AI automates the code it can check

Every "AI is coming for X" take names a specialization. Frontend first, because it's "just" UI. Then junior devs, then QA, then anyone who writes CRUD. Pick a job title, predict its funeral.

The data doesn't move by job title. Look at where AI actually writes the most code: frontend. JavaScript, TypeScript, HTML, and CSS make up 59% of the coding conversations in Anthropic's study of 500,000 of them — Python 14%, SQL 6%. If "the AI writes it" meant "the job dies," frontend engineers would be first in the ground. They're not.

So the specialization frame is broken. What predicts whether AI eats your work isn't what you build. It's whether the result can be checked.

The rule that actually predicts it

Andrej Karpathy put it in one line: traditional software automates what you can specify; AI automates what you can verify. You don't hand the model the steps. You check its output and keep the version that passes.

That makes verifiability the risk line. Work with a fast, automatic pass/fail — does it compile, does it render, does the test go green — is work AI can grind on cheaply, because it can try, check, and retry a hundred times a second. Work with no cheap oracle — will this schema hold at 50x traffic, is this the real cause of the outage, will a screen reader make sense of this — AI can't self-grade, so it can't self-improve.

Anthropic's own numbers line up. On Claude Code, 79% of conversations are automation — the model doing the task outright — versus 49% on the general chat product. Automation clusters exactly where the feedback is instant: compile, render, run the test. Frontend leads the language chart not because AI is better at it, but because a browser returns a verdict on every keystroke. That's why it's first — and it's also why the panic misreads what "first" means.

"Checkable" is a layer, not a job

Here's what the specialization takes miss: the checkable work and the judgment work live inside the same role. AI takes the checkable layer. What's left, in every specialty, is the judgment sitting on top.

Frontend is the cleanest example, because it's where AI writes the most and the split is most brutal. The model scaffolds the component, wires the state, matches the mock pixel for pixel. Then look at what has no pass/fail button. WebAIM's 2026 audit of the top million homepages found 95.9% fail basic accessibility checks, at 56.1 errors per page — both up from last year, as pages got more complex and more machine-generated. The pixels render. The accessibility, the cross-browser edge cases, the "why does this re-render 40 times" — no oracle, no automation, still yours.

Same cut, every specialty:

• Backend — AI writes the CRUD endpoint and the migration. Whether the schema survives real traffic, and which caching strategy won't corrupt state, has no green checkmark.
• DevOps / platform — routine Terraform and threshold tuning are checkable, and going. Reading an incident at 3am and knowing which dashboard is lying is not.
• Data / ML — the glue code is generated. The models still hang on company-specific, undocumented business logic, the least checkable thing in the building.
• QA — writing the assertion is trivial and automatable. Knowing which test actually matters for this system is the judgment that doesn't.

The market already repriced it — upward, not out

If AI were killing engineers, postings would fall with the rest of the market. They're doing the opposite. Indeed's Hiring Lab found US software-development postings up almost 15% since Claude Code launched in early 2025, while total postings fell 7% — the divergence in the chart below. The catch is where the growth went: 71% of it is senior roles, and 37% of the new postings name AI in the title. The junior, checkable rung thinned. The senior, judgment rung grew.

SignalFire's 2026 report says the same from the employer side. Across twelve of the biggest tech companies, total hiring is down 25% from 2019, but engineering hiring is down only 11% — and engineers went from 46% of new hires to 55%. Companies cut the org and kept the judgment layer.

None of this is a return to the old normal; software development postings are still about 27.5% below pre-pandemic. It's a reshape. The work that survived is the work with no pass/fail button, and it's clustered at the senior end.

What that means if you're the one job-hunting

Remote skews the same way, if not harder — the roles that clear a remote bar tend to be the ones employers trust to run without a babysitter, which is judgment work by definition. So for a remote engineer deciding where to point the next two years, the read is the same as the market's:

• Sell judgment, not stack. "React, Node, Postgres" is a list of things AI also knows. "I cut our p95 by 60% and can tell you which query did it" is not. Lead with the calls you've made, not the tools you've held.
• Treat AI fluency as table stakes. With 37% of new software postings naming AI, it's a filter now, not a differentiator. Expected if you have it, disqualifying if you don't.
• Aim above the checkable rung. The thin part of the market is exactly the part AI does best. If you're early-career, get to judgment work fast — on anything with real consequences — instead of racing a model on scaffolding speed.

The old question, "is my specialization safe from AI?", has no useful answer. The one that does: how much of my day survives without a pass/fail button? Whatever's left is the job now, and it's the part that's hiring.