A new YouTube account, with no watch history to bias the recommender, gets served AI-generated video in 21% of its first 500 Shorts: 104 out of 500, according to a report from video-editing platform Kapwing, which built the account and counted. Another 33% (165 of 500) met Kapwing’s definition of “brainrot”: low-effort content optimized for compulsive viewing regardless of who or what made it. The two categories overlap but aren’t identical, and the feed stayed clean for the first 16 videos before either started showing up.
That’s one measurement, on one platform, with one methodology: a single fresh account, one observer, YouTube’s algorithm in October 2025. A second, larger measurement comes from Pangram, an AI-text-detection company, which ran a browser extension across roughly one million posts on five platforms between April and June 2026 and found 41% of long-form LinkedIn posts (250 words or more) were fully AI-written, the highest share of any platform it checked. Medium came in at 31%, X at 25% (with another 23% flagged AI-assisted), and Substack lowest at around 10%. Pangram’s chief executive, Max Spero, called the number a lower bound, since people who install an AI-detection extension aren’t a random sample of the internet.
The two numbers measure different things by different methods and shouldn’t be averaged into one “AI slop rate.” What they agree on is the order of magnitude: on the platforms that have checked, somewhere between a fifth and two-fifths of the content a user is shown or posts is now machine-made, not touched up by a human but generated start to finish. Business Insider’s survey of seven platforms’ responses shows most of them converging on the same three tools: label it, let users dial it down, and stop recommending or paying for the worst of it. None has tried to ban it outright.
YouTube: the money, not the ban
YouTube doesn’t block AI video from the Partner Program; it blocks the categories of AI video that make advertisers nervous. On July 16, 2026, the company’s global head of trust and safety, Matt Halprin, walked through three buckets that lose monetization eligibility under the “inauthentic content” policy: generic or template-built videos, content built on emotionally manipulative formulas (his example: animals staged into apparent danger, then “rescued”), and AI personas posing as human experts on health, finance, legal or political advice. Halprin was explicit that the wording of the policy, renamed from “Repetitious Content” in July 2025, hadn’t changed; this was a clarification of an existing rule, not a new one.
The clarification followed a January 2026 sweep in which YouTube removed more than a dozen channels, some with millions of views, built on AI-generated cats and AI-generated Jesus content. Kapwing’s count puts a number on how much of that economy survived the sweep: 278 channels that post exclusively AI-generated video had, as of October 2025, collected a combined 63 billion views and 221 million subscribers, which Kapwing estimates at roughly $117 million in annual ad revenue, or about $421,000 per channel on average. The distribution is skewed hard toward a few winners: the single highest earner, an Indian channel called Bandar Apna Dost, drew an estimated $4.25 million a year on 2.07 billion views, ten times the per-channel average. By views, South Korea’s AI-slop channels led with 8.45 billion, ahead of Pakistan’s 5.34 billion and the US’s 3.39 billion. Separately, YouTube Music began labeling AI-generated tracks in October, letting listeners tap a player-screen label to see whether composition or vocals were AI-made; standard pitch correction such as Auto-Tune doesn’t count.
Substack bets on detection it lets you turn off
Substack’s approach, announced July 21, 2026, is a scan rather than a filter. The company partnered with Pangram to let readers run any post, note, reply or comment over 100 words through a detector that estimates how much of it is AI-written, live on web and iOS since that date, with Android to follow. Writers can add their own disclosure of what tools they used, or disable scanning on their own posts entirely, which means the tool measures what writers are willing to have measured. Substack’s chief executive, Chris Best, framed the goal bluntly: the company built the integration, he said, “to avoid slop and prevent Substack from turning into LinkedIn.” Pangram claims better than 99.9% accuracy and a false-positive rate near 1 in 10,000 for its newest model, though that figure comes from the company’s own testing, not an independent audit, and its hardest case, a human draft lightly edited by AI, is also its weakest one by Pangram’s own account. The company raised $9 million led by Menlo Ventures on July 29, bringing total funding to roughly $13 million, and counts universities, publishers and recruiters as customers alongside Substack and Quora.
LinkedIn builds the problem it now flags
LinkedIn’s position is harder to square than most, because the company actively promotes an AI button that polishes posts, and Pangram’s data shows the result: 41% of its long-form posts and 30% of its short-form posts fully AI-written, the highest of any platform the firm checked, with LinkedIn alone accounting for nearly two-thirds of all AI content Pangram’s extension detected across five platforms despite making up only a third of posts scanned. Pangram also found AI-written LinkedIn posts draw 45% less engagement on average than human-written ones, which is presumably why Steven Bartlett’s media company, FlightStory, dropped AI-drafted LinkedIn posts after concluding the human versions performed better. LinkedIn’s response has been to add a “Seems like AI slop” report button on July 30 and retire its “enhance your post” AI writer in favor of a proofreader billed as preserving a user’s own voice. One detail reporters flagged as ironic: Pangram’s own model marked LinkedIn’s announcement post about fighting AI slop as itself AI-generated.
The rest: dial it down, don’t push it, don’t pay for it
TikTok, built in large part around ByteDance’s own Seedance 2.0 video model, requires creators to label AI content, embeds an invisible watermark in some videos’ metadata, and since November 2025 lets users toggle how much AI content they see; in July 2026 it began testing detection aimed specifically at accounts that post AI spam to crowd out other creators. Pinterest, which has labeled AI-modified and AI-generated content since 2025, added a feed control for the same purpose; its chief executive, Bill Ready, told Business Insider in June that alongside AI’s genuine uses there’s “a bit of a counter movement to that with consumers.” Meta labels content its systems detect as AI-generated across Facebook, Instagram and Threads, but doesn’t automatically extend that label to AI-modified content unless it’s photorealistic video or realistic-sounding audio; in June it extended the label to ads built with AI, Meta’s own tools or third-party. Snap took the sharpest line: as of July 31, 2026, wholly AI-generated videos stay published on Spotlight but stop being eligible for recommendation; video shot by a human and edited with Snap’s own AI tools keeps full eligibility plus a transparency marker. None of these seven companies has published a number for how much traffic, revenue or engagement its anti-slop measures actually removed, which is the number worth watching once the clarifications have had a full quarter to bite.