Everyone is asking whether AI will take our jobs. I think that’s the smaller question. The bigger one is whether machines will overtake us as generators of knowledge itself — and I want to defend an unfashionably optimistic answer, through a mechanism I call the Circle of Stupidity.
The idea is simple. Today’s AI is trained on the recorded output of human intelligence; its brilliance is borrowed. That sets up a self-limiting loop. First, AI ingests the accumulated corpus of human knowledge. Second, as AI does more of our intellectual work, our incentive and capacity to produce genuinely new knowledge erodes — fewer people are paid to think from scratch, and more of the public record becomes machine-authored. Third, starved of fresh human insight and increasingly retrained on its own output, AI hits sharply diminishing returns and, in the limit, degrades.
If the loop holds, machine capability is capped by a ceiling that only humans can raise. Human primacy persists not out of sentiment, but as a structural feature of the knowledge economy.
What surprised me is how much recent research points the same way. Models genuinely collapse when trained on their own output (Shumailov et al., Nature 2024; Alemohammad et al.’s “Model Autophagy Disorder,” ICLR 2024). The stock of quality human text is projected to be largely exhausted by frontier training within a few years (Epoch AI). Returns to scale diminish by the very math of scaling laws. And the parts of cognition that create new knowledge — Polanyi’s tacit dimension, the embodied judgment behind Moravec’s paradox — were never written down, so they can’t be scraped.
I take the strongest objections seriously. Self-play systems like AlphaZero, and discovery engines like AlphaGeometry, GNoME, and AlphaFold, do seem to leap past humans without reading more of us. But each depends on a cheap, perfect verifier — the rules of chess, a symbolic proof-checker, physical synthesis, the Protein Data Bank. The Circle of Stupidity is a claim about open, unformalizable domains where no such oracle exists. And the result that model collapse is avoidable when real data keeps accumulating (Gerstgrasser et al., 2024) isn’t a refutation — it’s a statement of the thesis’s precondition: the loop stays open only as long as fresh human knowledge keeps flowing in.
So my conclusion isn’t triumphalism. It’s a bounded, domain-dependent human primacy — and a responsibility. If continued human relevance rests on continued human knowledge-generation, then protecting the livelihoods of original researchers, writers, and observers, and keeping human signal distinguishable in the commons, isn’t nostalgia. It’s critical infrastructure.
The Circle of Stupidity will save us only if we decline to complete it.
This is a work in progress. I genuinely want to hear where you think it’s wrong. Views and suggestions welcome, even (especially) contrary ones — write to me.