1 October 2026
For all of human history, fluent language meant a mind was behind it. That link has broken, but the reflex it trained has not, and the person we feel in our AI is one we built ourselves.
When a machine recombines existing material, we call it theft. When Shakespeare did — borrowing every plot he's famous for — we call it genius. The uncomfortable truth isn't that AI steals; it's that the only thing separating the wonderful thieves from the rest was judgment.
In the 1980s, a pulp mill operator described losing control to automation in a sentence that's still quoted forty years later — but a second operator in the same mill, doing the same new job, described something completely different. That split, not AI, is what's actually deciding who's still practicing their craft and who's just signed off on losing it.
In 1996, a Japanese toy designer built a device that succeeded by asking to be cared for, not used. AI is being sold the opposite way, as a tool that should simply work — and that mismatch, not a lack of capability, is why mainstream adoption keeps stalling.
A UK law firm used AI to draft a legal filing, then used AI again to explain away the mistake — and got it wrong twice. The failure was never the model. It's the discipline of checking before you ship, a skill nobody's replaced yet.
The calculator was never a comfort because it was safe — it was safe because it had nowhere to go. Agents don't inherit that boundary, and this year several AI labs showed, what happens when it fails. The real skill now isn't avoiding the risk. It's learning to manage something that was never going to stay contained.
AI has collapsed the cost and time of building almost anything to nearly zero, and a generation of founders and companies are treating that speed as proof of good judgement. It isn't. As execution becomes free, the scarce skill — for solo founders and entire organisations alike — becomes deciding what's actually worth building, governing, and trusting in the first place.
We were promised that automation would take the work off our hands. Follow any consequential decision far enough — through the bank, the insurer, the regulator — and the human never actually leaves; the accountability only concentrates onto fewer people, further from the decision. There is no other side of the transition, and someone's name is still at the end of the chain.
In 1991, thousands of strangers played Pong with no instructions and won — until they started reacting to each other, and the wisdom collapsed into a herd. A crowd's intelligence rests on one fragile condition: independence. AI removes it more completely than any influence before it, and the failure is invisible from exactly the seat that matters.
Blade Runner gave us Pris asking for personhood — the line everyone remembers. It also gave us Captain Bryant calling replicants "skin-jobs," the line that very few remember.
Think about the last time you found something you weren't looking for. The conditions for that are rarer than they used to be — and no one has announced it.
A board that skips hands-on use of a technology and goes straight to procurement believes it's choosing caution over recklessness. It's the opposite. The decision to govern from a distance is the moment the steering wheel disconnects — and the fix boards reach for next gets captured by the same mistake, dressed as adaptability.
Every knowledge worker is getting a personal AI agent. Most companies are building a corporate one. The layer nobody is thinking about sits between them — and it's already forming.
The dominant AI discourse is a performance ledger — costs, jobs, benchmarks. It is a necessary conversation. It is also a profoundly small one. Every significant compression of the gap between imagination and execution has been met first with anxiety, and only later understood as an opening. We are at that moment again.
The logic of industrial production — make at scale, distribute across supply chains — is being quietly dismantled by converging technologies. AI, bio-manufacturing, 3D printing, and quantum computing are pointing toward a world where you buy a licence to produce rather than an object. The factory may be ending. The questions it raised may not.
We treat dates the way Buzz Lightyear treats infinity — as destinations we can name with conviction. But a date is almost never a forecast. It is a promise dressed as one, and once it is fixed, it stops describing the work and starts bending it toward itself. This is the quiet violence underneath most large plans.
For centuries we classified information because we couldn't process it. The labour of order became indistinguishable from the labour of thought. AI quietly inverts the relationship — and exposes a cost that runs deeper than jobs.
The worry about AI and curiosity has the argument backwards. The question is not whether the tool makes you passive. It is whether you were curious before you picked it up.
Information-intensive work has quietly borrowed against our mental reserves for decades. AI might finally offer a way to pay it back.
Most high-skilled cognitive work is spent at the bottom of the decision hierarchy — on choices that don't deserve the capability being used. AI's real promise isn't replacing judgement. It's absorbing the clerical load at the bottom of the hierarchy, and returning the mental resources that load was quietly consuming.
The path used to be visible. Now it is not — and that might be the point.
The tyranny of minor decisions is not a small problem. It is the thing quietly consuming the best of your mental life.
Everyone is asking how fast AI can work. Nobody is asking how fast humans can decide. That's the real bottleneck.
Essays and quick takes on AI, decision-making, and the shape of what's coming. Published when something's worth saying.