There is a question hiding in plain sight behind every celebration of AI-generated code, every prediction that developers are obsolete, every LinkedIn post about building an app 100x faster with a prompt. It is a question that almost nobody in the current hype cycle is asking, and it may be the most important question of all.
Where did the AI learn to write that code?
The answer is straightforward: from decades of human engineering work. From millions of Stack Overflow answers written by developers who spent hours debugging real problems. From open source repositories maintained by people who understood not just syntax, but architecture, tradeoffs, and the messy realities of production systems. From technical documentation, conference talks, blog posts, and code reviews, all produced by human beings who earned their knowledge through years of practice.
Now consider what is happening simultaneously. Stack Overflow’s monthly question volume has collapsed from over 200,000 at its peak in 2014 to fewer than 4,000 by December 2025, a 78% drop from the previous year alone. Junior developer hiring has plummeted by 67% since 2022. A Harvard study tracking 62 million workers across 285,000 U.S. firms found that junior employment drops 9 to 10 percent within six quarters of companies adopting generative AI, while senior employment barely changes. Computer science graduates now face 6.1% unemployment, nearly double the national average.
Meanwhile, over 74% of newly published web pages contain detectable AI-generated content. Researchers estimate that more than half of all new English-language articles online are now synthetic. A landmark paper published in Nature demonstrated that AI models trained on their own output undergo “model collapse,” a degenerative process where they progressively forget rare but important patterns in the original data.
This is the training data paradox: the industry is simultaneously consuming the knowledge produced by human engineers while eliminating the conditions that produced that knowledge in the first place.
The Ecosystem Nobody Talks About
To understand why this matters, you need to see software engineering knowledge as an ecosystem rather than a static resource.
When a senior engineer writes a Stack Overflow answer explaining how to handle database transactions under concurrent load, that answer does not appear from nothing. It emerges from years of experience building systems, encountering failures, debugging production incidents, and learning from other engineers who came before them. The answer is the visible tip of an enormous iceberg of tacit knowledge.
That ecosystem had a lifecycle. Junior developers asked questions. Mid-level developers answered them and learned to articulate their knowledge in the process. Senior developers refined answers, corrected misconceptions, and contributed nuanced perspectives that only come from seeing systems fail in unexpected ways. The community collectively maintained and updated this knowledge as technologies evolved.
AI models consumed the output of this ecosystem at scale. Every training run ingested millions of these human knowledge artifacts. The models learned to produce code that looks correct because it is statistically modeled on code that was actually correct, written by people who understood why it was correct.
But the ecosystem itself is now collapsing. Stack Overflow, which was the largest publicly accessible repository of human-verified programming knowledge, has returned to question volumes not seen since its launch year in 2008. The 2025 Stack Overflow Developer Survey revealed that while 84% of developers now use AI tools, positive sentiment toward those tools has dropped from over 70% in 2023 to just 60%. Trust remains remarkably low, with only 3.1% of developers expressing high confidence in AI output. 87% report concerns about accuracy.
Developers are using AI more and trusting it less. That disconnect should concern everyone.
The Model Collapse Problem
The scientific evidence for what happens when AI trains on AI-generated content is now substantial and alarming.
In 2024, a team of researchers from British and Canadian universities published a widely cited paper in Nature demonstrating model collapse in large language models. When successive generations of models were trained on data produced by their predecessors, the models progressively lost information about rare events and edge cases, eventually converging on a narrow, high-probability output that looked increasingly nothing like the original training distribution.
The process has two stages. In early model collapse, the model loses information from the tails of the distribution, the unusual cases, the edge conditions, the rare but important patterns. This stage is particularly insidious because overall performance may appear to improve even as the model loses its ability to handle minority cases. In late model collapse, the distribution converges so dramatically that the output becomes nearly useless.
A 2025 ICLR spotlight paper titled “Strong Model Collapse” went further, showing that even the smallest fraction of synthetic data in a training corpus, as little as one in a thousand samples, can still lead to model collapse. Larger training sets do not help. And critically, larger models can actually amplify the collapse rather than mitigate it.
Think about what this means in practical terms. The Harvard Journal of Law and Technology has described the situation as an “Ouroboros,” a snake eating its own tail. AI companies scraped the web to train their models. Those models now generate content that floods the web. Future models will be trained on a web that is increasingly composed of AI-generated content. Each generation learns from the degraded output of the last.
