AI in interviews: where help ends and cheating begins

The short answer

AI in interviews is not one situation but three, and confusing them is expensive. Preparing with AI is fine. Keeping your own material within reach when the conversation is about your own work is fine. Passing a model's answers off as your experience is deception — the kind that unravels on the first follow-up question, and sometimes much later and much more painfully.

The line does not run along the technology. It runs along one question: does the tool replace the very skill being tested right now? A compiler does not make you an impostor, because on Monday you will still be working with a compiler. A model that speaks for you in an interview will not speak for you on Monday — and the gap between the impression and the reality is yours to close.

Why the question came up at all

A couple of years ago «AI in an interview» sounded like science fiction with a touch of spy movie. Today it is a toggle in an app, and the question moved from technical to ethical — and markets answer ethical questions more slowly than technology changes.

Three things happened at once.

1. Automation arrived from both sides. An algorithm screens the résumé, a model writes the outreach email, and the same model checks the take-home task. A candidate who has not met a live human across four stages draws the symmetric conclusion — and it is hard to fault the arithmetic.
2. The technical barrier disappeared. A prompt no longer needs a second monitor and an accomplice: an application window can simply stay out of screen capture using the operating system's own facilities. How exactly — below.
3. There are no rules. Job posts almost never say whether AI may be used during an interview. Where there is no agreement, each side invents one, and they invent differently: the candidate hears «not forbidden», the company hears «obviously not».

Hence the eternal comment-thread argument, where one side calls it cheating and the other calls it the new calculator. The argument is unresolvable until the cases are separated.

Three questions that draw the line

The check takes about twenty seconds and works without moralising.

Now the same three questions applied to real situations. The verdicts are not legal but engineering: they are about how things end.

A short formula settles almost every argument: if the tool stays with you at work, it is help; if it is needed only for the interview, it is substitution.

How «invisibility» works — and where it ends

Plenty of nonsense has been written about invisible windows, so here is the mechanism without the mystique. An application asks the operating system not to include its window in screen capture. On macOS that is one window property (sharingType = .none), on Windows one call (SetWindowDisplayAffinity with the WDA_EXCLUDEFROMCAPTURE flag). From then on the system's compositor simply does not hand that window to whoever is capturing the screen.

Three consequences follow, and they are worth knowing before rather than after.

  YOUR SCREEN                      WHAT THE OTHER SIDE SEES
  ┌──────────────────────────┐     ┌──────────────────────────┐
  │ call window              │     │ call window              │
  │ ┌──────────────────────┐ │     │                          │
  │ │ panel with a prompt  │ │  →  │        (empty)           │
  │ └──────────────────────┘ │     │                          │
  │ code editor              │     │ code editor              │
  └──────────────────────────┘     └──────────────────────────┘
        your monitor                    the shared stream

  NOT covered by the mechanism:
    • a camera pointed at you — gaze and pauses are visible
    • a phone, or a second person looking at your screen
    • proctoring software that sees the process list
    • your own behaviour in the conversation

First. It works against software capture — screen sharing, recording and screenshots. Nothing works against physics: a camera looking at your face faithfully transmits both the glance to the side and the three-second pause before every answer.

Second. An invisible window and an invisible program are different things. The application remains an ordinary application: it is in the process list and visible in the system monitor. Proctoring software, which you install yourself before an exam, looks exactly there. If you are asked to install such software, consider the rules explicitly announced.

Third. The mechanism has version requirements. On Windows, exclusion from capture appeared in Windows 10 version 2004; on older builds the toggle exists in the interface and does nothing — the window travels calmly into the broadcast. That is exactly the case where «I thought I could not be seen» turns into a story you tell for the rest of your life.

The tool itself — what it does before, during and after an interview, and why «second brain» is a precise description rather than a figure of speech — is covered separately in the AI interview assistant.

What actually gives a candidate away

Here is the real watershed, and it is rarely discussed: candidates are caught by behaviour, not by software. The interviewer does not see your screen — they see you, and the set of signals has been the same for years.

