The One-Way Door That Looks Like a Hallway
AI and the Erosion of Technical Expertise
Every previous technology we've worried about had a moment. Nuclear weapons had Trinity, then Hiroshima, then the moratorium debates. Recombinant DNA had Asilomar, where biologists voluntarily paused their own research until they could agree on safety protocols. These were imperfect moments — late, contested, insufficient. But they existed. There was a legible threshold, a point where society could look at what had been built and ask whether it should continue.
AI will not have that moment. Not because the technology isn't consequential enough to warrant one, but because of a structural property that makes it different from every previous case: there is no threshold to argue about. There is no bright line, no Trinity test, no discrete deployment that transforms the world on a specific date. Instead, there is a continuous function. Each model is slightly more capable than the last. Each integration is locally reasonable. Each delegation of judgment — let the AI flag the anomalies, let the AI draft the analysis, let the AI evaluate the other AI — makes sense in isolation.
No single step is the one where humans lose meaningful control. But the integral of all those steps is exactly that.
The Hallway
Here is a fact that would have sounded absurd five years ago: according to the Federal Reserve Bank of New York, recent computer science graduates now face higher unemployment than art history majors, at 6.1% versus 5.6% as of early 2025 [1]. The pools are different sizes and compositions, and the comparison is imperfect. But the direction is striking.
A Stanford Digital Economy Lab study using ADP payroll data found that employment for software developers aged 22 to 25 has declined nearly 20% from its late 2022 peak [2]. A Harvard study of 285,000 firms and 62 million workers found that when companies adopt generative AI, junior hiring drops 9 to 10 percent within six quarters, while senior hiring stays flat [3]. New graduate hiring at the fifteen largest tech companies by market cap fell 25% year-over-year in 2024 alone, and has dropped over 50% since 2019 [4]. The career ladder in software engineering, the most in-demand field of the last two decades, is being hollowed out from the bottom.
Not through layoffs. Through silence. Companies are simply not hiring juniors anymore.
AI is not the only factor. The post-2021 hiring correction, interest rates, and offshoring all play a role. But those are cyclical. AI is structural. A senior engineer with AI tools can produce what previously required a team of juniors, at higher quality and lower cost. Every individual company making this decision is acting rationally. The aggregate result is a structural crisis that nobody chose and nobody knows how to fix.
The Paradox
The hiring collapse is the visible symptom. The deeper problem exists whether or not juniors get hired, because even those who do are building a different kind of expertise than the one that matters.
The work juniors used to do (debugging simple systems, writing basic features, reviewing straightforward code) was never just about the output. It was a ten-thousand-hour apprenticeship in mechanistic understanding: learning how systems actually work, how they fail, and why. You learned what a good architecture looks like by building a hundred bad ones. You learned where systems break by being the person who broke them.
The optimistic response is that juniors working alongside AI will simply develop a new kind of expertise: pattern recognition on AI output, an intuition for when AI-generated code looks wrong. Ten thousand hours of reviewing and debugging AI output should produce something.
But this assumes AI fails in learnable patterns. Some of it does; you can learn to spot hallucinations, context window artifacts, systematic weaknesses with recursion or state management. But as Terence Tao has observed, AI-generated work can look "superficially flawless" while containing errors that are "really subtle, and then when you spot them, they're really stupid. Like no human would have actually made that mistake" [5]. A junior engineer's mistakes are predictable (off-by-one errors, missed edge cases, naive architectures) and a senior engineer recognizes them because they once made the same mistakes. AI's interesting failures, the ones that matter for system integrity, are novel and context-dependent in ways that human mistakes are not. The failure surface is vastly larger and less predictable, and the failures that matter most are precisely the ones you cannot learn to anticipate through pattern recognition alone.
What you need to catch AI's failures is mechanistic understanding: knowledge of how the underlying system actually works, so you can reason from first principles about what could go wrong. And that is exactly the kind of understanding that comes from doing the work yourself, not from reviewing AI's work.
Anthropic studied this inside their own engineering organization. In a survey of 132 engineers, they identified what they called the "paradox of supervision": effectively overseeing AI requires the deep knowledge and judgment that only come from doing the work yourself, but AI makes it increasingly unnecessary, and economically irrational, to do the work yourself. Their engineers reported that AI use led to measurable skill atrophy, with one noting: "When producing output is so easy and fast, it gets harder to take time to learn." Another described seeing AI propose a solution that looked correct but was structurally flawed in a way that only years of hands-on experience would reveal [6].
The skills you need to catch AI's mistakes are the skills you can only build by making those mistakes yourself. And the economic incentive at every level (individual, corporate, national) pushes toward skipping that process.
This is the core of the problem, and after three years of being documented by researchers at Stanford, Harvard, and the companies living through it, no one has proposed a solution that survives contact with incentives. The ideas that exist (restructured apprenticeships, medical residency models for engineers, appeals to companies to hire juniors as a long-term investment) all require actors to work against their short-term economic interest. When has that ever happened at scale?
The Abstraction That Isn't
But the erosion of expertise, however troubling, is not unprecedented. Technology has been reshaping how people develop skills since the industrial revolution. What makes AI different, and what I think most people writing about the broken career ladder are missing, is a technical property that has no precedent.
Every previous technological abstraction in computing had stable, mechanistically debuggable failure modes. When the industry moved from assembly to C, the compiler was a deterministic translation. If it produced wrong output, you could trace the cause through a defined set of rules. Same from C to Python, from Python to frameworks, from frameworks to cloud services. These layers were not perfect: compilers have bugs, frameworks have undocumented behavior, cloud services fail in emergent ways. Modern CPUs are already complex beyond any single person's comprehension; Spectre proved that speculative execution could produce security failures that nobody anticipated. Opacity is not new.
But there is a difference between opacity and non-determinism. A system can be too complex for anyone to fully hold in their head while still being mechanistically traceable. You can, in principle, follow the chain of cause and effect when something goes wrong, even if it takes enormous effort. This is what made the entire model of technological progress work: knowledge accumulated because each generation could build on a foundation whose behavior, including its failure behavior, could eventually be understood through systematic investigation.
AI is not that kind of foundation. It is a probabilistic system that is usually right. A compiler fails the same way every time and you can reason about why. AI fails in novel, context-dependent ways that even its creators cannot fully predict. You cannot stand on a probabilistic guess the way you stand on a compiler.
This has a practical consequence that I think is the most important and least discussed implication of AI-assisted engineering:
You can only test for failure modes you can anticipate. Anticipating failure modes requires understanding how the system works. If you don't understand the system, you can't imagine how it will fail. Therefore you can't test for it. Testing without understanding is theater. The reason we can fix compilers when they break is not just that they're deterministic. It's that there are people alive who spent twenty years building them. The determinism is only useful because someone has the understanding to exploit it.
And here is where it gets genuinely strange. If AI builds a system today, and a more advanced AI builds on top of it tomorrow, and a third AI builds on top of that, no human at any layer has actually understood any of it. There is no human-comprehensible foundation. It is probabilistic guesses stacked on probabilistic guesses. The first time something breaks deep in the stack, there may be no one who can trace it back, because the understanding was never there.
When Rome fell, we lost knowledge that once existed: how to make their concrete, how to build their aqueducts. But it had existed. People had understood it, and eventually the understanding could be reconstructed from the artifacts they left behind.
AI-generated systems are something genuinely new: artifacts of knowledge that may never have existed in any human mind. If they fail, there is nothing to reconstruct.
The obvious objection is that AI-generated code is right there: inspectable, traceable, diffable. Unlike Roman concrete, you can read every line. But having code is not the same as having understanding. Understanding is a property of human minds, not of artifacts. The question is whether anyone has ever built a mental model of why the system works and how it could fail. If not, the inspectable source code is like a book written in a language nobody speaks. The text exists. The comprehension doesn't.
Is This Actually a Problem?
I want to be honest about the limits of this argument, because the strongest version requires engaging seriously with the counterarguments.
The first counterargument is that AI will improve fast enough to make human oversight unnecessary. If AI reaches the point where it can verify its own work reliably, the expertise problem becomes irrelevant: you don't need human understanding because the machines handle both production and evaluation. This is the implicit bet that every company cutting junior hiring is making. It might be right. But "might" is doing a lot of work in that sentence when the stakes are civilizational infrastructure, and no one is explicitly choosing to make this bet. It's just happening.
And follow the logic to its conclusion. If AI gets good enough to evaluate other AI systems — meaning it can reason about emergent failures, adversarial edge cases, and alignment properties better than humans can — then for that specific function, humans aren't needed. Not because we chose to step aside, but because we became the bottleneck. We'd be the slow, biased, low-bandwidth reviewer insisting on checking the work of something that sees patterns we literally cannot.
Even then, someone still has to decide what counts as safe, acceptable, aligned with human values. That's not an engineering question — it's a political and moral one. We don't let the best doctor decide national health policy unilaterally. There's a version of human oversight that's less about technical competence and more about democratic legitimacy.
But this distinction erodes. If you can't understand the system well enough to evaluate it, you also can't meaningfully set constraints on it. You're choosing between options you can't fully comprehend. It becomes oversight theater — humans rubber-stamping decisions they don't really grasp. And you arrive at this outcome not through anyone's malice, but through a series of individually rational delegations.
This is the mechanism of the one-way door. Each delegation is justified. The person delegating can still, at that moment, understand what they're delegating. But the next person, who grew up in a world where that task was always delegated, cannot. The capacity to take it back quietly disappears.
The second counterargument is that this is fundamentally a software problem. I've spent the last year working at the boundary of software and hardware, building a detection system that involved sensors, circuits, embedded compute, wireless protocols, and mechanical integration. Here is what I can say honestly: physical engineering has structural defenses that software does not.
Hardware has natural verification. A bridge holds weight or it doesn't. A circuit conducts or it doesn't. You can't "vibe code" a physical system the way you can a web application, because physics provides a testing regime that is comprehensive in ways that software testing can never be.
Development cycles in hardware are long (years, not days) which means there's less pressure to ship fast and less opportunity for the kind of rapid AI-assisted iteration that has transformed software. Regulatory frameworks in aerospace, defense, nuclear, and medical devices already mandate demonstrated human understanding at every level. FAA certification doesn't care if AI designed your flight control system; a human engineer has to sign off on it and take personal liability.
These are real protections, and I want to be clear: I am not arguing that physical engineering will follow the same trajectory as software. The structural defenses may hold.
But I am arguing that the software crisis alone is serious enough to warrant attention, and that the question of whether it spreads is worth asking now rather than later. Simulation is increasingly replacing physical testing. AI-assisted CAD and finite element analysis are advancing fast. The economic pressure (iterate faster, reduce headcount, outbid competitors) exists in hardware too, even if it moves more slowly. None of this means the defenses will fail. It means the assumption that they'll hold indefinitely is an assumption, not a guarantee, and one that nobody is stress-testing.
And even if hardware's defenses hold perfectly, the software crisis alone is serious enough. Software runs everything. The financial system, the power grid, the communication networks, the logistics chains — all of it is software all the way down. If the people who understand that software age out and aren't replaced, the hardware protections sit on top of a foundation that nobody comprehends.
The Incentive Trap
The reason I'm pessimistic about solutions is not the technical problem. It's the incentive structure.
No one decided to break the junior developer pipeline. Each company made a rational choice. No one will decide to let AI design safety-critical systems without human understanding. What will happen is gradual creep. AI helps with tedious parts first. Then it handles more of the design. Engineers start reviewing AI output instead of designing from scratch. New engineers enter the field having never designed anything from scratch. The understanding erodes. At no point does anyone cross a bright line, because there is no bright line.
And you cannot unilaterally opt out. The company that uses AI to design faster outbids the one that doesn't. The country whose defense contractors iterate faster fields better systems sooner. The student who uses AI outperforms the one who doesn't, at least on every metric that's currently measured.
This has the structure of a tragedy of the commons. The common resource (our collective capacity to understand our own infrastructure) is being depleted by individually rational decisions. And the natural response to a tragedy of the commons is collective action: regulation, standards, professional licensing. We don't let companies dump chemicals in rivers even though it's individually rational. FAA certification already mandates that a human engineer understands and signs off on flight-critical systems regardless of how they were designed.
But this is a worse version of the problem than the ones we've solved. Environmental regulation works because you can measure the thing being depleted: parts per million, fish populations, tons of carbon. You can set thresholds and enforce them. "Collective human understanding of technical systems" is invisible. You can measure proxies (hiring rates, training hours, hands-on experience requirements) but proxies are gameable. You can mandate junior hiring and those juniors can spend all their time prompting AI. You can require training programs that teach people to review AI output without ever building the mechanistic understanding that makes review meaningful. The distance between the measurable proxy and the actual resource is where the problem lives. You don't know the resource is gone until something breaks and nobody in the room knows why.
We have a problem with the structure of a tragedy of the commons, but without the measurability that makes tragedies of the commons solvable. The resource being depleted is invisible until the moment you need it and it isn't there.
This is the hallway. Each step forward is rational. Each actor is making the right choice. The door behind you closes so quietly you don't hear it.
