Essay

A new MIT paper found the variable the AI-jobs debate keeps missing: the shape of the job

Comparative advantage has been the load-bearing assumption under every AI-and-jobs argument of the last three years: automate the routine, keep the skilled, and the market sorts itself by how hard a task is to do well. A paper posted to arXiv on June 14 by Mert Demirer, John Horton, and Peyman Shahidi at MIT, with Nicole Immorlica and Brendan Lucier at Microsoft (Immorlica splits her time with Yale), now accepted to this year’s ACM EC and carried as NBER working paper 34859, argues that the assumption breaks the moment AI can chain steps together. Their own words: “comparative advantage logic can fail with AI chaining.” That is a conditional, and the condition is the interesting part: where chains form, the old sorting logic stops deciding. Name the interests before leaning on any of it: two of the five authors are paid by Microsoft, whose products do the chaining, and the NBER page carries Horton’s own disclosure that he is a paid advisor to Anthropic, whose usage data the empirics run on. Keep reading anyway; the headline prediction was tested against a thousand placebo reshuffles rather than asserted, and the authors name their data’s limits more plainly than most of the coverage will.

Here is the model. Production is a sequence of steps. Each step can be done by hand, done with AI helping a person along, or done by AI end to end, no human in the loop. AI does not automate whichever steps happen to be easiest. It automates contiguous runs of steps, “chains,” and a firm decides how to bundle those steps into jobs by weighing the gain from specializing a worker narrowly against the coordination cost of managing more people doing narrower slices. Three predictions came out of that model, and all three held up in usage data: O*NET task workflows crossed with Anthropic’s Economic Index of real Claude conversations. Correlational, the authors say plainly, and the workflow orderings are themselves LLM-generated, but the patterns sit in the extreme tails of the placebo distributions: AI-executed steps cluster next to each other rather than scattering evenly across a job, a job whose AI-exposed steps sit spread out sees less AI execution than one where they sit bunched, and a step next to an AI-executed step is measurably more likely to be AI-executed itself. Automation spreads sideways, like a stain.

Run that model against the skilled specialist and the reassuring old story stops holding. A radiologist reading a scan and a claims processor filing a form used to sit at opposite ends of the skill ladder; comparative advantage said protect the radiologist, automate the processor. Chaining asks a different question: whether the steps around the scan-read sit contiguous enough to swallow whole, intake, image, read, report, in one unbroken run, or whether the processor’s job is stapled to three unrelated human judgment calls that break up any chain long before it reaches production scale. Skill still decides whether any single step can be automated at all. Adjacency decides whether the job around it gets swallowed.

I have made a version of this argument before, on different grounds. Range is the new moat: I have written that you must “broaden one’s horizon at a geometric rate, to stay relevant in the AI age.” The chaining paper hands that belief a mechanism it did not have. A career built as one long contiguous run of adjacent, similarly automatable tasks is exactly the shape a chain swallows whole. My own transcript made the case in the first person before I ever made it in prose: physics and math with computer science at MIT, coursework spanning ten departments on the same record. Accidentally, the correct shape.

Now the honest problem with turning that into a life plan, my transcript included. The paper’s chain lives inside one firm’s step-bundling decision for one job, not across a person’s stacked side projects in unrelated fields. Doing physics on Monday and running a trading model on Tuesday does not interleave anything. It hands you two separate jobs, and if either one is internally a clean contiguous chain, that one gets fully automated on its own schedule regardless of what you did the day before. Breadth across domains does not, by itself, reproduce the paper’s mechanism. That is a real gap, and I will not paper over it.

The strongest objection is sitting in payroll data. The paper observes usage, and nothing in its dataset shows a single person losing a job. The best displacement evidence anyone has, Brynjolfsson, Chandar, and Chen’s read of ADP payroll records, sorts by experience: workers aged 22 to 25 in AI-exposed occupations are down 16 percent relative, while experienced workers in the same occupations, same workflow, same adjacency, held steady. A job-shape variable cannot see that gradient, comparative advantage in tacit knowledge explains it directly, and I concede that as of today the market is still sorting the old way. Here is what the concession leaves standing: inside those same occupations, the work the juniors actually do is the contiguous run, the intake, the first pass, the routine read, and the ADP declines concentrate in occupations where AI automates rather than augments. The chain lens predicts which end of the job the swallowing starts from, and that is the end the gradient shows. Radiology makes the same point from the other side. It is the most chain-shaped workflow in medicine, and a decade after Hinton said radiologists were obsolete, average pay sits at $571,000, up 9 percent in a year, with 7,469 open positions. The chain through the reading room is essentially assembled. What stops it is a licensure regime that requires a human signature on the final read and a liability system that needs a named person to sue. Institutions cap chains too, and the model has no slot for that.

What survives needs the model’s fine print. In the model you do not draw the job boundaries; the firm does, and the paper’s long run is firms redrawing them the moment AI quality crosses a reorganization threshold. Scatter protects you only while the org chart holds still. The paper is even crueler about which human survives a finished chain: the verifier standing at its end, priced at a lower skill requirement and the wage to match. What no firm can redraw is the sequence of production itself. Past this point the paper goes quiet and I am extrapolating: weight what follows as my bet, not their result. The position I would buy is owning steps AI genuinely fails at, sitting in the middle of the sequence, because a chain that must pass through a step it cannot execute is capped there no matter how the boundaries get drawn. And the bar is genuine failure: the model pulls a step into the chain even when a human does it better in isolation, because verifying one long chain is cheaper than managing one more handoff. A step a human merely does better is chain food. The steps that hold are the ones that need a person to be trusted, held accountable, and kept in relationship with the people on the other end of the work, the “relationships, trust, accountability, judgement” I have argued are the last thing the falling ivory tower of book knowledge cannot cheapen. Braid those through the middle of your sequence and every chain caps early. Stack them at the end and you are holding the clipboard, at the price the model quotes for that seat.

Comparative advantage told you to go deep and trust the market to protect the deep. That market is still running, still sorting by skill. The chain assembling under it reads something else: adjacency.

Sources: