I have been noticing a pattern with YouTube over the past months.
Not a drop in volume. The opposite. More creators, more content, more production quality. But less signal.
Different channels, same ideas. Slightly reframed, repackaged, re-edited. The thumbnails are sharper. The intros are tighter. The fonts are nicer. The insight underneath is the one I already watched four versions of last month.
It is not that any single video is wrong. It is that I keep meeting the same idea in a slightly better outfit.
The Feed Got Louder, Not Smarter
Research on feeds shown to new users gives some context for that experience. Up to 33% of feeds shown to new users now consist of AI-generated slop or brainrot content [1]. That is not a stylistic preference. That is one in three videos in your recommendations coming from channels whose primary skill is volume.
The supply side explains itself. Production costs for AI content have dropped 30 to 50 percent, with creation time down 70 to 85 percent [9]. When the marginal cost of producing a video approaches the marginal cost of typing a prompt, you do not get fewer videos. You get more of them, optimized for the metrics the platform rewards, which are mostly engagement metrics rather than novelty metrics.
The feed grows louder rather than smarter. There is more to watch, more to be recommended, more to scroll past. There is not necessarily more to learn.
Did the Map Catch Up With the Bottleneck?
For most of the internet’s history, the problem was access. You needed to find the information, get to the source, talk past the gatekeepers. The interesting question was always “how do I get to the thing?”
That question is mostly solved. The harder one now is “do I already know the thing?”
The information-access distinction also appears in policy. The Norwegian government’s 2025-2030 Strategy for Strengthening Resilience to Disinformation explicitly codifies the shift [5]. Their phrasing is direct: “the challenge lies not in accessing information but in understanding and navigating it.” When a national government writes the bottleneck shift into policy, the shift has happened. It has just been institutionalized.
The economic cost has moved past anecdote. Information overload costs the US economy up to $1 trillion annually, with constant task-switching between platforms cutting productivity by up to 40 percent [3]. The story we tell ourselves — that watching one more video is approximately free — assumes attention is unlimited. It is not. Attention is the line item nobody puts on the balance sheet, and it has been running negative for a while.
What drove the inversion? Content generation turned out to be the lowest-hanging fruit in the entire generative AI stack. Across enterprise deployments, content generation achieves the fastest implementation cycles — around three months — with 33 percent productivity gains [4]. The same property that makes content the easiest AI use case to ship is what makes the supply curve nearly vertical.
The cost of producing content fell below the cost of filtering it. That is the structural inversion. Nothing else in the system changed — our attention budgets stayed the same, our hours stayed the same — and the imbalance fell entirely on us.
This is, in many ways, the consumer-side mirror of The Human API. The producer side is exhaustion from generating at AI volume with human throughput. The consumer side is the same arms race, watched from the other chair.
The platforms eventually notice. Sometimes after they have monetized the problem for a few quarters.
YouTube’s Enforcement Response
YouTube is not a passive observer here. In July 2025 the platform renamed its “repetitious content” policy to “inauthentic content” [2]. The shift was substantive. Enforcement moved from detecting near-identical uploads (a problem AI-generated content easily routes around — every video can be technically distinct while substantively identical) to evaluating creative authenticity. A channel could now be flagged even if no two of its videos were the same, as long as the content was template-driven and lacking meaningful human involvement.
Then, in January 2026, YouTube executed a mass enforcement wave: 4.7 billion views wiped in a single action, sixteen channels totaling 35 million subscribers removed or demonetized [2]. The context was a 2025 audit showing 278 channels producing exclusively AI-generated content had accumulated 63 billion views, 221 million subscribers, and an estimated $117 million in annual ad revenue [1].
The revenue shows the incentive to keep producing at that volume. One hundred and seventeen million dollars, per year, for content the platform was eventually forced to act against. The business model of saying the same thing 63 billion times turned out to be a real business model, right up until it wasn’t.
YouTube did not fix it. YouTube reacted to it. Platform enforcement is reactive by construction. The 4.7 billion views existed before they were wiped. They were consumed. The filtering cost during those months fell entirely on the people doing the consuming, which is to say, everyone scrolling.
When enforcement is the last line of defense, you have already paid the attention tax. The question is whether the algorithm above the enforcement layer is doing any better.
The Algorithm Optimizes for Engagement, Not for “Do You Already Know This?”
Recommendation algorithms should solve this. That is the most serious version of the counterargument. McKinsey’s own attention research shows that larger content volume correlates with higher attention when paired with strong recommendation engines [10]. More content, properly filtered, should mean better discovery, not worse. Algorithms exist precisely so humans do not have to do this manually.
This is fair. It would be true — if the algorithm were optimizing for what you do not already know.
It is not. The Norwegian disinformation strategy notes that algorithms favor polarizing content and blend credible and dubious sources indistinguishably [7]. The optimization function is engagement: did you watch, did you finish, did you click. There is no signal in the system for “I have already seen this idea four times this week.” The reward function is structurally orthogonal to novelty.
The secondary effects compound. AI-generated content labels reduce perceived accuracy, credibility, and sharing intent regardless of whether the underlying content is correct [6]. The defensive crouch the brain enters when a label appears is itself an attention tax. And when AI summaries are present, users forgo clicking through to the original source roughly 85 percent of the time [8]. The pattern is consistent: the algorithmic layer encourages shallower engagement with more content rather than deeper engagement with less.
The algorithm is not broken. It is doing exactly what it was built to do. It is just not built to do the thing you actually need.
Can the Platform Correct Itself?
Two more objections deserve attention; each one affects how useful this diagnosis can be.
The first is historical. Every major media shift — the printing press, cable TV, the early internet — triggered the same overload panic, and humans developed new filtering mechanisms each time. If YouTube’s slop problem is genuinely unsustainable, the platform’s own enforcement wave is exactly the correction the objector predicts. The system is self-correcting. Wait it out.
