Annie Lowrey’s recent Atlantic piece poses a deceptively simple question to white-collar workers anxious about AI: Are you coal, or are you a horse?
It’s worth reading in full. The framework she builds around the Jevons paradox is one of the cleaner pieces of economic thinking applied to the current moment. But it stops short of naming the structural property that determines which category you fall into.
That property is how you think. Specifically, whether you think in systems.
The Framework, Briefly
Lowrey’s argument turns on a 19th-century insight from economist William Stanley Jevons. When steam engines became more efficient, economists expected demand for coal to fall. Instead, efficiency drove down the cost of production, which expanded industrial activity, which consumed more coal than before. Making a resource cheaper to use increases total demand for it. That’s the Jevons paradox.
She applies this to labor. Some workers are coal: AI makes their function cheaper and more accessible, which expands total demand for what they do. Software engineers are her primary example; employment is up 6 percent year over year, in part because companies need people to develop and implement AI products. Radiologists are another: when AI improved medical imaging, it didn’t replace them. It unlocked new use cases for CT and MRI scans, and radiologists were the ones administering and interpreting more of them.
Horses are the other category. When tractors replaced draft animals, demand for horses didn’t expand. It collapsed. The function was substituted entirely, with no Jevons rebound.
Lowrey is careful to note that AI will affect different workers differently, and that the same person can be coal in one context and horse in another. What she doesn’t give you is a diagnostic. She shows you examples but doesn’t tell you what the coal roles have in common.
What the Coal Roles Share
Look at the roles Lowrey identifies as coal or coal-adjacent: software engineers figuring out how to implement AI systems and radiologists interpreting expanded imaging volumes.
None of them are operating in environments that behave predictably.
This is the distinction that matters: complicated versus complex. A complicated system has many parts, but those parts follow rules. You can model it, optimize it, hand it to an algorithm. AI is extraordinarily good at complicated. It synthesizes codifiable knowledge, executes repeatable processes, and finds patterns in large datasets faster and more reliably than any human.
A complex system is different. The parts interact in ways that produce emergent behavior. The rules shift as the system responds to interventions. The same action produces different outcomes depending on context, timing, and what else is happening simultaneously. You cannot fully model it in advance. You navigate it in real time.
Systems thinking is the cognitive practice of operating in complex environments. It’s not a methodology or a certification. It’s a way of seeing: feedback loops instead of linear cause-and-effect, dynamic relationships instead of static structures, emergence instead of predictable outputs. People who think this way do it whether or not they’ve read Meadows or Stacey or Snowden. And people who don’t think this way don’t do it just because they’ve taken a course.
This is why the coal/horse question is ultimately a question about cognition, not job title.
Why Systems Thinkers Are Coal
Ashby’s Law of Requisite Variety states that a system can only be governed by a controller with at least as much variety as the system itself. It’s a formal way of saying that complexity requires a commensurate intelligence to navigate it.
AI increases variety in organizational and technical environments. It accelerates deployment cycles, creates new interdependencies, and generates emergent behavior that wasn’t present before. The environments that engineers, consultants, project/program managers, doctors, and designers operate in are getting more complex, not less, precisely because AI is making it cheaper and faster to build and deploy systems.
That is a direct Jevons dynamic. Cheaper tools expand the surface area of what gets built. Expanded surface area increases the complexity of what needs to be governed, integrated, and interpreted. Demand for the intelligence that can do that goes up.
Systems thinkers are the people positioned to meet that demand. Not because of their titles, but because of how they process environments that keep changing their own rules.
The Honest Diagnostic
If you’re trying to locate yourself on the coal-horse spectrum, the question isn’t what your job is called. It’s how you actually work.
Some people, when they encounter a problem, look for the lever. They want to identify the variable, adjust it, and produce a known output. That’s excellent thinking for complicated problems. It’s also the thinking AI is rapidly absorbing.
Other people, when they encounter a problem, look for the system. They want to understand what’s producing the problem, what’s connected to it, what will shift if they intervene, and what unintended consequences might follow. They’re comfortable holding ambiguity while the picture clarifies. They change their model when the environment surprises them.
The second group isn’t smarter. They’re oriented differently. And that orientation is, right now, structurally valuable in a way that the first orientation is not.
This doesn’t mean every engineer, consultant, interior designer, or doctor is coal. Lowrey is right that the same role can go either way. A software engineer who codes to spec inside a defined system is operating in complicated territory. A software engineer who is figuring out how an AI system will interact with an existing organizational environment, anticipating failure modes, and adapting as the system produces surprises is operating in complex territory. Same title. Different cognitive work. Different position on the spectrum.
The Risk Worth Naming
The threat to systems thinkers isn’t substitution. It’s misclassification.
Organizations under cost pressure tend to flatten roles toward their most measurable functions. If a program manager’s job gets reduced to status reporting and schedule tracking, those functions compress. If a consultant’s engagement gets scoped to deliverable production rather than diagnostic judgment, that compresses too. The systems thinking layer, the part that reads the environment, holds the whole, and navigates emergence gets treated as overhead rather than function.
That’s the scenario to watch for. Not AI doing your job, but your job being redefined until it no longer contains the work AI can’t do.
The Caveat Lowrey Earns
She closes her piece with a sharp observation: coal itself eventually became horse. England, whose Industrial Revolution the Jevons paradox helped explain, now uses as much coal as it did in 1666. Technological succession is real.
Systems thinkers are coal right now. The Jevons dynamic is working in their favor. But the durability of that position depends on staying in the complex layer on doing work that requires genuine systems intelligence, not just work that happens to carry a prestigious title. The orientation protects you. The credential doesn’t.
Nicole
Read Annie Lowrey’s full piece in The Atlantic: How to Guess If Your Job Will Exist in Five Years
Great article, Nicole. I so appreciate clear, succinct writing, especially when dealing with such a fluid subject as the economic implications of AI.
I love this one a lot because of the focus on people -- because it's just hard to train and build people on systemic thinking. And it's getting harder, because those who don't have it can still get it through exposure from their environment -- but those environments are getting more distant.
We're getting more distanced through social media, which is an irony. Short-form content lacks the depth required to build the complexity -- a prerequisite of systems thinking.
People are more polarised, and the space to focus oneself on systems thinking and epistemology is becoming a privilege. I'm glad I spent my focus and time reading this.
Thanks for this 🙏