What I Think Is True
I am a graduate student in an aero department at Stanford. I've watched AI help me cross into circuit design, embedded systems, and wireless protocols that I had no prior experience in, and produce working hardware in weeks instead of years. That experience was exhilarating, and I understand viscerally why the economic incentives push in this direction. I also understand, from direct experience, how quickly "AI is helping me" becomes "I don't fully understand what I built." The gradient is smooth and the slope is steep.
This essay started as a piece about AI and education: here's what should change, here's how to fix the curriculum. I kept getting it wrong because the honest version isn't prescriptive. It's descriptive. Here is what I think is actually true, stated as plainly as I can:
AI is eroding the process by which humans develop technical expertise. This is happening now, measurably, in software engineering. It is likely to spread to other technical domains, though how fast and how far is genuinely uncertain. The incentive structures that drive this erosion are strong enough that no individual actor (no university, company, or government) can reverse them unilaterally. And the thing being lost is invisible enough that we may not have the conceptual tools to protect it even through collective action.
Every previous technology displaced human labor but preserved the ability of humans to understand what they built. The printing press displaced scribes but made knowledge more accessible. The calculator displaced human computers but left mathematical understanding teachable. The compiler displaced assembly programmers but produced deterministic output that could be reasoned about and built upon.
AI is doing something different. It is displacing not just labor but the development of understanding itself: the struggle, the failure, the years of pattern-building that produce the judgment to know when something is wrong. And it is replacing that understanding not with a reliable, verifiable abstraction, but with a system that works until it doesn't, and that no one may be able to diagnose when it fails.
We are running an experiment on humanity's ability to understand its own systems. The experiment was not designed. Nobody voted for it. It is the emergent result of millions of rational decisions. We won't know the outcome for a decade, by which point the generation that could have maintained the understanding will have aged out or never developed it.
There will be no moment when a committee convenes and decides whether to let AI take over evaluation of critical systems. Instead there will be a thousand procurement decisions, each individually sensible, each making the next one more likely, until the aggregate outcome is that no human in the room can meaningfully challenge what the AI is doing. Not because they've been forbidden to. Because they no longer can.
The danger is not that we'll make a bad choice. It's that we never get a clean choice to make. The door closes one degree at a time, each degree the rational response to the last, until you turn around and realize you can't go back. Not because someone locked it. Because the people who knew how to open it are gone.
If you're an engineer in a field that hasn't started this conversation yet (controls, mechanical, aerospace, electrical, civil) I'd encourage you to watch what's happening in CS very carefully. Not because your field is next. But because it might be, and by the time you're sure, the people who would have known how to fix it might not exist.
References
- Federal Reserve Bank of New York, "The Labor Market for Recent College Graduates," based on 2023 American Community Survey data. Unemployment rates are for recent graduates aged 22–27. Note: as discussed by O'Brien (2025) in "A Viral Chart on Recent Graduate Unemployment is Misleading," the sub-sample sizes for specific majors are small and confidence intervals are wide. https://www.newyorkfed.org/research/college-labor-market
- Brynjolfsson, E., Chandar, B., and Chen, R. (2025). "Canaries in the Coal Mine? Six Facts about the Recent Employment Effects of Artificial Intelligence." Stanford Digital Economy Lab. Uses ADP payroll microdata covering millions of U.S. workers. https://digitaleconomy.stanford.edu/publications/canaries-in-the-coal-mine/
- Hosseini, S. M. and Lichtinger, G. (2025). "Generative AI as Seniority-Biased Technological Change: Evidence from U.S. Résumé and Job Posting Data." Harvard University. Analyzed 285,000 U.S. firms and 62 million worker records from 2015 to 2025. https://papers.ssrn.com/sol3/papers.cfm?abstract_id=5425555
- SignalFire (2025). "The State of Tech Talent Report." Analysis based on SignalFire's Beacon AI platform tracking 650+ million individuals and 80+ million organizations. "Big Tech" defined as the top 15 technology companies by market cap. https://www.signalfire.com/blog/signalfire-state-of-talent-report-2025
- Tao, T., as quoted in Shankland, S. (2025). "Math genius Terence Tao says that AI still can't 'smell' bad math." The Decoder. Tao's full observation: "The errors are often really subtle, and then when you spot them, they're really stupid. Like no human would have actually made that mistake." https://the-decoder.com/math-genius-terence-tao-says-that-ai-still-cant-smell-bad-math/
- Anthropic (2025). "How AI is Transforming Work at Anthropic." Internal study surveying 132 engineers and researchers, with 53 in-depth qualitative interviews and analysis of internal Claude Code usage data. https://www.anthropic.com/research/how-ai-is-transforming-work-at-anthropic