Should candidates be allowed to use AI in technical interviews?

Should candidates be allowed to use AI in technical interviews? For most engineering roles, yes — if the job involves directing an AI coding agent day to day, testing someone without it measures a skill nobody will pay them for. But "allowed" isn't a single policy. Companies right now are split between an outright ban, a supervised allowance, and building AI directly into the assessment — and each choice says something different about what they're actually trying to measure.

Why a flat "no AI" rule measures the wrong thing

A ban assumes the skill worth testing is writing code unassisted. For a growing share of engineering work, that's no longer the skill in question. In the 2025 Stack Overflow Developer Survey, 51% of professional developers reported using AI tools daily — not occasionally, daily. An interview that removes the tool a candidate will reach for an hour into their first day isn't testing a cleaner version of the job; it's testing a different job that mostly doesn't exist anymore. (Our companion piece on interviewing engineers who use AI coding agents goes deeper on what to measure once AI is in the room.)

Which companies ban AI in interviews

The ban camp includes some of the most AI-invested companies in the world, which is the tension at the center of this whole debate. Goldman Sachs emailed campus recruits in June 2025 saying external sources, including ChatGPT or a search engine, were strictly prohibited during its HireVue interviews. Amazon has told recruiters to disqualify candidates caught using AI tools during an interview. And Anthropic — the company that builds Claude — still bars AI use during interviews and most assessments, even after it loosened its rules in 2025 to let candidates use Claude on resumes and cover letters. The common argument is that an interview is supposed to hear the candidate's own voice, and a company can want that even while it sells, or builds, the tool it won't let candidates use on the day.

Which companies allow or build around AI in interviews

On the other side, Meta began piloting AI-enabled coding interviews in 2025: candidates in a limited pool work through the assessment with an assistant like ChatGPT or GitHub Copilot available. The stated rationale is twofold — it's closer to how engineers actually build, and it makes the kind of covert AI use that bans try and fail to stop beside the point, since the tool is already on the table. Whether this expands past a pilot isn't public yet, but the direction is notable coming from a company with that much hiring volume.

Does banning AI actually stop candidates from using it anyway?

Mostly not, and that's the practical problem with a ban that has no way to verify compliance. Overlay tools that read a shared screen without appearing in the recording exist specifically to beat a watch-and-hope policy. Coderbyte's 2026 State of Hiring report put it in numbers: 81% of the organizations it surveyed said they'd dealt with candidates using AI without disclosure during an evaluation. A rule candidates can quietly ignore doesn't protect the signal you wanted the ban to protect — it just moves the AI use somewhere you can't see it and can't grade it.

What a defensible AI policy actually looks like

The choice isn't really ban versus allow — it's whether your policy matches what the round is for, and whether you can tell candidates about it in plain language before they start:

• Testing unaided fundamentals on purpose(compilers, concurrency, a narrow CS-theory round) — a no-AI rule is coherent here, as long as it's labeled as a fundamentals check and not presented as the whole evaluation.
• Testing the actual day-to-day jobfor a role where engineers work agent-first — AI shouldn't be banned or bolted on as an afterthought; it should be the tool the whole assessment is built around, with the rubric scoring how the candidate directs it.
• Testing judgment under ambiguityfor anything in between — allow AI, say so up front, and grade the process: the prompts, what got verified, where the candidate overrode a wrong answer. That's the one policy that doesn't depend on catching anyone, because it works whether AI use is visible or not.

The one policy that doesn't survive contact with 2026 is the unstated ban — no AI rule, no AI-off mode, no way to tell afterward — because candidates will fill that silence with whatever tool they already use, and you'll have learned nothing about how they used it.

Where DeftBench fits

DeftBenchis built for the middle policy above: candidates work in a live browser IDE with a real coding agent, every prompt and edit is captured as session telemetry with human and agent actions attributed separately, and the evaluation scores direction and verification against that record rather than guessing at whether AI was involved. If you're comparing that approach to an AI-permissive IDE test, the DeftBench vs CoderPad comparison breaks down what each one actually captures. Or talk to us about a pilot.