In 2025 a tool built to cheat on technical interviews became one of the most talked-about startups in tech. Its makers, Roy Lee and Neel Shanmugam, were suspended from Columbia University over Interview Coder, an "undetectable" AI tool for coding interviews, according to TechCrunch. They went on to launch Cluely, raising $5.3 million in April 2025 (TechCrunch) and a $15 million Series A led by Andreessen Horowitz that June (TechCrunch).
Whatever you think of the marketing, the effect on hiring is real: a category of tools now exists whose whole selling point is that the interviewer cannot see them.
What these tools do
Cluely and Interview Coder are the best known, alongside tools such as Final Round AI and LockedIn AI. They work in broadly the same way:
- Listen. The tool captures the interviewer's voice from the call audio and turns it into text.
- Ask. It sends the question — sometimes with the candidate's CV or a screenshot of a coding problem — to a large language model.
- Show. The answer appears in a translucent window over the candidate's screen, positioned near the webcam so the eyes don't move far.
The whole loop typically takes a few seconds.
Why they don't show up on screen share
Windows and macOS let an application mark its window as excluded from screen capture. It exists so that private windows — password managers, video players with protected content — stay out of recordings. Invisible interview assistants use the same feature. When the candidate shares their screen, the interviewer sees the coding problem and an empty desktop; the overlay is only on the physical monitor.
This is why "make them share their screen" is not a defence on its own, and why asking a candidate to show their desktop proves nothing.
The patterns they leave
The software can hide from capture. The person using it cannot hide from how reading and relaying an answer look.
- A flat delay. The transcription and model take a few seconds, so every answer starts after a similar pause, whatever the difficulty.
- Stalling phrases. "That's a great question…", repeating the question slowly, or "let me think about that" while the answer loads.
- Reading eyes, close to the camera. Overlays sit near the webcam, so the eyes stay almost on camera — but they move in the short horizontal sweeps of reading.
- Polished structure. Answers arrive as complete, numbered, textbook paragraphs, with no false starts.
- Collapse on a follow-up. A specific, personal follow-up ("why that index and not a composite one?") takes longer than the original question did, or gets a different generic answer.
- Code that arrives whole. A solution appears in large blocks with almost no editing, or is typed at an unnaturally even pace.
How to catch them
See the screen from outside. A second camera — the candidate's phone placed to the side — sees the monitor itself. A window hidden from capture still lights up the screen, and the camera picks it up. It also sees a phone on the desk.
Measure the timing. Compare the delay before simple and hard questions. People answer "Which language do you use most?" instantly. A relayed answer takes as long as everything else.
Track eye movement. Reading has a distinct rhythm: short jumps along a line, then a long jump back. An eye-movement check can flag sustained reading while the head faces forward.
Ask what the model doesn't know. Questions about the candidate's own projects, numbers and decisions; follow-ups that build on what they just said; a quick change of direction mid-answer.
Watch the code pad. Detect text that appears without keystrokes, and ask the candidate to talk through a line they just wrote.
Verify in person. A short final-round check built from the candidate's online answers closes the gap that any remote process leaves.
What about audio-only and voice-mode AI?
Some candidates skip the overlay and run a voice-mode AI assistant on a phone propped below the webcam. Screen share is irrelevant here; the phone hears the question directly. The same timing and eye-movement patterns apply, and a side camera sees the phone. A check for Bluetooth headsets and unusual audio devices helps too.
A fair process still matters
Some employers now allow AI help in parts of a technical round, and that is a legitimate choice. What matters is that the rules are stated to every candidate in advance, that monitoring is disclosed, and that a person — not a tool — decides what any signal means. Read more in our guide to detecting AI cheating in online interviews.
How Oximus Pro handles it
Oximus Pro pairs the candidate's phone as a side camera that sees the monitor and the desk, measures answer timing and eye movement on the candidate's device, detects pasted code and partial screen shares, and puts every signal on a timeline with pictures for your panel. Book a demo and we will run a technical round with you.
Frequently asked questions
Can Zoom, Google Meet or Teams detect Cluely?
Standard meeting software does not detect it. These tools are designed so their window is excluded from screen capture, so the shared screen looks clean. Detection needs a view of the physical screen, such as a second camera, plus behavioural signals like answer timing.
Is Interview Coder really undetectable?
It is marketed that way because its window does not appear in a screen share. The candidate's behaviour still changes: a steady delay before answers, reading eye movements and weak follow-ups. A camera that sees the monitor can also see the overlay.
Can HackerRank or other coding platforms detect AI assistants?
Coding platforms look at signals such as pasted code, typing patterns and tab switches. An overlay that never touches the code editor can avoid some of these, which is why live rounds benefit from a second camera and follow-up questions.
What should I do if I suspect a candidate is using Cluely?
Do not accuse them on a hunch. Ask specific follow-ups about their own work, review the timing and camera evidence after the round, and if the doubt remains, verify in a short in-person or supervised round.
