I tried, across a number of pieces published on this platform, to use a large language model to write in my voice and hold an investigative position. I used the best available tool. I tried repeatedly, with genuine effort, over an extended period. The result was the same every time. No matter how many billions of documents that model was trained on before I found it, it could not reliably qualify its answers with appropriate determinism. It hallucinated information and presented it as fact. It drifted from its investigative position as arguments ran long. It produced fluent, confident output that was, on inspection, sometimes covering for nothing at all.
After edits, the total time required to get from prompt to publish exceeded what it would have taken me to write the piece myself. And I am not a professional software developer. I am an engineer who writes. For a junior developer whose entire output is code, the failure modes are identical and the stakes are considerably higher.
The companies that fired their junior developers in the name of AI efficiency did not run this experiment. They accepted the premise without testing it. According to Forrester Research, 55% of them now report regretting it. The other 45% have not admitted it yet.
Consider what cannibalism actually does to a body. You still have your brain. You still have your judgment. But you have cut off your arms and legs, and after that you have no ability to feed yourself. The food comes from without, not within. You live off the energy stored in those limbs for a while. Then you die. The technology industry shed its execution layer and called it becoming leaner. Lean means the tissue that remains is functional and self-sustaining. What happened here was amputation. The judgment layer cannot feed itself without the execution layer beneath it. Institutional knowledge runs down. Seniors leave and cannot be replaced because the pipeline that produced them is severed. A severed pipeline does not recover on its own.
The structural failure modes of large language models are not bugs awaiting a patch. They are architectural features of how these systems work. They hallucinate with confidence. They produce output that is fluent at the surface and wrong underneath. They cannot reliably signal their own uncertainty. They drift from a sustained analytical position when the argument runs long. They do not know what they do not know, and they do not flag when they are operating outside their competence.
These are not reasons to avoid using AI tools. They are reasons to have someone in the room who can catch those failures before they ship. That person needs two things: domain knowledge sufficient to recognize when the output is wrong, and accumulated experience with the tool's specific failure modes to know where to look.
The junior developer sitting with AI output every day is building exactly that competence. Not through instruction. Through being burned. Through tracing a confident error back to its source. Through learning which categories of problem the tool handles poorly and which kinds of fluency are genuine versus decorative. That is a skill that did not exist before these tools existed. It is being built right now in junior seats across the industry. It is being discarded.
Critical evaluation of machine output is not the same skill as critical evaluation of human output. It requires different intuitions, built through different experience. The junior who has developed those intuitions is not replaceable by the senior who never needed them.
This is not a hypothetical. The senior developer who learned to evaluate human work, who built intuitions about where human colleagues cut corners or made assumptions, is not automatically equipped to evaluate machine work. The failure signatures are different. The tells are different. The senior has more domain knowledge but may have less calibrated skepticism toward output that looks exactly like competent work and isn't.
If the stated rationale for the layoffs was cost reduction enabled by AI automation, the economic logic points in one direction. Senior developers are compensated at multiples of junior salaries. The tasks most directly automated by current AI tools are execution tasks: writing boilerplate, generating test cases, producing first-draft code from specification. Those are tasks that occupy significant junior time.
They are also tasks that seniors were doing when juniors were unavailable. A senior who can now generate a first draft with AI has, in a narrow sense, become more productive at execution work. But that productivity gain does not justify the senior's compensation premium over a junior who could review and correct that same AI output with fresh eyes and lower salary expectations.
The math of the layoffs never added up on pure efficiency grounds. Senior developers are not cheaper than junior developers with AI assistance. They are more expensive, less cognitively flexible by most measures, and more institutionally protected. The decision to cut juniors rather than seniors was not an economic optimization. It was a political outcome dressed as one.
Seniors had the organizational standing to protect themselves and frame the narrative. Nobody asked whether AI could do what seniors do, because seniors were doing the asking. The junior layoffs were the answer to a question that was never put to the people most at risk of an honest answer.
