Essay

The user tactics that make frontier LLMs usable — ritual prompts, workarounds, laborious decomposition — aren't accumulating in a resistant repertoire; they're training data the platform sorts, generalizing what serves it and quietly closing off what doesn't.

Generative AI raises the bar for earning credit while leaving the bar for blame untouched — concentrating a structural crumple zone and a psychological loss of felt ownership onto the same exposed signature.

Trust in AI doesn't reset between rollouts — it depletes like a non-renewable resource with asymmetric withdrawal rates, and no repair strategy restores it once employees have reframed errors as integrity violations.

Before an agent can do anything, someone must make the workflow legible to it — a massive, skilled, unpaid investment in articulation that roadmaps never name and the most peripheral workers absorb.

Deploying agents from day one doesn't remove the foundational learning phase — it makes it invisible, producing workers who look productive but never built the judgment to supervise what they delegate.

Chat-style AI compressed the skill gap; agents reverse it — briefing rewards decomposition, specification and evaluation, the same managerial competences that educational capital has always distributed unequally.

The AI super-user advantage is not a tech phenomenon — it's a laundered version of Bourdieu's linguistic capital, and LLMs are the most efficient class-sorting machine we've ever built.

Working with AI didn't extend my mind — it changed the plane on which I think, and the gap between those two things matters enormously for anyone building systems that depend on people thinking well.

The constraint in AI-augmented work has already moved — not away from the technology, but onto the human who must direct, review, and judge what the machine produces.

GenAI adoption stalls not because people fear the technology, but because organisations keep solving the wrong problem — and each failed rollout consumes trust that won't replenish itself.

More than a third of feedback interventions degrade performance — not because the intent was wrong, but because most organisations treat feedback as a free good and skip the four design conditions that determine whether it helps or hurts.

When performance slips, the instinctive managerial response is tighter control — which systematically destroys the information environment needed to fix the problem it was meant to address.