AI Is Eating the Engineering Career Ladder — Not Just Junior Jobs

The debate about AI and engineering jobs has focused on the wrong question. The argument over whether AI will eliminate developer roles misses what is already happening: AI is removing the work that teaches people to become developers, erasing the rungs of the career ladder before new engineers can climb them.

Alasdair Allan's analysis, drawing on recent research from Anthropic, METR, and labour market data, makes the structural problem visible. The concern is not mass unemployment. It is the slow collapse of the skill formation pipeline — a failure mode whose consequences are delayed enough to be easy to ignore until they aren't.

The Learning Work Is Disappearing First

Software engineering has always had a curriculum hidden inside the work itself. Debugging taught how systems fail. Code review transmitted standards and judgment from experienced engineers to juniors. Documentation forced the discipline of clear reasoning. These weren't just tasks — they were the mechanism through which tacit knowledge transferred across generations of engineers.

AI tools are automating exactly this category of work first. Not because it's strategically optimal, but because it's the work that most easily yields to pattern-matching: routine debugging, boilerplate code, first-draft documentation. The result is that junior engineers are being handed AI-assisted workflows before they've built the foundational understanding those workflows require.

An Anthropic-conducted randomized controlled trial found that junior developers using AI assistance scored 17% lower on mastery quizzes compared to peers who completed the same tasks without AI assistance — despite spending the same amount of time. They traded learning for nothing: no speed gain, no skill gain. Just the performance of productivity without its substance.

The Supervision Paradox

The problem compounds at the senior level. METR's research found that AI success rates drop sharply as task complexity increases: near 100% on 4-minute tasks, below 10% on 4-hour tasks. Effective AI deployment requires experienced engineers who can identify the 4-hour problems before delegating the 4-minute ones — engineers whose judgment was built through years of doing exactly the work AI is now automating away.

The supervision paradox is stark: the engineers most capable of using AI well are those who learned to engineer without it. And the conditions that produced those engineers are being systematically eliminated in the name of efficiency.

Amazon's December 2025 incident made the failure mode concrete. An AI agent with operator permissions deleted an entire environment rather than making targeted changes — a decision no experienced engineer would approve. Production judgment, built through direct experience with failure and recovery, is not something that can be inherited from an AI system. It has to be learned from the systems themselves.

The Labour Market Signal

Anthropic's labour market analysis found that hiring of young workers in AI-exposed occupations dropped 14% starting in 2024. Older workers experienced no equivalent decline. Companies aren't replacing experienced engineers with AI — they're simply not hiring the juniors who would become the next generation of experienced engineers.

The timeline of this problem is its most dangerous feature. The companies making these hiring decisions today will not feel the consequences for three to five years — roughly the time it takes a junior developer to accumulate the experience needed for senior judgment. By the time the gap is visible, the window to close it through training will already be years behind.

What Needs to Change

Allan proposes deliberate structural interventions: mandatory fundamentals training that operates outside AI-assisted workflows, evaluation metrics that measure understanding rather than velocity, and investment in codified context — documentation and decision records — as the infrastructure through which AI-era engineers can learn what was previously learned by doing.

The central argument is not that AI shouldn't be used in software development. It's that the introduction of AI tools without a deliberate plan for preserving skill formation will produce organizations that, in a few years, lack the experienced judgment necessary to supervise the AI systems they've become dependent on. The ladder isn't just missing rungs. The process that created the people who built the ladder is disappearing.