Promoted to Checker

Before any of it was engineering, it was a moulding machine.

The first years at LEGO were spent as a polymer technician, standing at injection moulding machines and learning what plastic actually does when it cools – which is not quite what the datasheet says it does. Warp. Sink marks. The way a wall thickness that looks fine on a screen produces a part that fights you on the line.

Nobody called that a curriculum. It was one. Every piece of judgment used later, in engineering and then in leading engineers, was bought on that floor, one ruined part at a time.

That rung is the one now being automated first.

Ask practising engineers what they want from AI and the answer is remarkably consistent. They want the grunt work gone.

Drafting. Documentation. The standards binder. The end-of-year self-review. The email telling a client this will cost more than they hoped. In months of reading how mechanical, controls and manufacturing engineers actually talk about this – on Reddit, on PLC and eng-tips forums, on Hacker News – the desire to hand over the drudgery was close to universal, including among the hardest skeptics.

One mechanical engineer put the terms of surrender exactly: "If I could replace all my drafting, presentation compiling, and reports with AI I honestly wouldn't mind it so much. But touch my design freedom and there gonna be trouble."

That is a clean deal, and it is the deal the market is offering. It also has a problem nobody has named.

The drafting was not separate from the design freedom. The drafting is where the design freedom was earned.

The numbers do not say what either camp claims.

The role is not vanishing. The US Bureau of Labor Statistics projects mechanical engineering employment up nine percent through 2034. Two things sit underneath that headline. The previous edition projected eleven percent, so the forecast was quietly revised down. And BLS projects software developers at fifteen percent – the official government forecast has software growing faster than mechanical engineering, not slower.

So the profession is fine. The entrance to it is not.

New-graduate hiring is down about twenty percent since late 2019, and engineering majors took the steepest fall of the twenty most common majors, down twenty-five percent. Two separate large studies agree on the mechanism: at firms adopting generative AI, junior employment falls roughly nine percent within six quarters, driven by slower hiring rather than layoffs, and workers aged twenty-two to twenty-five in AI-exposed occupations show a sixteen percent relative employment decline while their senior colleagues at the same firms hold steady.

Both are working papers, not settled science. But they agree with each other and with the hiring data, and they describe a shape: the top of the profession is stable and the bottom is closing.

The profession has already run this experiment once, and the government wrote down the result.

Look up drafters in the same BLS handbook. Employment: little or no change. The stated reason is not offshoring or the business cycle. It is that CAD raised drafter productivity and "allow[ed] engineers and architects to perform many tasks that used to be done by drafters."

One rung was absorbed into the tooling, in plain sight, over about thirty years. It happened slowly enough that nobody wrote an essay about it. The engineers who had already climbed past it barely noticed, because losing it cost them nothing.

A controls engineer on a PLC forum said the rest better than any analyst has: "If I'd used AI to program that first traffic light sequence as an apprentice, I wouldn't know enough now to do the job that I'm paid to do."

The grunt work was never valuable as output. Nobody misses hand-drafting. It was valuable as tuition.

Now look at what the tools are actually good at.

Generation is not it. In a benchmark published this May, researchers gave frontier coding agents free-form engineering briefs and validated the output with finite element analysis. Across four hundred first attempts, not one produced an assembly that passed. A round of FEA feedback across another four hundred submissions added exactly one. A separate benchmark scores the cascade: roughly seventy-seven percent of generated designs execute as code, seventy-one percent are geometrically valid, and manufacturability – the lowest-scoring dimension of all – sits below half. It compiles, then it looks plausible, then it fails to be a part anyone could make.

Verification is where the deployment and the money actually are: design failure prediction, design-for-manufacturability checking, automated drawing review.

The market is buying the checker, not the designer. An engineer on Reddit got there first, as gallows humour: "Congratulations! You've been promoted from engineer to AI supervisor and double-checker."

Here is the part that should stop anyone planning to settle into that role.

In 2024 the National Society of Professional Engineers' board of ethical review considered an engineer who relied on AI-generated plans and specifications without a comprehensive verification process, and then sealed them. The board found he had not maintained responsible charge, and quoted the definition that decides it: "Reviewing drawings or documents after preparation without involvement in the design and development process does not satisfy the definition of Responsible Charge."

Read that slowly. Checking is not a position the rulebook recognises. The seal means you directed the work, not that you inspected it afterwards.

The rest of the profession has converged on the same line. NCEES adopted a position statement in 2025 stating that licensed professionals "must not use AI to practice outside of their professional competency." The Texas board, asked to write new AI rules, declined and said the existing ones already cover it: licensees "are ultimately responsible for any work product they sign and seal." ASCE put it most bluntly – AI cannot be held accountable.

The insurers moved faster than the lawyers. Standard general-liability forms excluding generative-AI-related claims took effect this January, including the products and completed-operations coverage that responds when a manufactured part injures somebody, and more than sixty insurance groups have filed to adopt them.

So the profession requires exactly the deep involvement that the tooling is designed to remove, and the way engineers historically earned the judgment to be involved was the work now being automated first.

The honest problem is that nobody has measured whether this actually breaks anything.

There is no study tracking whether engineers working alongside AI still acquire judgment over time. The apprenticeship argument is a theoretical model plus a labour-market correlation, and I would not accept that standard of evidence from a vendor. The strongest peer-reviewed result cuts the other way: in a field experiment with 776 professionals, one person with AI outperformed a two-person team without it, and less-experienced staff matched teams containing experienced members. Though the task was product innovation scored by graders. No tolerance failed. No part refused to mould.

Lisanne Bainbridge saw the shape of it in 1983, writing about automated process control: when manual takeover is finally needed, something has gone wrong, and the operator "needs to be more rather than less skilled than before."

Producing senior-quality output and becoming a senior engineer are not the same event. One is measured in a session. The other used to be measured in ruined parts.

We automated the first rung once already and called it progress, because the people who noticed had already climbed it.