AI slop, detection arms races, and why algorithms should optimize for lifetime human flourishing
Oct 7, 2025
Key Points
- Detection arms races against AI-generated content are unwinnable; watermarks and metadata can be trivially circumvented through screen recording or inpainting, making platform strategies built on detection a distraction from the real problem.
- Video platforms optimize for watch time and engagement metrics that keep users hooked on low-quality content, leaving them feeling depleted despite high completion rates and driving churn once dopamine fades.
- Platforms should optimize for lifetime human flourishing across decades, not quarters, since educational content shown to young users compounds into higher lifetime customer value than junk content today, but public-market earnings cycles prevent this.
Summary
AI slop is not the problem. The problem is slop, and the tools that generate it are almost irrelevant to the real issue facing algorithmic platforms.
The familiar objection treats AI-generated video as the culprit, flooding feeds with low-quality content that needs detection and separation. This misses the mark. Stephen Hawking doing a McTwist at the X Games on AI video is creative and worth watching. The actual issue emerges when platforms optimize for the wrong metric and push endless low-effort content at users regardless of its origin.
Detection will not solve this. Generative adversarial networks pit generators against detectors in an endless arms race. For images and video, detectors can theoretically embed invisible watermarks or use cryptographic signing at the hardware level, but these are trivial to circumvent. Film a screen displaying AI-generated content on your phone, and the metadata still says iPhone, but the video is synthetic. Watermarks can be inpainted away. Text detection is even harder because there is simply not enough signal in a sentence to reliably distinguish AI from human writing. Adversarial pixel-level tweaks can fool image classifiers entirely. Detection will never be airtight. Building platforms around the assumption that it will be is a distraction from the real problem.
The real problem is algorithmic optimization. Video platforms have evolved through three phases of metric alignment, each solving one problem while creating another.
Click-through rate YouTube ranked videos by clicks on thumbnails. This broke immediately because clickbait like "I bought a Bugatti" earned clicks without delivering watch time. Users felt duped.
Average view duration Platforms shifted to optimizing for watch time. This solved clickbait because low-CTR, high-AVD videos now thrived while high-CTR, low-AVD videos tanked. Users got higher-quality feeds. But the metric had a hidden cost. A user could watch an entire hour of engagement bait, finish with 100% completion, and feel like they wasted their time. The algorithm sees "loved it, watched to the end" while the user knows they watched junk. This drives churn. One person left TikTok for exactly this reason because the feed was candy, and after the dopamine wore off, so did the desire to return.
ARPU and churn mitigation Platforms now optimize for average revenue per user by keeping users coming back repeatedly. But this can still rot brains. A user can be kept highly engaged and monetizable for a year or two through content that damages their cognition, career, and life trajectory. Eventually they become depressed, lose jobs, drop out. They stop being good consumers.
The solution is a fourth phase: optimize for lifetime human flourishing, not over quarters or years, but across an entire human lifespan. The math is capitalist and straightforward. Show a kid brain-rot content and cheap ads today. Show them educational content about programming or business, and in twenty years they have a high income and can buy premium things. Discounting the future value back to today, the Ferrari ads shown to them later generate more lifetime revenue than the junk ads today.
No platform is structured to optimize this way. YouTube is twenty years old and the average human lifespan is seventy-eight. To fully validate this strategy requires a fifty-year feedback loop, one full cohort from childhood to adulthood showing which early content choices correlate with lifetime customer value. The data will not exist for decades. But that knowledge gap is not an excuse to optimize for something else. The guiding light should still be user flourishing over a human lifetime, even with rough estimation and incomplete signals.
One data point cuts against the US approach. ByteDance's Chinese version of TikTok reportedly surfaces science and math content to children while the US TikTok feeds them slop. ByteDance operates with a multi-decade time horizon and thinks in terms of national human capital, not quarterly earnings. The implication is that US platforms, constrained by public-market earnings cycles, are extracting attention today at the cost of customer lifetime value tomorrow.
When tested informally with creators and tech people, this pitch resonated, a sign the logic maps onto a real tension in how platforms think about their role.