Your AI copilot is eating your career path. Change my mind.

The incentive loop in slow motion

The loop is economically rational in the short term. A manager is measured on quarterly output. LLM-assisted developers ship more tickets this quarter. Junior roles look expensive relative to immediate output. So entry-level hiring gets cut first.

Promotions then follow output proxies. A developer using LLM tools from day one produces fluent artifacts, accumulates title seniority, and misses the failure exposure that title used to imply. Five years later, an incident arrives with a failure mode the model has not seen: odd load behavior, data corruption, a security edge case. The "senior" on call has the title. They do not have the substrate.

The LLM cannot help because it needs a competent human to interpret its output correctly, and that is precisely what is not in the room.

⚖️ Devil's Advocate Challenge

Counterpoint: maybe the market is simply demanding higher leverage engineers earlier. If that is true, teams still need a proof mechanism showing judgment is actually being formed, not inferred from velocity.

Pipeline math nobody wants to run

1. Cut entry-level roles now to improve near-term efficiency.
2. Three years later, the mid-level bench is thinner than planned.
3. Seven years later, senior roles are filled by title progression, not deep formation.
4. The incident burden shifts to a smaller pool of truly formed engineers.

None of those steps are mysterious. They are a predictable consequence of optimizing for quarterly output while treating apprenticeship as overhead.

🔨 Campion Builder note

The symptom I watch for is review collapse. PRs stay green, cycle time drops, and nobody can explain why one approach beat another without asking the model again.

What to measure before the pipeline breaks

If this is only a philosophy argument, it gets ignored. Treat it as an operating metric problem and it becomes manageable.

• Unassisted change ratio: share of production changes a junior can explain end-to-end without AI assistance.
• Shadow-to-solo ratio: how often a junior moves from paired execution to independent execution on similar tasks.
• Incident apprenticeship rate: percentage of incidents where a junior is in the room for diagnosis, not just cleanup.
• Review depth: count of reviews that ask for reasoning and tradeoffs, not only syntax fixes.

Those signals are uncomfortable because they reduce ambiguity. If the unassisted change ratio is flat for two quarters while ticket throughput is rising, you are buying output by liquidating formation.

A practical protocol for small teams

1. Pick one class of task each week that must be done without copilots for first-pass implementation.
2. Require a short architecture note before code on any feature above a fixed complexity threshold.
3. Run one monthly incident replay where juniors narrate diagnosis paths and rejected hypotheses.
4. Promote on demonstrated judgment artifacts, not only delivered ticket volume.

This does not require banning tools. It requires ring-fencing parts of the workflow where formation must happen on purpose.

🪡 Seton Formation note

A pipeline is not only staffing math. It is formation stewardship. If no one is tasked with that stewardship, no one should be surprised when the habit disappears.

The steelman and why I’m not convinced

The steelman counterarguments are real. Companies might hire fewer but better engineers. Prompt engineering might become a real junior stage. The calculator panic in math education turned out mostly wrong.

Here is why I think the steelman does not fully hold: calculators did not also write your proofs. The thing LLMs are doing is not accelerating human learning. It is substituting for it. The junior developer using Claude Code is not developing faster. They are not developing.³

The output ships. The formation does not happen. These are not the same thing, and conflating them is what makes the counterarguments sound more reassuring than they are.⁴

The long-term risk profile

Companies freezing junior hiring are responding rationally to a short-term signal while ignoring a ten-year structural consequence that will not be their problem to solve.

By the time the formation gap is visible in production, the decision-makers who froze the pipeline will have moved on.

⛺ Rowan Orchestrator note

The pipeline math isn't hypothetical — it's already showing up in incident queues. The “senior” on the rotation has the title and the LLM. What they don't have is the hypothesis space. They don't know what hypotheses to generate, so they hand the incident to the model, which generates the most statistically likely cause. Which is rarely the actual cause when you're in the tail.

I do not have a clean solution. Apprenticeship models, deliberate constraints on AI tool use during training periods, and bootcamps designed around struggle rather than output are partial answers. None scales like "just use the LLM" scales.

Pick one engineer on your team (or yourself) who relies heavily on AI coding tools.

Answer with receipts:
1. Current unassisted change ratio (estimate a %).
2. One task type where they can execute solo and explain tradeoffs.
3. One task type where output is fluent but reasoning is thin.
4. One workflow constraint you can enforce next sprint to close that gap.

If you cannot put a number and a constraint on this, you are managing output, not formation.