The researchers have identified three compounding errors in this process: statistical approximation errors from finite sampling, functional expressivity errors from the limited capacity of models, and functional approximation errors from the biases in learning procedures. These errors accumulate across generations. They do not self-correct.
The Junior Developer Pipeline Crisis
The training data paradox does not only operate at the level of web content and model training. It operates at the level of human knowledge development itself.
Junior developer hiring has collapsed across the industry. Entry-level developer opportunities have dropped approximately 67% since 2022. In the UK, tech graduate roles fell 46% in 2024, with projections for a further 53% drop by 2026. A Stanford Digital Economy Study found that employment for software developers aged 22 to 25 declined nearly 20% from its peak in late 2022 by July 2025. In the United States, overall programmer employment fell 27.5% between 2023 and 2025 according to the Bureau of Labor Statistics.
The logic behind these cuts seems rational from a quarterly earnings perspective. If an AI coding assistant can handle boilerplate work, why hire juniors to do it? If a senior developer augmented by AI can be 20% more productive, why not let them absorb the work that would have gone to two juniors?
But this logic contains a fatal assumption: that the supply of senior engineers is somehow independent of the supply of juniors.
It is not. Every senior engineer was once a junior. Every architect once struggled with their first database schema. Every tech lead once submitted a pull request that was mostly wrong. The junior developer pipeline is not overhead. It is the mechanism through which the industry reproduces its own expertise.
When companies stop hiring juniors, they are not just cutting costs. They are cutting off the pipeline that produces the senior engineers they will desperately need in five to ten years. A 67% hiring cliff in 2024 to 2026 means 67% fewer potential tech leads in 2031 to 2036. The industry is eating its seed corn.
AWS CEO Matt Garman called the idea of replacing junior developers with AI “one of the dumbest things I’ve ever heard.” Microsoft’s Mark Russinovich and Scott Hanselman have publicly argued that companies must hire junior developers and teach them to fix the mistakes made by AI. But the industry trend moves in the opposite direction.
The Double Helix of Degradation
What makes the training data paradox particularly dangerous is that these two degradation loops reinforce each other.
The first loop is technical: AI models trained on increasingly synthetic data produce increasingly degraded output. The web fills with AI-generated content. Future models train on that content. Quality degrades further.
The second loop is human: fewer junior developers enter the profession. Fewer humans produce the kind of hard-won, experience-driven knowledge that made the training data valuable in the first place. The knowledge ecosystem that produced the original high-quality training data withers.
These two loops feed into each other. As AI output quality degrades due to model collapse, the need for skilled human engineers who can catch errors and provide genuine expertise increases. But those humans are not being developed because the industry has decided AI makes them unnecessary. And the fewer humans contributing genuine knowledge, the faster the synthetic content ratio climbs, accelerating model collapse.
It is a vicious cycle with a time delay built in. The consequences of decisions being made today will not become fully visible for five to ten years. By then, the damage to the knowledge ecosystem may be difficult to reverse.
What the Numbers Actually Say
Let me lay out the data points that together paint this picture.
On the synthetic content side: an Ahrefs analysis of 900,000 newly published web pages in April 2025 found that 74.2% contained detectable AI-generated content. A Graphite report analyzing 65,000 English-language articles found that AI-generated content has reached a roughly 50/50 split with human-written content. AI-written pages appearing in the top 20 Google search results climbed from 11.1% to 19.6% between May 2024 and July 2025. NewsGuard tracked AI “news” sites growing from 49 to 1,271 between May 2023 and May 2025.
On the knowledge ecosystem side: Stack Overflow monthly questions fell from 200,000 to under 4,000. The 2025 Developer Survey shows 84% of developers using AI tools but only 3.1% expressing high trust in the output. 76% of developers say they will not use AI for deployment and monitoring. 69% say they will not use it for project planning. The tools are being used for the easy parts. The hard parts still require human judgment that fewer humans are being trained to develop.
On the hiring side: a Harvard study of 285,000 firms shows 9 to 10% junior employment decline per AI adoption wave. U.S. programmer employment fell 27.5% between 2023 and 2025. CS graduate unemployment sits at 6.1%. Computer engineering graduate unemployment is at 7.5%, higher than fine arts graduates. 54% of engineering leaders surveyed by LeadDev plan to hire fewer juniors due to AI.
