Every token in a training corpus is a fossil of social interaction.
Researcher Bright Simons writes that AI’s intelligence isn’t a property of architecture, parameters, or compute, but an inheritance from the social complexity of the civilization whose language it ingested.
He makes this vivid with a thought experiment: train identical transformer architectures on the texts of 3000 BC Egypt, then 300 BC Athens, then 1500 AD Florence, then the modern internet.
Each successive model gets dramatically smarter. Not because the tech changed. But because the civilization did.
AI capability depends on social complexity, but AI deployment systematically reduces that complexity… through cognitive offloading, homogenization of output, and the elimination of interaction-dense work.
The technology is gradually undermining the conditions for its own advancement. Its successes, not its failures, drive the spiral.
When models train on AI-generated text, the tails of the distribution vanish first. Minority viewpoints, rare knowledge, edge cases go extinct. This leaves fluent, plausible, hollow homogeneity. Shumailov et al. Nature, 2024.
So we have to stop assuming scaling is the binding constraint.
You won’t “win” the next decade with the highest AI utilization rates. You will win the next decade using AI to create more human interaction.
More friction.
Investing in what Simons calls “trans-mediation” and “high human interactionism.”
The edges of disagreement are where rare knowledge lives. Minority perspectives, awkward questions, unusual formulations, oral traditions, domain-specific judgment, studio critique, embodied practice, cultural memory. These are not inefficiencies. They are the tails of the distribution. And when those tails disappear, intelligence becomes smoother, faster, more confident, and less alive.
This is exactly why Fine Arts and the Humanities matter in the age of AI.
A playwright arguing over intention, a filmmaker defending a framing choice, a designer questioning the ethics of an image, a dancer thinking through the body, a historian challenging the premise of a dataset: these are not soft add-ons to technical progress.
They are part of the human machinery that generates meaning.
The intelligence that AI systems exhibit was never individual to begin with. It was forged in the spaces between people.
And if AI is trained on the residue of human collaboration, then the future of AI depends on whether we protect the spaces where that collaboration still happens.
An inheritance, consumed without reinvestment, eventually runs out.
