The Cognitive Collapse Thesis: How Polluted Information Loops Are Degrading Both Machine and Human Intelligence—And What It Means for Capital, Power, and Civilization
Substacksubstack.com
The Cognitive Collapse Thesis: How Polluted Information Loops Are Degrading Both Machine and Human Intelligence—And What It Means for Capital, Power, and Civilization
The choice is not between AI and human intelligence. It is between information ecosystems that sustain both and information ecosystems that degrade both. The research is clear on where current trajectories lead. The strategic question is what to do about it.
Substack • The Cognitive Collapse Thesis: How Polluted Information Loops Are Degrading Both Machine and Human Intelligence—And What It Means for Capital, Power, and Civilization
The winning strategies over the next decade will not be those that maximize short-term AI deployment. They will be those that build and maintain access to high-quality information flows—genuine human expertise, carefully curated training corpora, institutional knowledge that has not been corrupted by synthetic recursion.
Substack • The Cognitive Collapse Thesis: How Polluted Information Loops Are Degrading Both Machine and Human Intelligence—And What It Means for Capital, Power, and Civilization
This bet—that domain-specific human expertise matters more than volume of human annotation—represents one theory of how to maintain quality in an increasingly synthetic information environment.
Substack • The Cognitive Collapse Thesis: How Polluted Information Loops Are Degrading Both Machine and Human Intelligence—And What It Means for Capital, Power, and Civilization
Social media platforms face a structural contradiction that current valuations do not adequately price. Their business model requires maximizing engagement. The research demonstrates that engagement-optimized content degrades both the AI systems trained on it and the humans consuming it.
Substack • The Cognitive Collapse Thesis: How Polluted Information Loops Are Degrading Both Machine and Human Intelligence—And What It Means for Capital, Power, and Civilization
The web is no longer primarily human-generated text created for human readers; it is increasingly AI-generated text created for algorithmic amplification, mixed with human-generated text optimized for the same algorithmic reward functions.
As AI systems train on this degraded corpus, they learn to reproduce its characteristics—shallow reasoning,... See more
As AI systems train on this degraded corpus, they learn to reproduce its characteristics—shallow reasoning,... See more
Substack • The Cognitive Collapse Thesis: How Polluted Information Loops Are Degrading Both Machine and Human Intelligence—And What It Means for Capital, Power, and Civilization
The web is no longer primarily human-generated text created for human readers; it is increasingly AI-generated text created for algorithmic amplification, mixed with human-generated text optimized for the same algorithmic reward functions.
As AI systems train on this degraded corpus, they learn to reproduce its characteristics—shallow reasoning,... See more
As AI systems train on this degraded corpus, they learn to reproduce its characteristics—shallow reasoning,... See more
Substack • The Cognitive Collapse Thesis: How Polluted Information Loops Are Degrading Both Machine and Human Intelligence—And What It Means for Capital, Power, and Civilization
Subsequent research by Gerstgrasser and colleagues at Stanford and MIT (arXiv:2404.01413) provides important qualification: model collapse is not inevitable if synthetic data accumulates alongside real data rather than replacing it. The pathological scenario assumes complete substitution—thoughtful data curation can prevent the worst outcomes.
Substack • The Cognitive Collapse Thesis: How Polluted Information Loops Are Degrading Both Machine and Human Intelligence—And What It Means for Capital, Power, and Civilization
The finding is elegant and disturbing: when AI systems are trained on data generated by previous AI systems—a scenario increasingly common as synthetic content floods the internet—the models progressively lose their ability to represent rare and subtle patterns. The statistical tails of the distribution vanish. The models converge toward median,... See more