation studio with a decade of human work behind it, got flagged anyway.
Halving review cost is the line worth reading twice. That is the actual return on the project.
LinkedIn spent less. On July 30, chief product officer Hari Srinivasan added a “Seems like AI slop” option to the three-dot menu on any post. Flagged posts lose reach outside the poster’s own network, repeat offenders get a private note in their dashboard, and there is no appeal. Members supply the training labels for free. LinkedIn also deleted its own “Enhance your post” AI writer and replaced it with a proofreader that leaves the wording alone. Pangram scanned 1,002,627 posts across five platforms on July 9 and found 41% of long-form LinkedIn posts fully AI-generated. LinkedIn made up about a third of everything scanned and produced 62% of all the AI content Pangram flagged.
The Business Model Angle
Spotify paid the music industry $11 billion in 2025, up from $10 billion in 2024 and $1 billion in 2014. That payout is roughly two thirds of the revenue Spotify earns from music, allocated to rightsholders by their share of total streams rather than at a fixed price per play. The pool is a percentage of a number the slop does not touch.
Deleting 75 million tracks therefore adds nothing to what Spotify pays out. It changes the denominator, so the humans left in the pool take a larger slice of the same $11 billion. Spotify’s own cost of the flood was storage and ingest on 100,000 files a day, which rounds to a rounding error against a business that clears roughly 30% of recorded music revenue. The purge transfers money between suppliers. It does not create any.
YouTube Shorts runs the same shape with a sharper edge. Long-form creators take 55% of net ad revenue on their watch pages. Shorts creators take 45% of what the Creator Pool allocates them by view share, so YouTube keeps 55% on the format where the slop lives. Run Kapwing’s estimate through it: $117 million across 63 billion views works out to about $1.86 per thousand views on the creator side, implying roughly $260 million of gross advertising against those views, of which YouTube retains near $143 million. Those figures rest on third-party earnings estimates and a clean 45% assumption, so treat them as a scale check rather than a filing. The direction holds regardless. On the 278 channels the internet wants deleted, YouTube’s cut ran larger than the sloppers’ cut.
Set the platforms side by side and the size of each cleanup tracks how much bargaining power its suppliers hold, not how much slop it carries. Substack pays writers 90% and lets them leave with the email list, so Chris Best bought detection and put a score on every post. Spotify negotiates with three major labels who can pull catalog, so it announced the biggest number in the industry and shipped a spam filter, artist verification and SongDNA. YouTube’s three million monetized channels have nowhere paying better, so it accepts a sub-1% overturn rate and a Kurzgesagt incident as the price of halving review costs. LinkedIn’s suppliers post for free, so LinkedIn shipped a button and asked them to label the training data themselves. X, which pays out of an undisclosed pool at an unpublished rate, renamed the program.
The advertiser was never going to force the issue. The ANA’s Q1 2026 Programmatic Transparency Benchmark puts AI slop at 1.3% to 2.4% of open web programmatic spend and notes that it scores well on viewability, carries low invalid traffic and clears at higher CPMs than the alternative. Slop passes the tests advertisers buy on.
The Risk
Deezer’s consumption data is the weakest load-bearing plank here, and it is one company’s detector marking its own homework. Deezer applied for two detection patents in December 2024 and now licenses the technology to the industry, so it has a commercial interest in reporting both a large problem and an effective fix. If Spotify or Apple Music published stream-share data for AI tracks and it came back at 10% rather than 2%, the dilution argument would carry real weight and the cleanups would look like straightforward cost defense.
The pool argument also holds only while the pool stays fixed. Spotify’s label contracts include per-subscriber and per-stream minimum guarantees in some markets, and a large enough shift in stream mix can pull against those floors. Nobody outside the negotiating rooms can price that.
Camille François at Columbia put the case for the other side at a Slop Salon in Menlo Park this spring, arguing that the assortment of tactics signals a systemic problem with the information environment rather than a definable class of content to remove, and that platforms handling it poorly will watch people leave. Retention risk sits outside every number in this piece. Paul Shovlin at Ohio University framed the same trade: engagement rose first, liability arrived later. If users abandon a feed because they cannot find anything made by a person, none of the payout arithmetic saves it.
Here is the falsifier. If a major platform reports a measurable decline in session length, ad pricing or subscriber retention and attributes it to synthetic content, the cleanup becomes a defense of revenue and this reading weakens. Meta’s Q2 print, with impressions and price both climbing, is evidence in the other direction.
Quick Questions
Does removing AI tracks increase what Spotify pays artists? No. Spotify pays roughly two thirds of its music revenue to rightsholders and splits it by stream share. Removing tracks changes who gets what, not how much leaves the building.
Is AI music stealing meaningful royalty money? On the only platform publishing the numbers, less than the coverage suggests. Deezer puts AI music at 1% to 3% of streams and says 85% of those streams are fraudulent and demonetized before payout.
Why does video have a slop problem when music barely does? Listeners choose songs. Rankers choose Shorts. Slop earns wherever an algorithm picks the next item from an unlimited pool.
Does YouTube lose money on slop channels? Not on Shorts. Creators receive 45% of their Creator Pool allocation and YouTube keeps the rest, so slop views pay YouTube more than they pay the uploader.
Is AI slop hurting ad prices? No sign of it. The ANA puts slop at 1.3% to 2.4% of open web programmatic spend with strong viewability and low invalid traffic, and Meta’s average price per ad rose 12% year over year in Q2 2026 on 14% more impressions.
What would actually reduce the supply? Paying for something other than impressions. Every cleanup described here leaves the payout basis alone.
The Business Model Analyst Take
These platforms did not get flooded. They posted a standing order for unlimited content at zero origination cost, priced per unit of attention, with no quality gate at the door. For twenty years nobody could fill that order at scale. Generative tools closed the gap in about eighteen months, and the order is still open.
The cleanups are a returns policy. Useful, expensive to run, and aimed at the last step of a process the platform designed. Google’s paper gives away what the project is really worth by leading with review cost cut in half, because that is the only line in this whole story where a platform books a saving.
Deezer did the one thing that changed the outcome, and it was not deletion. Pulling AI-tagged tracks out of algorithmic recommendations and editorial playlists cut off the demand side, and consumption sits at 1% to 3% while uploads pass 50%. Deletion runs against a flow of 90,000 tracks a day. Delisting runs against the mechanism that made the flow worth funding.
Every operator running a marketplace should read the payout formula as a product spec, because that is how suppliers read it. Pay per impression and you will receive impressions. Pay per stream and you will receive tracks over 30 seconds. The people who built these systems are the customer of last resort for slop, and the invoice for that arrives as moderation cost, false positives on people like Kurzgesagt, and a supplier base that watched the platform spend a year deleting what its own formula ordered.
