I was the geeky kid who was obviously supposed to study computer science. I didn’t. I studied History and Philosophy of Science (HPS) instead, and for most of the two decades since, I quietly filed that decision under “mistakes I got away with.” I went on to co-found a software company. I wrote a third of the original product myself. The degree, I assumed, was decoration.
Then The Economist published this chart.
Look at where computer science sits: among the most exposed to AI, with the share of new graduates in full-time work down roughly nine points in two years. Now find philosophy: top left, above the zero line. The majors least exposed to AI barely moved. The most exposed fell off a cliff on average (not every bubble: accounting sits at high exposure and is fine), and The Economist’s time series dates the drop to after ChatGPT’s launch. Yes, 2022-24 was also the rate-hike hiring crunch, but the crunch hit everyone; the slope against AI exposure is the part the rate cycle can’t explain.
My first reaction was a laugh. My second was more uncomfortable, because the chart is not really saying “humanities won.” It is saying something more precise: the vocational skill got commoditized and the meta-skill didn’t. Writing Python was always the automatable layer. Reasoning about why you’d build the thing, what counts as true, and where a system’s assumptions break: all of that lives one layer up from the code. That is the layer AI is pushing everyone into. And that, almost embarrassingly, is what History and Philosophy of Science drills.
First, a confession
I was not a good student. I skated the essays. I did not leave university with Popper and Kuhn lovingly annotated on my shelf. If the value of my degree had been the grades or the content, I would have graduated with nothing.
The value was the mode of thinking, and the strange thing about modes of thinking is that they seep in even when you’re coasting. Sit in enough seminars where the only question that matters is “how do you know that’s true?” and the question installs itself. It waited twenty years for its moment.
Everyone treats “how do I become a strong AI player” as a tooling question. Which model, which prompt, which agent framework. It is not a tooling question. It is an epistemics question.
Epistemics just means the study of how you know what’s true. AI produces infinite fluent plausibility. It will hand you a confident answer on anything, beautifully structured, entirely wrong some meaningful fraction of the time. The scarce skill of this decade is not generating output. Generation is now free. The scarce skill is telling confident-sounding from correct, at speed and at scale. There is an entire academic discipline about exactly that problem, and it has been sitting in the course catalogue the whole time, mislabeled as useless.
Popper, twenty years late
I run StoreHub AI-first. My actual day is spent directing fleets of AI agents: they draft, build, analyze, and review, and my job is to decide what to trust. Somewhere along the way I noticed that the operating system I’d built for this was not engineering doctrine. It was falsificationism with a command line.
Karl Popper, the philosopher of how science catches its own errors, had one core move: a claim is only as strong as the test that could break it. In my setup, that stopped being seminar material and became literal, executable text. Every piece of work an agent claims is “done” must ship with its proof in the same turn: the test that passed, the render that was checked, the number reconciled against a second source. No proof, not done.
When agents report findings, I spawn other agents whose only job is to refute them; a finding survives by resisting refutation, not by sounding right. When I change the instruction files that govern how my agents behave, I don’t ship the change because the new wording reads better. I run the old rules and the new rules against the same tasks and keep whichever one actually behaves better. A rule earns its place by surviving attempts to kill it.
None of that came from a software textbook. It is a philosophy of science exam answer, wearing a terminal.
Thomas Kuhn, the historian of how science changes its mind, supplied the other half. His account of paradigm shifts is a four-beat story: an anomaly shows up, the incumbents explain it away, the anomalies pile up, and then the whole framework flips at once. Afterwards, everyone pretends they saw it coming. If you want a description of how established companies are responding to AI right now, that is it, beat for beat. Most people are still doing what Kuhn called normal science: working harder inside the old paradigm. The operators winning right now are the ones treating the anomaly as the signal.
What you can actually train
The honest version of my story is not “I chose wisely at 20.” I didn’t. I chose awkwardly, coasted, felt dumb about it, and got the benefit anyway because thirteen years of running a company forced the rigor in. You don’t need the degree, and you certainly don’t need to be good at it. The muscles are trainable directly, and here is what changed: they used to cost willpower, which is why almost nobody kept them up. Now you can delegate the bookkeeping to the machine and keep only the judgment. Three moves:
Move 1 · Popper
Ask every AI answer for its falsifier
Don't run the checks yourself. Make the AI name what would prove its own answer wrong, then go run that check before reporting back. An answer that can't state its falsifier is plausibility, not knowledge. Wire it into the workflow as a standing rule, so "done" always arrives holding its proof and you never have to remember to ask.
Move 2 · Calibration
Have the AI keep your decision journal
Not a notebook. Since your real calls increasingly happen in conversation with AI anyway, instruct it to log each one as it happens: the claim, your prediction, the one thing that would prove you wrong, a review date. Then have it score the record monthly and surface where your confidence ran ahead of your hit rate. The journal keeps itself; you just read the scorecard. It is the seminar question pointed at your own track record.
Move 3 · Kuhn
Steelman the option you're rejecting
Before committing to a decision, spawn an agent whose only job is to argue for the path you're about to decline, as hard as it can. This used to be an exercise in self-discipline; now it costs one prompt and two minutes. It forces you to look at the evidence your current paradigm is quietly explaining away, which makes it the cheapest disruption radar available.
The bad student’s consolation
If you’re watching AI eat the skill you spent years mastering, the chart above reads as a threat. I’d offer a different reading. The layer that got automated was always the automatable layer. What’s left over is the part that was always the actual job: deciding what’s true, noticing when the framework is breaking, and knowing where your model of the world stops working. Those muscles don’t care what your degree says. They care whether you train them.
The 20-year-old who felt dumb didn’t waste the degree. He just hadn’t installed the rigor yet. Twenty years later, with real decisions to practice on and machines generating plausibility by the ton, the excuse is gone.
So the question isn’t whether you studied Plato. It’s whether you can tell confident from correct.