• Rhythm. An even three-to-four-second pause before every answer, including the easy ones. A live person answers unevenly: «tell me about yourself» comes at once, «the worst technical decision of your life» comes thoughtfully.
• Gaze. Steady reading away from the camera, synchronised with the start of the answer. On its own it proves nothing; together with rhythm it is enough.
• Style. A perfectly structured answer to an unstructured question. «Firstly, secondly, thirdly» in response to «so what do you make of this task?» is written prose, not speech.
• Emptiness inside the detail. The right terms with not a single concrete fact: no numbers, no system names, no «we rolled that back on a Friday evening».
• Failure at depth. The follow-up «why that way and not the other, and what broke afterwards» ends the conversation in thirty seconds. A model will invent a plausible answer, but not the one that was yours.

The practical conclusion is unexpected: the risk comes not from the tool but from trying to look like somebody else. A prompt listing facts about your own project produces none of these signals — you are saying your own things in your own words. An answer generated on your behalf produces all of them at once.

Where AI belongs in a conversation without caveats

The «assistant during a call» niche was born around interviews, but its dullest scenarios turned out to be the most useful. There is no grey area here and nothing to hide.

How to prepare so the prompt is not needed

The strongest position in an interview is having an assistant and barely using it. That comes from preparation rather than heroism; here is the working version.

1. Collect eight to ten of your own stories. Situation, task, what you personally did, how it ended. With numbers. Ninety per cent of behavioural questions are those same stories from different angles.
2. Say them out loud and then read the transcript. Not re-read what you wrote — read the transcript of your spoken speech. Filler words, broken sentences and forty-second monologues are visible only there; we covered the technique in transcripts as a language tool.
3. Put the facts in one place. Stack, versions, data sizes, deadlines, system names — everything that evaporates under stress. That is the legitimate cheat sheet: the facts are yours, the wording is yours.
4. Run a mock interview with AI as a hostile interviewer. Ask it to dig: «why this and not that», «what broke», «what would you do differently». That is exactly where candidates fall apart.
5. The night before, speak rather than read. The skill tested in an interview is spoken. Reading notes trains reading notes.

What the interviewer should do

The other side of the coin. The «catch them red-handed» race is unwinnable: AI detectors are wrong often, and demanding that nobody look away turns an interview into an acting exam. Something else works — questions a model cannot carry.

The last point is uncomfortable but fair: if an interview task is solved by a model in ten seconds, and at work it will be solved by a model anyway, the problem is with the task rather than with the candidate.

Where Whisperer fits

We build an assistant that listens to the conversation, keeps a transcript and prompts from your own material — and it works on top of any service: Zoom, Meet, Teams, Telegram, Discord. No bot joins the call and none appears in the participant list: audio comes from your computer, which is also why the roster does not change; we covered the three ways to capture a call without a bot separately.

The prompt panel stays out of screen sharing — by the mechanism described above, limits included. We call it a feature for people who lead a conversation and would rather not show the other side their kitchen: notes, quotes from documents, a draft of a thought. It is not an invisibility mode that does the work for you.

What genuinely helps in a conversation where the substance is yours:

• prompts rest on your knowledge base — your documents, notes and facts, not internet generalities;
• a live transcript runs in real time, so you are not reconstructing the conversation from memory afterwards;
• analysis of what is on screen — code, a diagram, a slide — when the discussion is about a picture rather than text;
• a no-logs mode: the conversation is not written to the database at all when the topic calls for it;
• simultaneous voice translation if the other side speaks another language (Max plan);
• 60 minutes a month free — enough to test the scenario rather than a story about it.

And what we do not promise. Whisperer will not pass an interview for you, will not make you an engineer one level up, and takes no responsibility if you break announced rules. We think the honest position pays better for both sides: a product sold as a way to fool an employer lives until the first probation period — and you and I still have to work on Monday.

Frequently asked questions

What to do next

In short: stop choosing between «cheat» and «suffer». The real fork is elsewhere — between looking like a strong candidate for one hour and being one starting Monday.

The easiest place to see how an assistant behaves in a live conversation is not an interview but an ordinary work call: nothing is at stake, and you immediately see whether it keeps up and whether it lands on topic. Create an account — the first 60 minutes each month are free — run your next meeting with it and look at the transcript afterwards. And prepare for the interview so well that you never need the prompt. That, incidentally, is the best way to win it.