The collapse of AI software engineering is more complicated than it looks

AI coding tools were supposed to be the ultimate cheat code for software engineering: faster output, lower costs, and leaner teams. Instead, they’ve created a mess of buggy code, burned-out seniors, and a disappearing path for junior developers—while doubling as a convenient story to hide Big Tech’s mounting debt problem.

AI code is faster—but not better

On paper, AI-assisted coding looks incredible. Many teams report that AI can generate code 4–10x faster than traditional development. In theory, that should mean more features, more products, and more profit.

But speed at the point of generation doesn’t equal real-world efficiency. The real question is: how much of that code survives, and at what cost?

The rise of code churn and fragile software

One of the clearest warning signs is code churn—the rate at which code is rewritten, heavily modified, or thrown away. Ideally, churn stays low, because every rewrite is time and money spent fixing old work instead of shipping new value.

According to GitClear’s analysis of 211 million lines of code, churn has more than doubled since teams started adopting AI coding tools. In 2021, churn sat around 3.3%—about 33 lines out of every 1,000 eventually needed to be rewritten or discarded. By 2025, that number had jumped to 6.87%.

That means more time spent refactoring, debugging, and stabilizing AI-generated code. The promised efficiency gains at the front of the pipeline are being eaten alive further downstream.

Security risks baked into AI-generated code

The quality problem isn’t just about bugs—it’s about security. Veracode’s analysis of 100–150 large language models (LLMs) found that 45% of AI code-generation tasks introduced at least one known security vulnerability.

In some languages, it’s even worse. Java-based code generated by AI fails security checks more than 70% of the time. Python does better, but still shows vulnerability rates around 38%. That means nearly half of AI-written code is risky out of the box, before it even hits production.

Security flaws at this scale don’t just slow teams down. They raise the risk of breaches, compliance failures, legal exposure, and expensive firefighting after release.

AI speed isn’t turning into shipped software

Even if the code is messy, you might hope that at least more features are shipping. But the data doesn’t support that either.

A 2026 study by research firm DX looked at more than 400 companies. Between November 2024 and February 2026, these companies increased their use of AI tools by 65%, expecting a big jump in shipped software.

The median pull request (PR) throughput per developer—the number of completed PRs—rose by just 7.76%. In other words, AI helped developers type faster, but it didn’t meaningfully increase how much production-ready code actually made it over the finish line.

When AI makes software more expensive

AI coding tools were marketed as a way to cut engineering costs. Instead, they’re turning into cost multipliers.

A 2026 industry analysis projects that the total cost of ownership (TCO) for AI-assisted software will be 30–40% higher by 2027 than for traditional development. GitClear’s data suggests that by year two of using AI for coding—if teams don’t manage it carefully—maintenance costs can climb to four times previous levels.

Why? Because fixing AI-generated bugs is hard, slow, and expensive. Once low-quality code is merged into a large codebase, every change ripples through multiple systems. If it slips into production, you add incident response, security fallout, and regulatory risk on top.

In the pre-AI era, the bottleneck was often writing the initial code. Now, AI has simply moved the bottleneck: review queues are clogged, security checks take longer, and senior engineers are forced to spend more time validating that AI-generated code actually works.

Whatever savings appear at the typing stage are burned in review, debugging, and remediation.

AI didn’t replace developers—it changed what they do

Despite the hype, AI hasn’t eliminated the need for developers. It’s changed the shape of the work and where the demand is.

Instead of replacing humans, AI has created a new kind of demand: experienced engineers who can understand complex systems, spot subtle bugs, and clean up messy code. The problem is that the industry tried to get there by cutting out the bottom of the ladder.

The hollowing out of junior developers

From 2022 to 2025, tech companies laid off tens of thousands of developers. But those cuts were not evenly distributed. Juniors—who usually handle boilerplate tasks, testing, and initial bug-fixing—were hit hardest.

Faced with a choice between paying an $80,000 junior salary or a $100/month AI subscription, many companies chose AI. Data from SignalFire and ZipRecruiter shows that entry-level tech postings dropped by roughly 25–35% between 2023 and 2024. Big Tech cut new-grad hiring by about 50% compared to pre-pandemic levels, and by 2025–2026, entry-level postings were still down around 29%.

At the same time, the definition of “entry level” quietly shifted. Roles labeled as entry-level now often require 2–5 years of experience. For true beginners, the job market is roughly a third smaller than it was in 2022.

If you want a deeper dive into how this dynamic is affecting the craft of engineering itself, it’s worth reading why AI won’t replace real software engineering.