This is partially right. YouTube is correcting. What it misses is the floor. Previous media shifts increased volume, but production still required meaningful human effort, which created a natural quality threshold. When content generation costs 30 percent of what it cost two years ago and ships in three months [4], the floor drops out. The correction still happens. It just happens months or years after a wave of redundant content has already been consumed. The lag is the tax.
The second objection is more uncomfortable. The pipeline I am about to describe requires technical infrastructure most knowledge workers do not have. If the prescription is “build a custom AI-powered filtering pipeline,” the advice serves perhaps five percent of the audience and abandons the rest.
That objection is correct, and worth leaning into rather than dodging. The pipeline is not the recommendation. The pipeline is the evidence. When a working professional finds it worth building custom infrastructure to filter free content, the bottleneck has structurally moved. The takeaway is the decision framework, not the code.
I will admit the irony. Writing an article about content redundancy that might itself be the seventh version of a familiar idea you have already seen elsewhere this week. If it is, the diagnosis is at least internally consistent.
Filter Before You Consume, Not While You Consume
The pipeline I built treats content as a stream rather than a library. Anything that looks potentially interesting gets sent into it. The content is transcribed, parsed, and matched against an existing body of knowledge — not statically, but as a comparison. The question the system asks is simple: is there anything here I do not already know, or that reframes something in a meaningful way?
If the answer is no, the content gets filtered out. If there is something new, it surfaces as a small set of insights. Only when those insights are strong enough does it justify going back to the original video.
I tried NotebookLM before this. It is genuinely useful — for one domain at a time. The moment you try to scale it across the full set of inputs a working professional consumes, the overhead of managing notebooks becomes its own system. Filing the inputs starts to consume the attention the filing was supposed to protect. The tool stopped being a filter and started being a different version of the thing it was supposed to replace.
The infrastructure is not what matters. What matters is the inversion. Most of us were taught to consume content and then decide whether it was useful. The bottleneck shift inverts that order: decide whether content is likely to be novel, then consume.
You can apply the principle without building anything. Three categories cover most of what counts as novelty:
- New data — a statistic, a case study, a finding you did not have.
- A reframe — a perspective that changes how you think about something you already knew.
- A contradiction — a claim that challenges a belief you held as settled.
If a piece of content does not fit at least one category after a 60-second triage of title, creator track record, and explicit claim, it is recognition, not learning. Recognition is pleasant. The brain rewards it the same way it rewards genuine novelty, which is exactly why it is so easy to confuse the two.
There are domains where this filter fails on purpose. When you are genuinely new to a topic, you do not have a baseline to filter against. Volume consumption is the right move there, because you are building the baseline that future filtering depends on. Premature filtering in a new domain is its own failure mode — the filter only works once you know enough to filter for.
For the domains you have been following for a year or more, the filter is the practice.
So What Now
If you have been treating your YouTube watch list as evidence of learning, take a look at the last ten videos you completed. Articulate, in one sentence each, what they added to what you already knew. The ones you cannot summarize were not learning. They were familiarity wearing a thumbnail.
You do not need a pipeline. You need a baseline. Define, for the two or three domains you actually follow, the five to ten things you already hold as settled understanding. Then ask of any incoming content: is there new data here, a reframe, or a contradiction? If not, skip it. The volume on YouTube is not going down. The cost of your attention is not going down either.
The interesting question is not whether YouTube is still teaching anyone anything. The interesting question is whether you can tell.
References
[1] Kapwing Research. (2025). AI Slop Report: The Global Rise of Low-Quality AI Videos. Kapwing. https://www.kapwing.com/blog/ai-slop-report-the-global-rise-of-low-quality-ai-videos/
[2] TechCrunch. (2025). YouTube prepares crackdown on ‘mass-produced’ and ‘repetitive’ videos. https://techcrunch.com/2025/07/09/youtube-prepares-crackdown-on-mass-produced-and-repetitive-videos-as-concern-over-ai-slop-grows/; Flocker. (2026). YouTube Inauthentic Content Policy: AI Enforcement Wave 2026. https://flocker.tv/posts/youtube-inauthentic-content-ai-enforcement/
[3] Causes, consequences, and strategies to deal with information overload: A scoping review. ScienceDirect / Elsevier (2024). https://www.sciencedirect.com/science/article/pii/S2667096824000508
[4] Business Impact of Generative AI. International Journal of Advanced Research (2025). https://www.journalijar.com/uploads/2025/04/6826c8fc76a80_IJAR-51221.pdf
[5] Norwegian Government. (2025). Strategy for Strengthening Resilience to Disinformation (2025-2030). https://www.regjeringen.no/en/documents/strategy-for-strengthening-resilience-to-disinformation-20252030/id3109255/
[6] How Users Perceive and React to Labeled AI-Generated Content. International Journal of Human-Computer Interaction / Taylor & Francis (2026). https://www.tandfonline.com/doi/full/10.1080/10447318.2026.2618553
[7] Norwegian Government Strategy (above) + McKinsey & Company. (2024). Winning the battle for consumer attention. https://www.mckinsey.com/industries/technology-media-and-telecommunications/our-insights/the-attention-equation-winning-the-right-battles-for-consumer-attention
[8] Neobadger. (2025). AI Overviews: The Shifting Sands of Content Consumption. https://neobadger.com/resources/ai-overviews-the-shifting-sands-of-content-consumption/
[9] Business Impact of Generative AI. International Journal of Advanced Research (2025) and Future of the Creator Economy Report 2024. https://www.epidemicsound.com/blog/the-future-of-the-creator-economy-report-2024/
[10] McKinsey & Company (2024). Winning the battle for consumer attention (above).