← Data Into Labor · All essays

Data Into Labor · Essay 05

The Overlap

The Intelligence Overlap: Why A.I. Got the Job Market Wrong

July 2026 · AJ Orbach

By now AI was supposed to have put a lot more people out of work.

It was not a crazy prediction. A model can write code, make an ad, summarize a contract, diagnose a bug, and pass tests designed for smart humans. Those are things people get paid to do. If a machine can do them in seconds, you would expect companies to need fewer people.

But that is not what has happened, or at least not at the scale people expected. There is real pressure in some places, especially on young workers in jobs full of routine digital tasks. The aggregate effect is still surprisingly small.^[1] The usual explanation is timing. Firms move slowly, and new technology takes years to spread. This is surely true.

But I suspect it is not the main reason. The job-loss prediction assumed that machine intelligence would grow into the same shape as human intelligence. So far it has grown into a different shape.

Circles

Most predictions about AI and work begin with a picture of two circles. The large circle is everything a human can do. The smaller circle is everything a model can do. As models improve, the smaller circle expands inside the larger one. Every job it covers disappears.

This picture treats intelligence like height. A person or a model simply has more or less of it. Benchmarks encourage us to think this way because they turn ability into a score. But a job does not hire a score. It hires an odd collection of abilities that have to work together under real constraints.

Anyone who uses these models for real work runs into this quickly. A model solves a hard programming problem, then misses what seems like the obvious implication of a short email. It explains a tax rule perfectly, then applies it with great confidence to the wrong case. Tasks that look equally difficult to us can sit on opposite sides of its competence. This is what researchers call the jagged frontier.^[2]

The picture I keep coming back to is two coastlines instead of two circles. Human intelligence has one shape and model intelligence another. They share a large and growing area, but the boundary is full of narrow inlets and strange islands.

The boundary matters more than the area. A model may cover ninety percent of the visible work in a job. If the uncovered ten percent determines whether the result is safe, useful, legal, persuasive, or even pointed at the right problem, the person stays.

The last ten percent becomes the whole job.

Weights

The shapes differ partly because humans and models learn differently.

A language model takes a large fraction of the recorded world and compresses patterns from it into billions or trillions of numerical parameters called weights. The weights are not tiny documents, and the model does not retrieve one when it answers. They work together to shape a probability distribution over what should come next. Calling this recall is convenient, but slightly misleading. It is closer to reconstruction from a lossy compression of human expression.^[3]

I do not think this makes the model uncreative. If it combines old ideas in a configuration nobody has produced before, the result is new in the ordinary sense of the word. Human creativity also recombines things. There is no rule saying the ingredients of a new idea must themselves be new.

What differs is where the creativity appears. A model has absorbed more styles, examples, arguments, and fragments of code than any person could. It can move across that space with astonishing speed. But it does not feel the pressure that gives human thinking its shape.

I would not say humans learn only from first principles. Most of what we know is borrowed from other people. The difference is that reality can force us to throw the borrowed answer away and start again. The machine breaks. The customer will not buy. A child asks one more why. We have bodies, goals, reputations, deadlines, and consequences. These do more than supply information. They decide what matters.

A model can generate one hundred plausible ideas. A person still has to notice that the ninety-seventh fits this company, this customer, this week. The hard part has moved from making possibilities to caring which one survives.

Bundles

Jobs are bundles of tasks, and task automation is not the same as job automation. Economists have been pointing this out for years.^[4] The distinction matters more with AI because the tasks it automates are often the most visible ones.

We call someone a copywriter because we can see the copy. The job also includes learning what the product actually does, extracting preferences the client cannot state, rejecting claims that will cause trouble, noticing that the brief is wrong, and accepting responsibility for the result. We call someone a programmer because we can see the code. The job also includes deciding what should be built, finding the constraint nobody wrote down, fitting a change into an old system, and recognizing when a passing test proves the wrong thing.

Jobs are usually named after their most legible output, not their hardest responsibility.

When AI removes part of the bundle, the rest of the job does not remain still. The worker spends less time producing a first draft and more time specifying, choosing, checking, connecting, and explaining. Firms also attempt work they could not previously afford. When software becomes cheaper, more software gets written. When analysis becomes cheaper, people ask more questions. Demand is not fixed.

None of this saves every job. The overlap is widest in work that is routine, digital, and low in context. Entry-level roles often contain a lot of exactly this work, which may explain why some of the earliest employment pressure is appearing there.^[5] The more immediate danger may be a broken apprenticeship ladder. If AI does the beginner tasks, how do beginners become experts?

A broken ladder can do real damage even if the economy still has plenty of work.

Plumbers

Travis Kalanick recently proposed a useful thought experiment. Imagine that everything in the world has been automated except plumbing. In that world, plumbers do not become a tiny remnant of the old economy. They become the constraint on the new one. Every new house, factory, and city waits for them. As everything else gets cheaper and faster, plumbing gets more valuable.^[6]

Intelligence has plumbing too.

It is the work between the obvious outputs: figuring out what a vague request means, finding the missing context, dealing with the exception, persuading someone to act, and checking whether the result survived contact with reality. These abilities are sometimes called soft skills because they are hard to measure. In practice they are load-bearing.

You can see the bottleneck move as soon as output gets cheap. A manager can generate ten plans in the time it once took to write one, but the company may still be able to execute only one. An engineer can produce code faster than the system can safely absorb it. The scarce part was not always the thing we had named the job after.

If a model can do everything except decide what should be done and whether it worked, those two tasks expand until they fill the day.

This is why a technology can be astonishingly capable and still create less unemployment than expected. It accelerates the overlapping work, which pushes more economic activity toward the non-overlapping work. The human part becomes the bottleneck, and bottlenecks attract resources.

Scaling

The frontier will move. Models will get larger, and the systems around them will get better at remembering, using tools, and checking their own work. Companies will also redesign jobs around the shape of the machine. A complete AI system may automate a workflow that a raw model could not.

So mass displacement is possible. But it requires a stronger condition than most forecasts admit. It is not enough for a model to perform the central task in a demonstration. The whole reliability-weighted bundle has to fit inside the frontier, or the surrounding system has to cover every important gap, at a cost low enough for firms to remove the person.

A real breakthrough could satisfy that condition. A system that learned continuously from the world, handled unfamiliar failures, planned for a long time, and knew when it did not know would cover much more of the human coastline. If that happens, the labor question changes quickly.

If progress comes mainly from doing more of what we do now, my guess is that the shared area will grow faster than the boundary smooths out. Models will do more tasks and make individual workers far more productive. They will also keep failing in ways that require human judgment, context, and ownership. Jobs will be rebuilt around those failures.

This would produce a strange-looking economy: AI everywhere, output rising rapidly, some career ladders collapsing, and most people still working. From inside a single task, the technology will look like total replacement. From the labor market, it will look more like a continuous reorganization around the jagged edge.

Boundary

We asked whether AI could do human work. That question was too abstract. A company removes a person only when every binding part of that person's job lies inside the overlap, the system is reliable enough to trust, and the extra output does not create more demand elsewhere.

For many jobs, that is still not true.

AI will still take a great deal of work. Some skills will become cheap. Some careers will become harder to enter. People whose whole job already lies inside the overlap will be displaced. But while model intelligence remains jagged, the economy will keep reorganizing around the parts that stick out.

The machines do not need to think like us to be useful. They only need to overlap with part of what we do. But usefulness and replacement are not the same thing.

Everything outside the overlap becomes the job.

← Back to all essays