In a recent poll, 80% of Americans think that it is likely that AI could cause unemployment to rise sharply over the next five to ten years.
According to a new book, predictions like that are based on a fundamental misunderstanding of what white-collar workers actually do all day.
I’ve spoken to one of the authors, Luis Garicano, about the book called Messy Jobs: The Work that AI Cannot Reach. He is an economist and professor of public policy at the London School of Economics.
You can read the interview below. It was first published in the Danish business daily Børsen on September 15. This is a version slightly edited for an international audience.
On the surface, it might have looked like the task had been completed in two weeks.
A consulting firm had been asked to redesign a frequent flyer program for an airline, but months after the analysis had been delivered, the proposal still hadn’t gotten off the ground.
Now, the airline’s different management layers had to buy into the plan. Following feedback, the proposal was continually adjusted. The program’s commercial partners, such as hotel chains and credit card companies, also had to green-light it.
A major effort to reach consensus and then implement the program was underway, and it would take six months from the proposal’s conception until the first customers earned points under the new program.
The story is told by Luis Garicano and is based on the experience of a consultant he recently spoke to.
Garicano is an economist and professor of public policy at the London School of Economics. He is also co-author of the book Messy Jobs, subtitled The Work That AI Cannot Reach.
For the professor, the consultant’s work is a good example of a “messy job.”
“Essentially we think a lot of jobs are messy jobs. And that type of work is far from being within reach of computers,” Luis Garicano says.
The Spaniard wrote the book together with Jin Li and Yanhui Wu, both professors of economics at the University of Hong Kong.
We will come back to how exactly the trio defines a messy job, but the phenomenon is the reason they believe “doomsday predictions” of AI-induced mass unemployment are wrong.
The technology will affect the labor market, yes, but people making such predictions do not understand what most white-collar workers do all day, they argue.
excitech has read the book and interviewed Garicano to understand how he and his co-authors reached their conclusion.
A mistake to generalize
One somber prediction about AI’s impact on the job market came from Anthropic CEO Dario Amodei in May of last year:
The technology could eliminate half of all entry-level positions for white-collar workers within one to five years and push unemployment up to 10–20 percent.
When Amodei made his prediction to Axios, unemployment in the United States stood at 4.2 percent — now, 16 months later, it has fallen to 4.1 percent.
According to Luis Garicano, Dario Amodei and other Silicon Valley figures make the mistake of taking developments in their own field — programming — and generalizing them to all knowledge work.
“Machines are amazing at coding, but why? Coding is a simple, single task. You can easily verify whether the code works. You do a few tests, see it’s working, and if it is, you let the computer do it.”
But it is a fallacy to assume that all knowledge work similarly has an end product for which it is possible to objectively measure whether the task has been completed correctly.
“They don’t put enough weight on the irreducible messiness of the human world.”
Garicano has spent most of his career studying how technology affects organizations. For example, he co-authored a major analysis of how the rollout of IT systems in American police departments between 1987 and 2003 affected their productivity.
Throughout his studies, the professor has repeatedly seen how slowly change occurs.
“People don’t just adopt technology immediately. There is a lot of friction in that process,” he says.
Tasks or jobs
According to Garicano, understanding how AI will affect the labor market requires making an important distinction between jobs and tasks. A job consists of several different tasks — like the consultant, who has to conduct an analysis but also attend meetings and present their work.
Fundamentally, Garicano says there are two factors that determine whether a job is messy.
The first is whether the tasks that make up the job are difficult to separate from one another.
He cites sales as an example. There is a significant cognitive component involving understanding the customer and what product they are looking for. The salesperson can use AI to help with this, but it’s still something they will also have to do themselves.
“In order to meet the potential customer for dinner, I need to be able to think ‘Oh, he wants this other thing’, and then remember that we have this other product.”
The second factor that makes a job messy is whether it involves managing a complex network of human relationships.
The book gives the example of a chief engineer at a factory who, among other things, has to hire new employees, negotiate with executives about implementing proposals, communicate with the municipality about obtaining building permits, and so on.