Adapting to AI as a senior means confronting something uncomfortable: some of what you spent a decade learning is now partially automated. Your value proposition has to shift from execution to judgment, from output to evaluation, from knowing how to write the code to knowing when the code is wrong. That is a genuine identity threat. Most organizations gave seniors a way to avoid confronting it by pointing at juniors instead. The seniors took it.
Klarna replaced 700 customer-facing employees with AI in 2024, announcing the move as a productivity breakthrough. Quality declined. Customers revolted. The company quietly began rehiring humans. The AI did not fail dramatically. It failed at the margins, in the judgment calls, in the cases that required understanding what a customer actually meant rather than what they literally said. Those margins were invisible until they weren't.
Klarna is not an outlier. Forrester predicts half of all AI-attributed layoffs will be quietly rehired, but offshore or at significantly lower salaries. The pattern is consistent: lay off workers for AI capabilities that do not yet exist at the required reliability level, discover the gap when quality degrades, fill it with cheaper labor rather than admit the original decision was wrong. The institutional memory of the error is processed into a lesson about implementation speed rather than a lesson about what you owe to people whose lives you treated as rounding errors in a productivity calculation.
Entry-level tech hiring decreased 25% year over year in 2024. Employment for software developers aged 22 to 25 declined nearly 20% from its peak in late 2022. Entry-level hiring at major tech companies dropped more than 50% over three years. These are not contested figures. They are the documented, quiet collapse of the on-ramp into the technology industry for an entire generation.
Most of this did not happen through layoffs. It happened through hiring freezes. A layoff is a moment. You can point to it. You can be angry about it. You can tell the story. The hiring freeze means the phone never rings and nobody owes you an explanation. An entire cohort of people who did everything right, who acquired the skills the industry was explicitly recruiting for, simply found the door had been closed before they arrived.
Those people are not waiting. A mathematical mind can be retrained. Pattern recognition, systematic thinking, comfort with ambiguity resolved through logic, tolerance for debugging a problem until it breaks open: these are not software skills. They are cognitive skills that software development selected for. They transfer.
Nursing is the most pointed example because the shortage is acute and the public image of the profession obscures how much of it is clinical, procedural, and analytical. ICU nursing, surgical nursing, radiology, anesthesia support, health informatics, medical coding. None of those roles primarily require bedside manner. The nurse monitoring a ventilated patient in a neurological ICU is reading data streams, interpreting trends, catching deviations from baseline, and escalating appropriately. That is debugging with higher stakes. The cognitive profile is identical. Healthcare is desperately short of people with exactly that profile, and the technology industry just trained and then expelled a generation of them.
If 500 developers retrain as nurses, the nation's hospitals have 1,000 jobs waiting for them. The technology industry will have 500 fewer people who know how to catch what the AI gets wrong. Both of those things are true simultaneously.
The window to recover this talent is closing. The people who left are reskilling. Some will not come back at any price, having concluded, reasonably, that an industry willing to treat an entire generation as expendable during a transition it mismanaged is not one that has earned their loyalty. The companies now quietly rehiring are discovering that the market has moved. The talent they need is either gone, employed elsewhere, or available only at rates that make the original cost reduction calculation look worse in retrospect than it did at the time.
The companies that cut junior developers in the name of AI efficiency made a political decision and called it a strategic one. The organizations that recognize that distinction now have a narrow window to act before the talent pool permanently reallocates to fields that treated it better.
PolicyTorque recommends the following:
The technology industry told a generation of analytical thinkers that their skills were obsolete before testing whether that was true. It is now discovering, in real time, that it was not true and that the cost of the error is compounding. The talent that left to become nurses, paralegals, data analysts, and clinical coders is not sitting idle waiting to be called back. It is building expertise in fields that needed it more and treated it better.
You can't start eating your own limbs and call it a weight loss plan. At some point you have to account for what you consumed.