Cluely and Parakeet AI: How Recruiters Can Spot AI Cheating in Interviews
A third of recruiters now say they see AI used live during interviews, according to Greenhouse. The tools behind it are easy to buy, quick to install and built not to show up on a shared screen. Two of the best-known are Cluely and Parakeet AI.
This guide is for the person on the other side of the call. It explains what these tools do, why screen sharing won’t reveal them, the signs interviewers notice, and the questions that make hidden help useless. It ends with a short AI policy you can paste into your interview invitations.
We make Realtimer, a real-time AI assistant for sales, property and recruiting calls, and our terms forbid using it to cheat in exams or job interviews. Facts about other products come from their own websites and from named reviews, linked in the text. Checked October 2026.
What Cluely and Parakeet AI do
Cluely
Cluely grew out of Interview Coder, a tool its founders built at Columbia University to help candidates in technical interviews; we tell that story in our guide to AI cheating in job interviews. Today its website presents it as a meeting assistant: a desktop app that listens to the conversation, answers when you press a shortcut, can read what’s on your screen, and writes notes afterwards. It works alongside Zoom, Google Meet, Microsoft Teams and Webex.
The same page calls it “completely undetectable” and says it never shows up in shared screens or recordings.
Parakeet AI
Parakeet AI is built for job seekers. Its website offers a call assistant that gives real-time answers during interviews, alongside a resume editor and mock interviews, with models from OpenAI, Anthropic and Google. It runs on desktop, in the browser or on a phone, and it’s paid with credits per session or a subscription.
Reviews describe it working with Zoom, Meet, Teams, Webex and coding platforms such as HackerRank and LeetCode, and say its marketing claims it stays hidden from screen sharing (Toolradar, InterviewCoder review).
Why screen sharing won’t show them
Both macOS and Windows let an app mark a window as excluded from screen capture. Screen sharing, recording and screenshots skip it, so the candidate can share their whole screen and you’ll see a clean desktop with the answer still in front of them. Presentation tools use the same feature to keep speaker notes private.
The other route needs no software trick at all: a phone or tablet propped under the webcam, running an assistant that listens through its own microphone. Nothing on the computer changes.
That leaves what you can hear in their answers.
Eight signs interviewers notice
No single sign proves anything. People pause, look away and read notes for innocent reasons, and nervous candidates, non-native speakers and neurodivergent people can show several of these. What experienced interviewers notice is a pattern across the whole conversation. Amazon’s internal guidance, as reported by ITPro, lists some of the same signals.
| Sign | What it looks like |
|---|---|
| The same pause before every answer | A steady two or three seconds before each reply, even to “how are you?”, followed by fluent, complete sentences. People think at different speeds for different questions; a tool takes about the same time for all of them. |
| Eyes that read | Small left-to-right movements, usually just off the camera, while the candidate talks. Reading looks different from remembering: remembering looks up or away, reading tracks a line. |
| Answers shaped like a list | “There are three key factors here: first…” for every question, with headings in the spoken answer. Real answers wander, correct themselves and skip steps. |
| Textbook words, none of their own | The vocabulary of the topic without the details of the candidate’s life: no system names, no colleague’s role, no number they remember. |
| Strong first answers, weak follow-ups | The opening answer is polished; the second question about the same story gets vague, or the candidate restates the first answer in different words. |
| Repeating the question back | “So you’re asking how I would handle…” before most answers. It buys time for transcription and generation, and it gives the tool a cleaner version of the question. |
| Typing while you talk | Keyboard sounds during your question, or hands moving below the camera, when nothing on the call needs typing. |
| Fluency that doesn’t match the CV | A junior candidate explaining architecture trade-offs like a principal engineer, then struggling to describe their own last project. |
Seven questions and tests that make hidden help useless
Aim for an interview where the answers can only come from the candidate’s own head. It’s also the interview that tells you the most about them.
- The follow-up ladder. Pick one project from the CV and ask three levels down: what you built, what went wrong, what you would change. Each step needs memory a model doesn’t have.
- Interrupt and redirect. Halfway through an answer, ask a narrower question: “Wait, who decided that?” A person adjusts mid-sentence. A candidate reading a generated answer often loses the thread or starts reading the next block.