By contrast, your job is vulnerable if it consists of a single isolated task where it is possible to objectively verify whether it has been completed correctly.
Examples might include technical translation, drafting standard contracts, or simple bookkeeping tasks where the numbers have to add up.
Garicano emphasizes that this is not about cognitive complexity. Chess, for example, is a complex game, but AI had already surpassed humans at it when IBM’s supercomputer Deep Blue defeated the legendary Garry Kasparov in 1997.
“The machine could learn chess very easily because it is verifiable and a single task.”
If you have that kind of job, you may have to rely on regulation for protection.
Thanks in part to resistance from labor unions, for example, 20 years passed between the invention of the automatic elevator and the point when the last manual elevator operator lost their job.
Another lifeline is demand for human authenticity.
Garicano and his co-authors give the example of a sommelier in a restaurant. A guest could easily handle the task of matching a wine with a dish, a budget and a preference using a wine-recommendation app. But the sommelier provides something else.
“The sommelier is not selling information. She is selling a performance: the story about the vineyard, the gesture of pouring the first taste,” they write in the book.
Three types of knowledge
But even in messy jobs, couldn’t you simply have an AI attend every meeting, listen in and make decisions?
No, says Garicano, arguing that many of the decisions that may appear to be made in meetings have actually already been settled elsewhere.
There may be political reasons or internal power struggles that are not reflected in what is actually said during the meeting.
“You could record a meeting and train an AI model on the recording, and you would completely miss what is going on.”
Garicano calls this strategic knowledge. It is like selling a home. You know the minimum price you are willing to accept.
“This knowledge you’re not going to tell the buyers. In a context of work, there is a lot of information that is strategic and that is really withheld on purpose for incentive reasons.”
Another type of data that can be difficult to obtain — and therefore difficult to use to train an AI model — is what the Hungarian philosopher and economist Michael Polanyi called tacit knowledge: We know more than we can or are willing to tell.
“If I had to explain to a machine how I ride a bicycle, I wouldn’t really be able to,” says Garicano.
Like the senior consultant who knows that a client will reject a particular proposal without being entirely able to explain why.
Finally, there is a third factor that can make knowledge difficult to share: It often does not come into existence until we have the conversation.
“Marketing thinks something, sales thinks something, they start chatting and in the conversation the knowledge is created.”
Not an information problem
Another limitation that AI cannot solve, according to Garicano, concerns power.
London has an acute housing shortage. But the root cause is not a lack of information: We know perfectly well how to build more housing.
What prevents apartment buildings from going up has more to do with regulation, opposition from local communities, bureaucracy and political power struggles.
AI can draft a planning review, but “it cannot convince an environmental group to drop its lawsuit or negotiate with a council member who faces reelection,” the argument goes.
I point out to Garicano that there was recently a study by the UK AI Security Institute and Oxford University showing that AI systems were better at persuading people on political issues than even professional canvassers.
“It’s not so much about can I persuade you that I’m right. It’s that I have my interests, you have yours. You give me a tool that is making me better at complaining, I’m going to be better at stopping building.”
If productivity is to increase, those bottlenecks need to disappear. Instead, they often become stronger as technology improves.
“Somebody who saw the French nuclear build-out in the ‘70s would have said the world’s energy problem is going to disappear in 20 years. No, we are even worse than then. It’s not just about the technology.”
It’s about politics, which Garicano argues cannot simply be reduced away.
Take the messy job
So what is the professor’s advice to people who want to future-proof their careers?
Whether you already have a job, are being offered one, or are a student, Garicano says the crucial thing is to consider whether the job has an output that can be systematically verified.
“Does this job have an output which is systematically verifiable as correct or incorrect externally? If there is, and you can train a machine in this way, and that’s the basic thing you’re doing all day, then the job is really at risk.”
“If the job is messy and you’re doing many things and it has some relational components and those are difficult to peel off, then the job is not at risk.”
So the advice is: Try to take the messy job. Again, he emphasizes that he is not talking about complexity, but rather jobs that involve many different things.