From pyramid teams to diamond teams

Before AI, most engineering orgs looked like a pyramid: lots of juniors at the base, fewer mid-levels, and a smaller group of seniors at the top. Juniors learned by doing, gradually taking on more complex work as they gained experience.

Now, the structure is more like a diamond. There are very few juniors, a swollen middle of mid-level engineers, and a rapidly growing demand for seniors. By early 2026, senior roles made up over 41% of all IT job postings, with roles requiring 5–10 years of experience growing by 40% and 10+ year roles by 27% in 2024 alone.

On the surface, this looks like a hiring recovery. In reality, it’s a “senior boom” hiding a “junior bust.”

The new job of the senior engineer: janitor-in-chief

In the old model, juniors handled much of the grunt work: writing boilerplate, running tests, and cleaning up obvious issues before code reached senior reviewers. That meant seniors could focus on architecture, design decisions, and the thorniest problems.

Today, with roughly 45% of AI-generated code containing known security issues and code churn nearly doubled, the job has flipped. Senior engineers are spending more and more of their time reviewing, rewriting, and securing AI-written code.

They’re still highly paid and highly skilled—but much of their day now looks like janitorial work. Unsurprisingly, burnout is rising, and some teams are grinding to a halt under the weight of review and maintenance. Stories like why one senior engineer quit GitHub over AI coding agents are becoming less rare and more like a preview of where the industry is heading.

The “junior death spiral”

By slashing entry-level roles, companies have created a long-term structural problem: they’ve broken the talent pipeline.

Traditionally, juniors gain experience by fixing bugs, writing tests, and gradually taking ownership of more complex systems. Over 3–7 years, they grow into mid-level and then senior engineers.

Without that entry point, the industry is heading toward a “junior death spiral.” In 2–3 years, we could see a severe shortage of mid-level engineers because there simply weren’t enough juniors entering and progressing through the system.

That raises an obvious question: if there are no juniors gaining experience today, where will tomorrow’s seniors come from?

Not every company has made this mistake. Organizations that chose to augment juniors with AI—using tools as force multipliers rather than replacements—are already seeing better results. For companies that hollowed out their teams, though, AI is acting less like leverage and more like a force divider.

Behind the layoffs: AI hype and hidden debt

So if AI isn’t delivering the promised efficiency, why are companies still leaning so hard into it—and cutting so many humans in the process?

The answer has less to do with productivity and more to do with money. Specifically: cheap debt, rising interest rates, and massive AI bets that haven’t yet paid off.

During the pre-pandemic “cheap money” era, interest rates were near zero. Big Tech used that environment to borrow aggressively—both to fund growth and to finance huge stock buybacks that pushed share prices higher.

When lockdowns turbocharged digital demand, companies like Microsoft and Amazon doubled down, assuming that behavior would stick. They went on hiring sprees and started building out enormous AI infrastructure—much of it financed with more debt.

Then the pandemic ended, inflation hit, and central banks responded by hiking rates from near 0% to around 5%. That “cheap” debt suddenly got expensive.

$1.65 trillion in hidden obligations

The real shock isn’t just the amount of debt, but how much of it is effectively hidden. The “Big Five” tech giants are sitting on about $1.65 trillion in off-balance-sheet liabilities—commitments that don’t show up clearly as traditional debt.

Meta alone holds around $420 billion in such obligations, nearly three times its reported debt. Oracle’s off-balance-sheet commitments have exploded 30x to roughly $273 billion. Alphabet, Microsoft, and Amazon account for the rest.

Most of these liabilities are long-term leases for data centers, hardware purchase commitments, and similar infrastructure deals. They’re not optional. They’re fixed payments that have to be made, regardless of how much AI revenue actually shows up.

By keeping these obligations off the main balance sheet, companies make their leverage look lower than it really is. That can mislead investors and credit agencies, inflating valuations and underestimating risk—an echo of the creative accounting that helped bring down Enron.

Why “AI efficiency” is the perfect cover story

Given this backdrop—rising rates, massive hidden obligations, and AI investments that haven’t yet delivered—it’s not surprising that Big Tech is desperate to cut costs.

They need to trim expenses to service their commitments and keep investors happy, all while continuing to pour money into AI infrastructure. That’s a tough sell if you admit you simply over-hired or misjudged demand.

But if you say you’re cutting jobs because of “AI efficiency”? The market reacts very differently.

When Salesforce cut 8,000 jobs in 2023 and cited “over-hiring” and macroeconomic weakness, its stock fell by about 32% over the following months. When Nike laid off around 800 people to “enhance automation” amid inventory and sales issues, its share price dropped nearly 35%.

