Narrative-first, evidence-driven

A hiring manager has fifteen open roles.

Three are posted as junior engineer jobs.

On paper, that sounds normal.

In practice, she fills them with people who already have five or six years of experience. The boilerplate work, the starter bugs, the rough first draft stuff, the documentation, the test scaffolding, a lot of what used to train new people now gets done by senior people with AI help.

The junior posting stays up.

The junior path doesn’t.

She doesn’t think of this as some grand policy choice. She thinks of it as being practical.

Nobody has to stand up in a meeting and say, “Let’s break the ladder.”

You can break it just by making a hundred practical choices in a row.

What’s happening

AI is taking over a lot of the work that used to make up the first rung of IT.

Not all of it. But enough of it to matter.

The work humans keep is more likely to be the work that requires trust, judgment, access, and accountability. The work AI takes first is often the work people used to learn on and that makes the ladder thinner at the bottom.

If you don’t protect the learning path on purpose, “human review” starts sounding better than it really is. The person is still there, the approval step is still there, but the skill underneath it may be getting weaker.

That’s the deeper problem: this isn’t only a jobs story, it’s a judgment story.

Why it’s happening

The pressure is pretty simple: if a senior engineer with AI assist can do in one afternoon what used to take a junior a couple days then the manager sees speed, lower cost, fewer handoffs, and fewer open reqs to fight for.

From that seat it looks rational and that’s what makes it dangerous. The damage doesn’t come from some cartoon villain. It comes from a system rewarding short-term throughput and treating learning as overhead.

Then the gate shifts. The question stops being “Can you do some starter work?” and becomes “Do you have the standing to be trusted with the harder call?”

That’s a much tighter gate. Once that happens a few things follow fast:

1. entry gets harder
2. mid-career growth can flatten
3. fewer people have enough real reps to know when the machine is wrong
4. oversight gets weaker

That’s the part I keep coming back to: you don’t get meaningful oversight by putting a human name on the form, you get it by having people who actually know what they’re looking at.

What we know, and what we don’t

The evidence here is early. The whole ladder hasn’t collapsed.

Plausible from available evidence: AI is taking over a meaningful share of the tasks that used to train junior workers, entry-level hiring is tightening in technical fields, and organizations are moving faster on efficiency than on protecting the learning path. The mechanism is consistent and the incentives are clear. But we don’t yet have robust longitudinal data showing ladder collapse at scale.

Unknown: whether new AI-adjacent roles are opening at anything like the same rate as starter roles are shrinking, whether those roles are stable and well-paid, and whether today’s junior squeeze becomes tomorrow’s mid-level squeeze as the cohort that missed the early reps works its way up. Answering those questions would require role-mix and compensation data tracked over time.

This is exactly the kind of problem where waiting for perfect proof may be too late to avoid long-term damage. The mechanism is visible. The response doesn’t require certainty about the final scale.

Who benefits, and who carries the risk

In the short run, the organization can get real gains:

• teams will move faster
• senior people will produce more
• leaders will get the productivity story they wanted

Those are major gains. But the risk goes somewhere.

The first risk lands on junior workers because the door narrows.

The next risk lands on mid-career workers if their growth turns into cleanup, prompt handling, and machine-checking without enough independent work to keep building real judgment.

Then the delayed risk lands on the institution itself.

Because later, when the AI output is wrong, brittle, insecure, unfair, or just sloppy, there may be fewer people left who can really tell.

The company may win the sprint and still lose the bench.

What good looks like

The good version isn’t hard to picture:

• use AI for the drudge-work
• let it help with repetitive scaffolding, rough drafts, routine cleanup, and other low-value grind

Don’t cash all of that saved time out as more output. Use some of it to protect the part that actually builds people:

• mentorship
• pairing
• real debugging
• design discussions
• unassisted reps

Work where someone has to think through the problem and not just supervise the answer.

That’s the line I care about: use the tool to reduce grind without hollowing out the path into the work.

If the machine does the starter work and nobody protects the learning, then later human oversight carries the label without the substance. The review is still there, it just no longer means much.

What to do

At the team level, intentionally protect learning work. Not someday or “when things calm down.” Name it, schedule it, and treat it like production work.

That can mean unassisted tasks on purpose. It can mean pair-review rotations. It can mean protecting a slice of junior work for actual skill formation instead of squeezing every saved minute back into the sprint.

At the institution level, stop settling for “a human signs off.” That’s too weak. Ask better questions:

• how much time is actually being spent reviewing AI-generated work?
• how often do people override the tool?
• are junior roles shrinking?
• are “junior” jobs being filled by experienced people without saying so directly?
• can an applicant tell when AI screened them out?
• can they appeal it?

Those are much better tests of whether human oversight still means something.

If a company or public institution buys AI for hiring, evaluation, or technical review, the usual checks still have to hold: people should be told what happened, given a real reason, able to inspect the record, and able to reach a human who can actually change the answer. Oversight has to be more than a name attached at the end.

How to talk about it

I wouldn’t lead with “AI is destroying jobs.” That’s too broad, and it misses the sharper point.

I would say it like this:

• I’m not arguing against better tools. I’m arguing for keeping the ladder intact.
• If we use these tools to cut grind and help people do better work, great.
• If we use them in a way that slowly removes the training path for the next generation, then we are borrowing from the future to make this quarter look better.

That feels like the right frame to me because it lowers the heat without lowering the standard.

Most leaders aren’t trying to wreck their teams, they’re trying to get speed under pressure.

The point is whether they’re getting that speed in a way that leaves them stronger or weaker in the long term.

What to ask

Where do people learn now?

That’s the question underneath this whole case. If the old starter work is going to the machine, where is the new path to judgment, trust, and real competence? Watch for the moment when AI is sold as support, but the actual effect is that fewer people ever get enough reps to become strong reviewers themselves.

If nobody has a good answer to that then the ladder is probably not shifting, it’s thinning.

One steady action to take this week

Pick one piece of technical work this week that stays intentionally unassisted and treat it as skill-building work, not inefficiency.

That’s small enough to do right now. It’s a good test of whether your team is still building judgment or just accelerating output.

Action ladder

Short term

• Managers: Protect one piece of intentionally unassisted work this week and treat it as training, not wasted throughput.
• Workers: Notice where the learning reps are disappearing. If AI is absorbing the starter work, ask where people are supposed to build judgment now.

Medium term

• Teams: Add one protected practice structure this quarter - pair-review rotations, unassisted reps, or a reserved slice of junior work that is kept for skill formation.
• Leaders: Track more than output. Watch junior-role shrinkage, override rates, and whether “junior” work is quietly being reassigned upward.

Long term

• Executives and public institutions: Fund apprenticeship paths, not just productivity tooling. If the first rung disappears, the later bench disappears after it.
• Workers, managers, and educators: Push for systems that treat training as production capacity in development, not as optional overhead to be cut under pressure.

Related reading

• When Building Gets Cheaper but Breaking In Doesn’t
• When Your Customers Can Build What You Sell
• When Cheap Content Changes Who Gets Heard
• Case study: AI and the IT Ladder Collapse