Using AI in an interview: who sets the rule

The rule belongs to the interview, not to the tool

The question is usually asked as though it had one answer somewhere, and finding it were a matter of searching harder. It does not. A hiring process is private, run under whatever conditions the employer chooses to set. Those conditions can forbid outside help of any kind, can say nothing at all, or can require you to use a model and score how you did it. All three exist right now, at companies of similar size, in the same industry, in the same month.

Which is why no vendor can answer this for you, including us. A tool can tell you what it does and what it does not conceal. It cannot tell you what your interviewer permits, and any tool that implies otherwise is selling you a feeling rather than information.

What four large employers say, as of September 2026

• Amazon: prohibited, with disqualification named. Internal recruiting guidance reported in 2025 bans generative AI during interviews and calls teleprompter-style tools that feed real-time prompts an unfair advantage. The stated consequence for candidates found using one is removal from the process.
• Goldman Sachs: prohibited. The bank has asked candidates not to use ChatGPT or any outside source during interviews, on the stated grounds that it wants to hear applicants in their own voice.
• Anthropic: allowed to prepare, not during the live round. Its candidate guidance invites you to use Claude on your application and to prepare, and is explicit that there is no AI assistance during a live interview unless the interviewer says otherwise, because the point is to see how you think in real time.
• Google: piloting the opposite. Through the second half of 2026 Google is testing a code comprehension round in which candidates use Gemini on an existing codebase, with interviewers scoring prompt quality, verification of the output and debugging. It is a pilot on selected US teams for early and mid-career roles, and it sits alongside standard rounds where assistance is not part of the format.

Other large employers have run comparable experiments in both directions — some bringing back in-person stages specifically because of remote assistance. Both movements are happening at once.

The pattern underneath the policies

The disagreement looks less arbitrary once you notice that the two camps are measuring different things. A round that forbids assistance is measuring what you hold and how you reason without a net — which is a real signal, and the one that transfers to a whiteboard, a customer call or a production incident at two in the morning. A round that supplies a model is measuring something newer: whether you can frame a task, read the output critically and catch it being wrong.

That makes the company the wrong unit. The same loop can contain both kinds of round, which is exactly what Google’s pilot describes. “Does this employer allow AI” is already too coarse a question; the answerable one is “does this round”.

“Can the interviewer tell?” is really two questions

The first is whether anyone can see the tool. Nobody can promise you an answer to that, in either direction, and a promise that they can is the clearest signal you are being sold something. The second question is the one that actually decides interviews: whether the answer you gave looks like an answer you built.

Karat, which has run and reviewed more than half a million technical interviews, publishes the signals its interviewers are told to watch for. None of them involve finding software. They are behavioural: frequent glancing off-screen, a complete solution arriving in one uninterrupted pass with no visible iteration, large blocks of code appearing at once, an explanation that does not match the code it belongs to, no sense of edge cases or failure modes, and jumping straight to an optimised approach with no reasoning in between. Amazon’s reported guidance adds the same family of cues — typing while a question is still being asked, answers that sound rehearsed, eyes tracking something else on the screen.

This is worth sitting with, because it reframes the whole worry. The risk is not that a window gets found. It is that a fluent paragraph after a long pause does not sound like the six minutes of conversation that preceded it, and interviewers notice a break in pattern long before they notice a tool.

What the employer side is buying now

There is also a straightforward commercial answer forming. Honorlock, long established in exam proctoring, announced employer-facing interview products on 22 September 2026: identity verification before the session, blocking of unauthorised AI assistants and remote access software, flagging of suspicious activity during the call, and timestamped session reports handed to the employer afterwards. Deepfake impersonation and remote-control tools are named alongside answer suggestion as the things being targeted.

Independent write-ups of this category are candid that webcam-first designs miss assistance delivered on a second device, which is presumably why the products keep growing new signals. That contest will continue, and it is not a contest a candidate has any reason to enter. If an interview is proctored, the rule has been stated about as loudly as a rule can be.

How many people are doing it anyway

The numbers vary enormously depending on who counted and what they counted, so treat them as a range rather than a figure. Interview-integrity vendors report flag rates climbing through 2025 and into 2026, with one dataset covering roughly 19,000 interviews putting it near two in five overall and higher in technical roles. Self-reported surveys land much lower — around one in five workers admitting to using AI during a live interview.

Vendor flag rates and self-reports measure opposite biases, which is roughly why they differ by a factor of four. What none of them is in dispute about is the direction, and it is large enough that employers have rewritten their processes over it. That is the part a candidate can actually observe.

Finding out the rule for your own interview

• Read the invitation and the assessment page. Prohibitions are usually stated, and stated badly — a line about “your own work” in the third paragraph counts.
• Ask the recruiter in writing. One email before the loop. The reply is the answer, and it is also the record.
• Treat silence as no. If nobody has said assistance is permitted, assume it is not. That is the working default for 2026 and it costs you very little.
• Ask per round, not per company. A take-home, a live coding round and a behavioural conversation can each have a different answer at the same employer.
• Separate preparing from performing. Almost nobody objects to using a model to prepare. Nearly all of the disagreement is about the live round.

Where we stand

We build an interview copilot, so we are not a neutral party and it would be odd to pretend otherwise. What we can do is be consistent: Trippi Cue does not hide itself, we make no claim that it cannot be noticed, and we build nothing aimed at working around proctoring. Those are not features waiting on a roadmap — they are decisions.

The uses we build for are the ones that survive being described out loud afterwards: following an interviewer in a language you know less well, catching a question that bad audio ate, keeping a long answer from wandering. The rest is yours to read and yours to decide, and our full position, including what we will not build, is on the responsible use page.

Read next

• Interview Prep Tools vs Help During the Call
• Why an Interview Copilot Should Be a Browser Extension
• What Is an Interview Copilot and How Does It Work?