May Your Job Be Messy

If your job isn't messy, A.I. can do it -- if not now, then soon

Source: London School of Economics

All my career, I steered clear of messy jobs. Jobs that involved managing people, negotiating, deepening relationships with business partners, taking risks, prodding organizations in new directions. Not for me. My favorite thing was finding out new stuff, more or less on my own, and then telling people about it. It could be challenging at times, but there was a clear beginning, middle and end to each piece of work. It was fun and a good living.

Times have changed. Artificial intelligence has exceeded the skills I spent decades mastering. It finds out things far faster than I ever could, and writes in seconds. The writing tends to be formulaic, but it’s serviceable for most purposes. And it’s free, or nearly so. I sometimes feel like an expert archer the day he encounters an enemy with a musket.

Luis Garicano has an explanation and a solution that applies to you as a current or future worker. The Spanish economist says that you really don’t want a job that A.I. is good at doing. You want a job that A.I. is bad at doing. And that, he says, is a messy job. “Messy Jobs: The Work That A.I. Cannot Reach,” is by him, Jin Li, and Yanhui Wu. It was published last month.

I just watched a fascinating presentation that Garicano gave as part of Markus Brunnermeier’s Markus’ Academy series at Princeton University. I recommend watching the whole thing, although it contains a fair bit of econ talk.

Garicano has a Ph.D. in economics from the University of Chicago and is a professor at the London School of Economics. He’s also a past member of the European Parliament.

The book’s core move is to take an old piece of economics — Ronald Coase’s 1937 theory of the firm — and apply it one level down, to the job instead of the company. Coase asked why firms exist at all, instead of everyone simply contracting with everyone else in the open market. His answer: coordinating through markets is costly, so it’s often cheaper to bundle work inside a firm.

Garicano, Li, and Wu ask the same question about a single job: why do wildly different tasks — diagnosing a patient, filing the paperwork, reassuring the family — end up bundled into one person’s role instead of being split across specialists? For the same reason: Bundling avoids coordination costs. But with A.I., coordination costs are low enough that unbundling can start to make sense.

In his exchange with Brunnermeier, Garicano recounted the story of a graduate student who told him she could finish a Ph.D. in 10 days instead of four years using A.I. Garicano pointed out to her that everyone else can do the same thing, so a thesis of that caliber is no longer worth what it was. (That’s analogous to a piece of journalism that was once pretty special becoming commoditized.)

Garicano talked about an autonomy threshold. Below it, A.I. is an assistant or partner. Above it, A.I. completely takes over. The messier the work, the higher the threshold is.

Tier one is a single, clean, verifiable task: a coder writing a function in a well-specified program, a junior lawyer drafting a routine contract clause, a writer doing whatever the heck writers do. Here the autonomy threshold is already being crossed.

Tier two is a bundle of tasks, some automatable and some not: a nurse, a store manager, a factory manager. Here the cognition part, which A.I. can do well, is fused with social, physical, and relational work that people are (still) better at. These jobs are still mostly below the autonomy threshold.

The highest, tier three, is careers that involve a bundle of relationships, not just tasks: an entrepreneur, a senior manager, an implementer. Their value rests on authority, trust, and being human. An example is rebuilding an organization around a new technology. You have to change the tasks, the job boundaries, the pay structures, and the reporting lines — all at once, with no way to test the pieces separately first. You have to have tacit knowledge and you have to be trusted. Not things that A.I. is close t having.

In 2016, Geoffrey Hinton — one of the founders of deep learning, who went on to share the 2024 Nobel Prize in physics — said radiologists should stop training immediately, because A.I. would soon outperform them at reading scans. Yet today radiology is one of the highest-paid, hardest-to-fill medical specialties. That’s because a radiologist doesn’t just read a scan (tier one); she talks to the patient, consults the surgeon, trains residents, and signs the diagnosis, legally, with her name on it (tier three).

A.I. can take over a job two ways. Sometimes it captures the whole bundle. More often, it peels off parts of the bundle one at a time until only a shell is left. When A.I. takes over the part of a job that used to require real expertise, more people can do the remainder, so wages get driven down.

Economists Gary Becker and Kevin Murphy, keying off Coase’s work, once said you’ll never find a historian who specializes in just a few years of history. Garicano, taking the concept to an extreme in his talk, imagined a historian who was an expert on May 13, 1775. No matter how much depth that specialization would produce, the cost of coordinating with the May 12 and May 14 specialists would swamp the benefit. Coordination costs are what cap how far specialization goes — and by extension, what keeps unrelated tasks bundled inside one job. A.I. changes that math, because coordinating with an A.I. collaborator is nearly free.

It’s tempting to think “my job is unpredictable and social, so I’m safe.” Garicano pushes back on that. A store greeter or a front-desk receptionist has a genuinely messy job — every visitor wants something different — but they still risk getting replaced by a kiosk, because the cognitive demand underneath the mess is low. Messiness protects a job only when it’s paired with cognitive difficulty.

So far, there’s no mass unemployment, but junior roles are disappearing. Junior employees have always done the unglamorous grunt tasks as an implicit trade: low-value labor now in exchange for the on-the-job training that turns them into a senior professional later. A.I. is very good at exactly that scut work, and once it’s worth nothing, juniors have no way to “pay” for the training that used to come bundled with it. Garicano told Brunnermeier that universities will have to pick up training that used to happen on the job, or new graduates will find themselves paying for internships instead of being paid for them.

A sliver of good news is that as cognition gets cheap the same way artificial lighting and food did, what we spend the freed-up money on is increasingly “human stuff” — status goods, authenticity, live performance, a sommelier’s recommendation, the stuff people pay a premium for specifically because a person made it. People still play chess and watch others play even though computers can trounce even grandmasters.

The question to ask is not “will A.I. take my job?” but “what, in my job, is still genuinely scarce?” The answer is probably the messy part.