Compare that to Meta. In 2023, it framed massive layoffs as part of a “Year of Efficiency” and a pivot to AI. The stock surged more than 190% over the next year. When Alphabet cut 12,000 jobs in early 2023 while highlighting AI integration, its stock jumped about 5% in a single day.

A 2026 study of S&P 500 companies found that more than half of firms announcing AI-linked layoffs saw their stock prices rise, while generic cost-cutting often led to declines. The message is clear: if you say you’re cutting staff to redirect capital into high-margin AI, investors hear “future profits.”

AI washing: when everything is “because of AI”

This incentive structure has led to a new form of corporate spin: “AI washing.” Any restructuring, austerity measure, or layoff can be reframed as an AI transformation.

Outplacement firm Challenger, Gray & Christmas reported that by May 2026, AI was the most cited reason for job cuts—a category that barely existed two years earlier. A Deutsche Bank analysis in July found that mentions of “AI disruption” on earnings calls jumped 310% in the first half of 2026 compared to late 2025.

When Amazon CEO Andy Jassy said in early 2026 that AI would mean “fewer people doing some of the jobs that are being done today,” the statement coincided with a 16,000-person cut. Meta laid off 8,000 workers and shifted another 7,000 into AI roles, again framing it as a necessary move for an AI-first future.

In many cases, “AI efficiency” is less about real productivity and more about giving Wall Street a story it wants to hear.

Three possible futures for AI and software engineering

Where does this all go from here? Based on data from Goldman Sachs, MIT, and the Bank for International Settlements, there are three broad scenarios.

1. The optimistic scenario (around 15% probability)

In the best case, the current mess is just the “installation phase” of a new technological revolution. By late 2027, AI models could achieve more reliable agentic reasoning—moving beyond code snippets to designing architectures, coordinating systems, and even self-healing their own mistakes.

If that happens, today’s buggy code and churn problems might largely disappear. Productivity could finally spike, global GDP could see a major boost, and that $1.65 trillion in hidden obligations would look like a smart early bet. Displaced workers might move into high-value “AI orchestration” roles, managing fleets of agents instead of writing every line by hand.

2. The crash scenario (around 25% probability)

At the other extreme, the AI boom could end like the dot-com bubble—only bigger. A few high-profile bankruptcies triggered by off-balance-sheet liabilities could spark a credit crunch, popping the AI bubble and wiping out trillions in paper wealth.

Data centers could be abandoned, long-term energy contracts might become toxic assets, and the “junior death spiral” could harden into a permanent structural decline. With no entry-level roles and seniors burned out maintaining brittle AI-enabled systems, the tech sector could shrink by as much as a third, dragging the global economy into a deep recession.

3. The messy middle (most likely)

The most probable outcome is somewhere in between: not a spectacular crash, but a long, grinding deflation of AI hype.

Over time, markets will stop rewarding companies just for saying “AI” and start demanding clear, verifiable earnings from AI products. A minority of companies—maybe 5–12%—that used AI to augment humans instead of replace them will start to show strong, durable returns.

The $1.65 trillion in hidden liabilities will keep being serviced, but will act as a drag on growth. Big Tech will face years of austerity to pay down commitments while trying to maintain innovation.

Meanwhile, software development will settle into a more or less permanent “diamond” structure: small teams of elite engineers supported by AI tools, a shrinking pool of mid-levels, and very few true entry-level positions. The industry will be left with higher structural costs, a broken talent pipeline, and a debt hangover that could take a decade or more to clear.

What this means if you’re a developer

For developers—especially juniors or those trying to break into the industry—this shift is more than an abstract macro story. It directly affects your career path and risk exposure.

If you’re early in your career and your company’s leadership starts talking loudly about “AI efficiency,” it may be a sign that cuts are coming, not that your workload is about to get easier. In many organizations, AI is being used to justify headcount reductions, even when it’s quietly increasing the work required from the people who remain.

On the other hand, companies that use AI to support and accelerate human developers—not replace them—are more likely to build sustainable teams and long-term value. They invest in juniors, treat AI as a tool rather than a magic solution, and design processes that control churn instead of amplifying it.

The real meaning of “AI efficiency”

“AI efficiency” isn’t just a buzzword. It’s a signal that we’re in the middle of a massive economic and structural shift in how software is built—and who gets to build it.

Behind the scenes, AI coding tools are increasing code churn, raising security risks, and pushing up long-term costs. At the same time, Big Tech is using AI as a convenient narrative to justify layoffs and mask the pressure of enormous, often hidden, AI infrastructure bets.

The technology will keep improving. Some teams will figure out how to use it well. But for now, if you’re a junior developer and your leadership starts praising AI as the key to efficiency, it may be less a promise of a smarter future—and more a warning that it’s time to start planning your next move.