Author's note: In 2023 I wrote AI Isn't Coming for Tech Jobs—Yet for Asterisk, along with a companion piece laying out the model. The core thesis held: Whatever the disruptions of AI, the number of tech jobs grew on net. My simulations suggested a median of 5.1M jobs in Computer and Mathematical Occupations by 2025. BLS came in at 5.26M. I still think AI will create more jobs than it eliminates in the near term, but I am less certain now than I was then, and my uncertainty grows the further out I look. This article is my best guess at why we will not see mass unemployment in white-collar work before the end of the decade.
On July 2, Scott Alexander published “The AI Superforecasters Are Here.” Humans had taken the top two places in a recent tournament and an AI system came in third. Given the luck involved, Scott called the best humans and AIs “a statistical dead heat.”
A week later, I saw an announcement from a platform where they paid top forecasters to tackle policy-relevant questions that they were closing down the platform and moving to a project-based model.
While they didn’t mention AI in the announcement, you don’t have to be a superforecaster to suspect it might have had something to do with it.
So if AI can replace forecasters, why won’t other jobs follow? Looking at why forecasting work is being disrupted is instructive.
Forecasting Was Never the Bottleneck
Give a group of top forecasters a clear question and they quickly converge on a reasonable probability, but forecasting didn’t catch on as part of organizational decision-making because it requires the client to do the hard part of figuring out the right questions to ask in the first place.
An executive wants to know: “Will the war spread?” A forecaster needs a precise question with a deadline and resolution criteria. Unfortunately, making the question resolvable can move it away from the thing the executive cares about. This might be potential shipping disruptions. They could be caused by a blockade, an invasion, or a cyber attack. Writing the right question requires understanding the decision it is supposed to inform, the options the organization has, and what information would cause it to change direction. Often, no one person knows.
Bringing in top forecasters creates work and mental overhead. It’s like hiring an intern for a job you do not know how to perform. By the time you specify the task, provide the context, and explain what a useful answer looks like, you might discover that you have higher value options than doing the essential pre-work to make the task possible in the first place.
Automation Moves the Bottlenecks
Unlike forecasting, most jobs unblock bottlenecks.
Consider radiologists. Even if AI interprets scans perfectly, someone still has to resolve ambiguous findings, communicate results, and take responsibility for the report. The job shifts toward the decisions and exceptions that remain.
Software follows a similar pattern. Producing code is only one step between an executive saying “improve the customer experience” and customers using a working product. As code gets cheaper, requirements and organizational approval become the constraints. Developers spend less time writing code and more time clearing whatever blocks the project.
Jobs are bundles of tasks built around bottlenecks. Automating one task may reduce headcount, increase demand, or both.
Human Data Warehouses
Most organizations are like giant, inefficient databases whose most important information exists only in people’s heads.
Consider this hypothetical business situation:
You need the Watkins contract? Debbie worked on it last year. She knows the person at the client who can approve a change. Do not use the official portal; nobody checks it. Debbie gets too many emails, but if you walk by her desk she will tell you everything.
Management theorists call this the garbage-can model: organizations are not machines translating goals into action. Problems, solutions, people, and opportunities slosh around until some combination produces a decision.
Even perfect requirements elicitation from AI does not solve this. Senior leaders are often “big picture” people. They say “improve the customer experience,” delegate the project, and move to the next meeting. They are not going to spend an afternoon playing twenty questions while an AI converts that direction into a specification.
The people willing to answer may not have authority. The people with authority may not know the constraints. The requirements may not exist in any one person’s head, but only emerge when legal, engineering, and the client react to one another. To replace a human role, AI needs to decide who to overrule, when to contact clients, and be trusted to take action when there’s a real risk of being wrong. Organizations built around human accountability will be slow to delegate those powers.
Sometimes that friction protects the organization from its leaders. Imagine leadership decides to raise prices on the company’s largest client. The contract has renewed automatically for a decade because nobody at the client company has to think about it. An employee knows changing the price would trigger a client review, and the client would likely discover that competitors are better. An employee who slow-walks the increase until it’s too late to make the update and management moves on might save the business. On a process map, that employee looks like an obstacle. In reality, they’re part of the organization’s immune system.
Conclusion
To replace a human role, an organization needs to give AI the appropriate context, access to the right people, permission to make tradeoffs, and the authority to act
AND
the right people will have to respond when AI requests things from them.
Most organizations are not even remotely set up to do that.
This is why AI may transform millions of jobs while eliminating very few of them.
Emphasis on the yet - at least, that’s what history would suggest (see also the fate of 18th-century hand spinners: meettheuk.substack.com/p/what-does-the-fate-of-hand-spinning)