The most uncomfortable software study of the last year did not come from a job board or a layoff tracker. It came from METR, which ran a controlled experiment on experienced engineers doing real tasks and found they were 19% less productive when using AI tools — while predicting they were faster.
Sit with the second half of that. The slowdown is interesting. The fact that they could not perceive it is the actual finding. When the AI is doing the typing, your sense of your own competence drifts away from your real competence. And the interview, increasingly, is the place that gap gets measured.
Atrophy is quiet until it is loud
Skills do not announce their own decline. You still ship features. The PRs still merge. The AI fills the gaps so smoothly that nothing feels missing — until you are in a setting where the AI is gone or where the interviewer pokes one layer past the autocomplete, and the floor is not there.
You do not forget how to code. You forget how to derive. The pattern still looks familiar, but the reasoning that lets you adapt it to a twist has gone soft from disuse. In normal work you never notice, because you rarely hit the twist. An interview is nothing but twists.
This is why Gartner predicts half of organizations will require AI-free skills assessments by 2026, and why the in-person rounds got harder even as AI got allowed. Companies read the same studies you did. They are no longer willing to assume that "ships code with Copilot" means "can reason about code."
Where it shows up in the room
Atrophy has tells. Interviewers running modern rounds know them, and they probe for them on purpose.
| The tell | What it reveals |
|---|---|
| Reaches a working solution but cannot explain the tradeoff | Pattern was retrieved, not reasoned |
| Stalls completely when the edge case is twisted | Adapted from memory, not understanding |
| Cannot estimate complexity without running it | Lost the underlying model |
| Describes what the code does but not why | Read the output, never built the intuition |
None of these mean you are a bad engineer. They mean a specific muscle — unassisted derivation — has not been loaded in a while. The good news is that it is one of the most trainable muscles there is.
The fix is not "stop using AI"
That advice is both impractical and wrong. AI is part of the job. The fix is to deliberately keep a sharp unassisted baseline alongside your assisted daily work, the same way a pilot keeps manual-flying hours even though the autopilot does most of the flying.
Use AI to go fast at work. Keep a regular practice where you go slow with the AI off. The interview measures the second one, and so does the day everything breaks in production.
The market is now built to find the gap
For a few years, the gap between assisted and unassisted skill was invisible — nobody tested for it, so it did not cost you anything. That window is closing. The AI-free round, the comprehension round, the harder onsite — all of them are instruments for measuring the difference between what you can do with the model and what you can do without it.
The engineers who stay hireable in 2026 are not the ones who avoid AI. They are the ones who use it daily and still keep the unassisted version of themselves in shape — because the interview, and eventually the job, will ask to meet that version directly.
Keep it sharp. It is the part of you that does not show up in the PR, but absolutely shows up in the room.
Rubduck is spoken interview practice with a live AI interviewer that pushes on your reasoning, not just your output — so the unassisted, derive-it-yourself muscle stays loaded. Practice the version of you the interview is actually built to measure. Start your free sessions →