- Ask for the number. “What was the conversion rate before and after?” People who did the work remember the number, or remember why there wasn’t one. A model invents a round one.
- Change the audience. “Explain that as you would to a new hire on your team.” Rewording on the spot for a different listener is hard to prompt quickly.
- Change the constraints. In technical rounds, change a requirement halfway: a smaller budget, a missing dependency, ten times the traffic. Ask what breaks first.
- Work in a shared document. Ask the candidate to sketch an outline or a query in a shared doc while talking it through. You see the thinking happen, including the false starts.
- Ask for an opinion they can’t look up. “What’s something in your field most people get wrong?” or “What on your CV would you remove now?” There’s no standard answer for a model to reach for.
All of these depend on memory, judgement and changing course mid-answer. A language model is good at a clean first answer to a standard question and bad at all three.
What not to do
Don’t accuse anyone on the call. A wrong accusation is unfair to the candidate, and in some places a legal risk for you.
Don’t rely on an AI detector alone. They produce false positives, and the people most often wrongly flagged are the ones who already find interviews hardest.
Recording the call to review it later needs consent: many jurisdictions require everyone’s agreement before a call is recorded or transcribed.
Whatever extra step you add because of a hunch, add it for everyone in the round, not only for the candidate you suspect.
If you suspect a candidate is using AI
Move to the tests above and stay friendly. If your policy allows it, ask the candidate to share their entire screen or show their desk briefly, and say you ask everyone. Then, after the call, write down what you observed rather than what you concluded: “paused about four seconds before each answer; couldn’t say what the conversion rate was after the redesign listed on the CV.”
If the role matters, add a live round: in person, or on video working in a shared document. Google and Deloitte UK both brought back in-person rounds for some roles, and for remote hiring a live working session gets you most of the way there.
An AI policy you can paste into interview invitations
Amazon’s rules work partly because candidates confirm them in advance. A short statement tells honest candidates what’s allowed, gives you a fair basis for decisions, and means nobody can say they weren’t told. Adapt this one:
Using AI in our interviews. You’re welcome to use AI to prepare: practise questions, research the company and polish your CV, as long as everything in it is true. During live interviews, please answer without AI assistants, on this device or any other; we want to hear how you think. For take-home tasks, we say in the task whether AI tools are allowed, and we ask you to tell us which ones you used. If anything is unclear, ask us before the interview.
If your own process uses AI, for example for scheduling, note-taking or screening, say so in the same message. It’s increasingly a legal requirement, and it’s hard to ask candidates for openness you don’t offer.
Spotting AI in interviews: FAQ
Can you detect Cluely or Parakeet AI through screen sharing?
Usually not. Tools like these use windows the operating system excludes from screen capture, or run on a second device, so a shared screen looks clean. Detection depends on how the candidate behaves and answers, not on what you see on their screen.
Are AI detection tools reliable?
Not on their own. They produce false positives, and a wrong flag can end a real candidate’s application. Use them, if at all, as one input next to better questions and a clear process.
What should I do if I suspect a candidate is using AI?
Don’t accuse anyone on the call. Move to questions that need the candidate’s own memory, write down what you observed (not what you concluded), and add a live or in-person round if the role matters. Apply the same steps to every candidate.
Is it legal to reject a candidate for using AI in an interview?
Generally yes, if your rules said AI help wasn’t allowed and the candidate agreed to them. Base the decision on your process and documented observations, not on a guess, and check local employment law where you hire.
Should we ask candidates to share their whole screen?
It helps against open browser tabs, but not against windows excluded from capture or a second device. If you ask, ask everyone, explain why, and keep it short; it’s less useful than good follow-up questions.
Do nervous candidates show the same signs?
Some of them, yes. Pauses, looking away and reading notes all have innocent explanations, especially for non-native speakers and neurodivergent candidates. That’s why one sign proves nothing and the response is a better question, not a verdict.
Are in-person interviews the only real fix?
No, but they work. Google and Deloitte UK reintroduced in-person rounds for some roles. For remote hiring, interviews built on the candidate’s own experience and live work in shared documents get you most of the way.
Help on your side of the call
If you run interviews and want quick, accurate answers to candidates’ questions about the role, salary band and next steps, from your own documents, with notes after the call, see Realtimer for recruiters.