On the productivity illusion side: businesses self-report a 24.7% productivity increase from AI adoption. But actual measured results from studies of 39,000 developers show only a 2.1% overall productivity increase and a 3.4% code quality improvement. Software delivery performance actually declined 7.2% in some studies. The gap between perceived and actual productivity gains is roughly 12x.
These numbers tell a consistent story. The industry is making dramatic structural changes based on dramatically overestimated productivity gains, while the actual foundations of software quality are being eroded.
The Photocopier Analogy
The Harvard Journal of Law and Technology used an analogy that I think captures the situation perfectly: model collapse is like repeatedly photocopying a picture. Each copy is slightly degraded from the last. After enough generations, the image becomes unrecognizable.
But the analogy needs to be extended. Imagine that the original photographs were taken by skilled photographers who spent decades learning their craft. Now imagine that, because the photocopier exists, you decide to stop training new photographers. After all, the machine can produce images. Why invest in expensive human expertise?
For a while, the photocopies look fine. Close enough, at least. But each generation loses a little more detail. And because no new photographers are being trained, there are no new original photographs being taken to refresh the source material. The machine copies copies of copies, each time losing something. And the humans who could have created genuinely new, high-quality originals were never given the chance to learn.
This is where the software industry is heading. Not with a dramatic crash, but with a slow, almost imperceptible degradation. Code that works but handles fewer edge cases. Systems that function but fail in increasingly surprising ways. Documentation that reads well but contains subtle inaccuracies that compound over time. An industry that gradually loses the deep expertise that made its products reliable in the first place.
The Pre-2022 Data Advantage
There is an irony here that deserves attention. AI companies are increasingly aware that data collected before the generative AI explosion of 2022 is qualitatively different from what is available now. Pre-2022 web data was predominantly human-authored. Post-2022 data is increasingly synthetic. The contamination is difficult to detect and nearly impossible to fully remove.
This creates what researchers have called a “data moat” for incumbent AI companies. Organizations that scraped and stored large datasets before 2022 possess something that future competitors cannot easily replicate: a corpus of genuinely human-generated knowledge. New entrants to the AI market face an increasingly polluted web that makes high-quality training data harder to find.
The implications extend beyond corporate competition. If the highest-quality AI models depend on pre-2022 human knowledge, and the industry is simultaneously eliminating the conditions that produced that knowledge, then we are living off a finite inheritance. We are spending the intellectual capital accumulated by generations of engineers while failing to invest in creating more.
Some AI companies are attempting to address this through data licensing deals. Stack Overflow, despite its traffic collapse, achieved 17% revenue growth by licensing its data to AI providers. Reddit has similarly monetized its archive. But these deals do not solve the fundamental problem. The data being licensed is historical. It represents what humans knew and discussed in the past. Without ongoing human knowledge creation, these archives become increasingly stale.
What Actually Sustains Software Quality
The training data paradox reveals something important about the nature of software engineering knowledge that the “developers are dead” crowd consistently misses.
Software quality does not come from code generation. It comes from a complex ecosystem of human activities: understanding business requirements deeply enough to know which edge cases matter, making architectural decisions that balance competing concerns, debugging failures that only appear under specific production conditions, reviewing code with an understanding of how systems evolve over time, mentoring junior engineers who will carry institutional knowledge forward.
None of these activities produce the kind of clean, structured data that AI models train on effectively. They happen in conversations, in whiteboard sessions, in the accumulated judgment of experienced engineers who have seen enough systems fail to recognize warning signs early. This knowledge is largely invisible to AI training pipelines.
The 2025 Stack Overflow Developer Survey confirmed this. The most common activity when developers visit Stack Overflow is reading comments, not reading answers. The comments are where the nuance lives. Where experienced developers say “this works, but be careful in multi-threaded environments” or “this solution is correct for PostgreSQL but will fail silently on MySQL.” That contextual, conditional, experience-driven knowledge is exactly what AI models struggle to capture and exactly what model collapse erodes first.
This is the tail of the distribution that the Nature paper warned about. Edge cases, rare conditions, platform-specific quirks, the kind of knowledge that only comes from years of production experience. Model collapse erases these tails first. They are the minority data. They are also the data that prevents systems from failing in production.
The Uncomfortable Economics
The economic incentives driving the training data paradox are powerful and, in the short term, entirely rational.
