The World Is Human. The Cap Table Isn't.

Ali Ansari, writing on X:

The quality, safety, and alignment of AI models are directly tied to the well-being, clarity, and motivation of the experts contributing their judgment.

The piece traces AI's arc from hand-coded rules through statistical learning to today's frontier models trained on structured expert judgment. The historical sketch is solid. The argument that per-step accuracy compounds into task-level failure — 90% accuracy across 20 sequential decisions yields a 12% success rate — is the kind of arithmetic that should be tattooed on the forearm of every executive approving an "agentic AI" rollout. That alone is worth the read.

But the essay keeps reaching for a grander claim: that AI is becoming more human. It isn't. It's becoming a better lossy compression of certain human outputs. That's a meaningful and commercially important thing. It is not the same thing.

The recruitment example is supposed to illustrate "human-first" design. An AI agent posts roles, conducts outreach, interviews candidates, writes evaluations — and the human "ultimately approves the offer." That's not oversight. That's a rubber stamp with better lighting. If the agent shaped the pool, ran the interviews, and wrote the scorecard, the human is ratifying decisions that were already made three steps upstream. Whether that counts as "control" depends entirely on implementation details the piece never gets into.

The strongest part of the argument — that expert training labour is undervalued cognitive infrastructure — runs headfirst into the part the piece won't say out loud. The companies spending hundreds of billions on compute are not, in practice, treating human judgment as infrastructure. They're treating it as a cost to be optimised away. Ansari describes the world as it should work. The gap between that and how it actually works is the whole story.

"The world is human" is a nice closing line. But the companies building these systems aren't optimising for humanity. They're optimising for inference-cost-per-task. The world may be human. The cap table isn't.