Jagged Disempowerment

Most talk about AI and work assumes a slow fade. Roles erode, tasks thin out, the lights dim gradually. That’s the macro story. It’s true enough to be comforting. Gentle enough not to activate the lizard brain.

On the ground it won’t feel like that. Progress comes in jolts. One week your workflow is intact. The next, AI progresses and eats it in a single gulp.

Because with AI, capabilities don’t advance evenly; they lurch in clusters.

That’s why I’ve been thinking about the idea of jagged disempowerment: the lived experience of automation when breakthroughs arrive unevenly, wiping out specific jobs and workflow clusters overnight instead of dissolving them gradually by degrees.

It’s based on the concept of the jagged S-curve (or the jagged frontier of AI capability), the realisation that when it comes to AI, some tasks leap forward while their — often simpler — neighbours barely move.

But when the S-curve shifts, it swallows.

We’ve already had previews. When Google switched Translate to neural nets in 2016, quality jumped across dozens of languages overnight. A human-led cottage industry didn’t slowly decline; it lurched, swiftly double-tapped by advances in natural language processing upon the arrival of ChatGPT.

That same pattern is projected forward.

Epoch AI’s recent report shows that if scaling continues through 2030, we should expect further jolts: coding assistants that go from autocomplete to autonomously fixing software; mathematical models that flesh out proofs rather than just sketching steps; biology systems that can answer lab protocol questions outright.

Zoom out and the curve still looks smooth. That’s the gradual in gradual disempowerment: capability accrues, institutions re-optimise, and leverage declines because labour is needed less to function. The economy, culture, and the state keep working — just with fewer humans in the loop. The long arc of erosion.

However, zoom in and you get the staircase. Irregular thresholds. Some tasks activate inside the S-curve while others lag behind. That unevenness is what many workers will feel: not a slope, a series of drops as domains suddenly find themselves caught by the eye of the Sauron.

Works in Progress released a good deep-dive last week looking at how tales of radiologies demise have been greatly exaggerated. Salaries rose, residencies filled, throughput expanded. Why? Only a slice of the job is — for the time being anyway — cleanly “AI-able”. Liability, accreditation, and integration friction keep a human signer on the line. And when scans get cheaper and faster, we order more of them. Elastic demand soaks up capability instead of spitting people out.

That pattern is real — but it’s not universal. Most junior work for instance lacks those moats. Drafting, summarising, reconciling, templating: the very tasks that built the apprenticeship layer are the ones models eat first, hence the “jobocalypse” in graduate hiring.

Some analysts describe the “diamond firm,” where the base of the organisation thins as junior jobs are automated but the middle swells with integrators and validators.

However, there’s no reason to think AI will stop eating its way up the ladder. Yes we will have humans in the loop for some time, but as I’ve written before, that is but a shoestring stopgap. Managers will be hollowed out as soon as coordination and reporting are agentified. In this instance, the pyramid could fold in on itself entirely, ultimately collapsing the structure (or at least resizing it by a negative magnitude) rather than simply reshaping it.

• Macro: institutions gradually untether from human dependence via gradual disempowerment.

• Micro: workers experience shocks when specific task clusters cross capability thresholds via jagged disempowerment.

• Meso: firms tilt toward diamonds as the base erodes — but those, too, are temporary forms, already beginning to collapse.

What to do with that? Radiology shows the lesson: aim for roles where no more than a third of the work sits inside today’s S-curve of AI capability. Let the other two-thirds be strategy, interaction, and judgement — the messy bits still outside its reach. Because when the curve shifts, the inside third can vanish overnight. If that’s all your job is, you fall with it. If it’s only a slice, you can absorb the shock.

Every day more tasks are pulled into its gravity, and both diffusion and advancement isn’t accidental — it’s the explicit aim of governments and firms racing to deploy AI at scale.

This being the case, policy needs to stop pretending the path is a gentle slope. Plan for cliffs. Build insurance mechanisms for sudden transitions. Make benefits portable so they move with people. And, for the time being at least, train workers to supervise and steer systems, not compete with them on the same ground.

Radiology is fine. Good. It proves buffers exist. But the headline should not lull anyone. The curve may be smooth from a distance. Up close, it’s teeth.