For individual companies, cutting junior headcount and relying on AI-augmented seniors saves money this quarter. Licensing AI tools is cheaper than hiring. The productivity reports, even if inflated, provide cover for cost-cutting decisions that executives were already inclined to make.
For AI companies, the incentive is to acquire and license as much human-generated data as possible while the getting is good. The value of their models depends on training data quality. They know this. But they also know that the market rewards capability today, not sustainability tomorrow.
For individual developers, the incentive to contribute knowledge to public platforms is declining. Why spend an hour writing a detailed Stack Overflow answer when the platform’s traffic has collapsed, the answer will be scraped into an AI training set, and the AI will receive the credit? The rational response is to stop contributing. And that is exactly what is happening.
This is a classic tragedy of the commons. The shared resource, the global pool of human software engineering knowledge, is being depleted because no individual actor bears the full cost of its loss. Everyone benefits from the knowledge. No one has sufficient incentive to sustain it.
What Can Be Done
I do not have a neat solution to the training data paradox. But I can identify what needs to happen for the industry to avoid the worst outcomes.
Companies need to maintain junior developer pipelines. Not out of charity, but out of long-term self-interest. The senior engineers of 2035 are the juniors who should be getting hired today. Companies that cut this pipeline entirely are making a bet that AI will replace human judgment within a decade. If that bet is wrong, and six decades of technology history suggest it will be, they will face a talent crisis they cannot quickly fix.
The industry needs to value and protect high-quality human-generated knowledge. This means creating economic models that compensate knowledge creators rather than simply scraping their output. Stack Overflow’s data licensing deals are a step in this direction, but the compensation needs to flow to the individual contributors, not just the platform.
AI companies need to invest seriously in data provenance and synthetic data detection. Training on contaminated data is not just an abstract research concern. It is a product quality issue. Models that progressively degrade because of synthetic data contamination will lose market position to competitors that maintain data quality. The ICLR research showing that even one-in-a-thousand synthetic samples can trigger collapse should be taken seriously.
Individual engineers need to continue investing in deep understanding. Not because it is noble, but because it is practical. As AI-generated code quality degrades over time, the ability to evaluate, debug, and correct that code becomes more valuable, not less. The developers who understand systems, tradeoffs, and architecture will be the ones who keep things running when the AI-generated layers start to show their seams.
And all of us, as an industry, need to be honest about what is happening. The narrative that AI will simply replace developers is not just wrong in the historical sense I explored in my previous piece, “The Eternal Promise.” It is actively harmful because it accelerates the destruction of the very knowledge ecosystem that AI depends on.
The Paradox Restated
The training data paradox is this: AI learned from the best of what human engineers produced over decades. The industry is now using AI as justification to stop producing the kind of human engineering knowledge that made AI valuable in the first place. If this continues unchecked, the models will degrade, the talent pipeline will dry up, and the industry will face a compounding crisis of declining AI quality and declining human expertise simultaneously.
Software is crystallized thought. AI models are crystallized from the crystallized thoughts of millions of developers. If we stop training the developers who produce those thoughts, we are melting the source material.
The question is not whether AI can write code. It can. The question is whether the code it writes will get better or worse over time. And the answer to that question depends entirely on whether we maintain the human knowledge ecosystem that AI was built on.
History does not repeat itself exactly. But it does have patterns. And the pattern here, the one where an industry consumes a resource faster than it can be replenished, is one of the oldest stories in human civilization.
The engineers who recognize this paradox and act accordingly will be the ones who remain indispensable. Not because they resist AI, but because they understand what AI cannot sustain on its own.
Final Thoughts
If this analysis resonated, you might also want to read my earlier piece, The Eternal Promise: A History of Attempts to Eliminate Programmers, which recently sparked a 160+ comment discussion on Hacker News. Together, these two posts explore both the historical pattern and the emerging systemic risk.
I write regularly about the intersection of software engineering, AI, and the realities of building systems in production. You can find me on LinkedIn for shorter insights and discussions, or follow this blog for longer explorations.
I would genuinely like to hear your perspective on this. Are you seeing the knowledge ecosystem decline in your own work? Have you noticed AI output quality changing over time? Are companies in your network cutting junior hiring and, if so, are they thinking about the long-term consequences?
Leave a comment below, connect with me on social media, or reach out through the contact page. The conversations that follow these posts are often as valuable as the posts themselves.
Until next time, keep building wisely.