The In-Interview AI Use Decision, Scored Round by Round

You are staring at a calendar invite for one interview round, and you need to decide one thing before you join: can you use an AI assistant during it, and if so, how do you say so without it reading as cheating. This guide is for candidates who want a per-round decision, not an abstract ethics debate. It hands you a scoring model, four verdicts, and a disclosure sentence you can send.

The reason this is hard is that the answer moved. Permission used to be a company property - Amazon bans it, some startup allows it. Now it is a round property. Google bans AI in its standard loops while piloting Gemini in one code-comprehension round. So the honest answer to "is using AI in interviews cheating" is: it depends on the round in front of you, and you can score that round in about twenty minutes.

Why permission is now a round property, not a company one

Permission has migrated from the company to the round, which means a recruiter's company-level answer can be true and still wrong for your specific session. Employers now redesign scoring per round rather than per brand, so one loop can require an AI assistant while the next forbids it.

Look at the split. Meta introduced an AI-enabled CoderPad round in October 2025 for mid-level (E4) and senior (E5) Software Engineer roles, and pairs it with a separate AI-free round. Canva goes further and makes AI mandatory in Backend, Frontend, and Machine Learning engineering rounds. Google bans AI in standard loops but is reported to be piloting a Gemini-in-CoderPad format for code comprehension. Amazon bans it outright, with disqualification stated as the consequence. Goldman Sachs bans it in recorded video interviews.

Read the table twice. Meta and Google appear on the "allowed" side and the "banned" side of their own process. That is the whole problem in one grid. A candidate who hears "Meta is fine with AI" from a friend, and applies that to the AI-free half of the loop, has just made a company-level assumption about a round-level rule.

The reason this migration matters for scoring is that it invalidates your cheapest source. You cannot ask "does this company allow AI" and act on the reply. You have to ask "does this round allow AI" and get the round named back to you.

What the scored round is actually testing

An AI-permitting round is not testing whether you can type. It is testing AI fluency: prompting, verifying the output, and reasoning about it out loud. That is why underusing the approved tool can count against you as much as overusing a banned one.

The market signal here is loud. In 2026, according to Dice, 71% of tech job postings in the US require some form of AI fluency, a 181% year-over-year increase. When an employer builds a round where the tool is required, the round is a job proxy: they are watching how you drive the assistant, catch its mistakes, and explain your choices. Hand yourself the answer and stop thinking, and you fail the thing they are actually grading.

This skill is thinner on the ground than the hiring frenzy suggests, and it is geographically skewed. In Refolk's index of professional profiles, the exact skill these rounds test is underrepresented, which makes demonstrated fluency a genuine differentiator rather than table stakes.

The practical read: if a round allows AI, treat that as an invitation to show fluency, not a shortcut to skip the work. The verdict "use openly" is not permission to coast. It is a scored event.

The four verdicts, and the signals that produce each

Every round resolves to one of four verdicts, and the verdict is a function of two things: how strong the permission signal is, and how exposed you are if you are wrong. Below is the map, then the scoring.

• Use openly. The round is AI-required or explicitly AI-allowed in writing, and the platform supports it. Drive the tool, narrate your reasoning, verify its output.
• Use and disclose. AI is plausibly fine but not explicitly named for this round, or you are on a platform where a detector could false-positive you. Use the tool and state that you are, in one plain line.
• Prep-only. No written permission, a standard live round, or a video screen where the norm is unassisted. Use AI to prepare a story bank and rehearse, then close it during the round.
• Do not use. A proctored lockdown, an explicit ban like Amazon's, or a recorded round like Goldman's HireVue. Any assistance breaks a rule you accepted, not just a norm.

The two axes that produce the verdict are permission strength and detector exposure.

Note where "do not use" sits: high detector exposure with weak or silent permission. A proctored assessment lands there by default. When you explicitly accepted full-screen lockdown and process monitoring, using assistance violates rules you agreed to, which is a different category of problem than breaking an unwritten norm.

Permission strength decides whether you can; detector exposure decides how badly it goes if you are wrong.

Read the detector surface before you decide

Detectors watch three surfaces, and knowing which one your platform uses tells you what "use and disclose" actually protects against. In-browser telemetry is the most common and cheapest bucket: tab switches, window focus loss, clipboard pastes, keystroke timing, mouse movement.

HackerRank runs tab-proctoring, copy-paste tracking on by default, and AI plagiarism detection cited at 93% accuracy on typing patterns. CodeSignal computes a "Suspicion Score" from paste events, typing-speed anomalies, and code-similarity checks. CoderPad logs pastes, plays back keystrokes, and can use the webcam, but it monitors the editor, not your whole desktop. Async video tools like HireVue read speech pace and hesitation and can log blur, but cannot see another screen.

Two things to carry from this table. First, the webcam catches the off-screen glance but not the on-screen read: it cannot tell whether you are reading the question or reading an overlay positioned next to it. The tell is the glance, not the read. Second, and more important, none of these is the real detector.

The most reliable detector is conversational. It is one unscripted follow-up about code you did not write. An overlay defeats tab and screenshot telemetry, but it cannot help you explain, in your own words, a design decision you never made. Reading a generated paragraph produces every signal at once: uniform latency, scanning eyes, documentation-style phrasing, and zero depth on the follow-up.

CoderPad is the one platform with a documented per-pad toggle: an interviewer enables or disables AI Assist for an individual pad from Pad Settings, and the organization default does not gate availability. This is the surface Meta uses and Google is reported to be piloting. If your round runs on CoderPad and the interviewer says AI is on, that is a strong, verifiable permission signal.

The procedure, run in under thirty minutes

Run these seven steps in order before you join the call. Each has a clear "done" state, and the whole sequence fits in the gap between scheduling and the interview.

Step one carries the most weight, so use a script that reads as process-minded rather than nervous. Asking costs nothing and carries no stigma. Assuming is the expensive option in both directions: assume permission and you risk disqualification, assume prohibition and you underperform.

title: The pre-round permission ask
note: Send to the recruiter after scheduling. It reads as a candidate who thinks about process, and the reply removes the only version that ends badly.
Hi [name], quick logistics question before the [round name] round.

Are any of the rounds designed for me to use my normal tools, or should I plan to work unassisted throughout? Happy either way - I just want to prepare correctly.

Thanks.