On AI and work

Automation was always going to come for some jobs. The more interesting question is which parts of which jobs we've chosen to keep — and why those choices tell on us.

The wrong half, automated first

We keep building tools that take over the parts of work people found meaningful — the drafting, the designing, the composing — while leaving the administrative overhead largely untouched: the approvals, the status updates, the meetings about the meeting. If automation followed value to the people doing the work, it would go the other way. That it so often doesn't suggests the tools are optimized for a different value — usually the employer's measurable output — rather than the worker's experience of the job. That's not a scandal, but it is a choice, and it's worth naming as one.

The second employer

A growing share of work now runs through software that doesn't just help you do the job — it measures you doing it, ranks you against your peers doing it, and quietly reshapes what ‘doing it well’ means, all without ever appearing on an org chart. That software is, functionally, a second employer: one with real authority over your day and none of the accountability of the first. The less visible that authority is, the less anyone thinks to question it.

The skill you didn't know you were losing

Every tool that removes an effort also lets a muscle go slack — the mental map you no longer build because the app routes you, the argument you no longer have to construct because the summary hands you a conclusion. Not all atrophy is a loss; plenty of effort is worth automating away entirely. But some of it is the effort itself that was building the skill, and a tool that removes the effort quietly removes the skill too, long before anyone notices it's gone. Worth asking, deliberately, which muscles in your own work are worth exercising on purpose.

What doesn't automate

The tasks that resist automation longest tend to be judgment under ambiguity, accountability for a decision, and the kind of trust that only accumulates between people over time. None of those show up cleanly in a productivity metric, which is exactly why they're easy to undervalue in the moment and expensive to have lost later. If there's a practical takeaway across these essays, it's this: the parts of a job that are hardest to measure are often the parts most worth protecting on purpose. The habits of trust that make delegation to a machine reasonable — or not — are explored further in On AI; the civic-scale version of “whose time gets protected and whose doesn't” is On technology and society.