Can interviewers tell if you're using AI? What's actually detectable in 2026

By The Assistly team · June 27, 2026

Every AI interview assistant on the market promises the same thing: it's "undetectable." Some of that is true. A lot of it is marketing. This is a straight, sourced breakdown of what an interviewer or a proctoring tool can actually see in 2026 — so you can make an informed decision instead of trusting a landing page.

How "invisible to screen share" actually works

The real mechanism is mundane, and it's the same one Netflix uses to show a black rectangle when you screenshot a movie. A well-built overlay asks the operating system to exclude its window from the screen-capture pipeline:

• Windows has a window-level display-affinity setting (available since Windows 10 version 2003). The Desktop Window Manager then leaves that window out of every capture surface, so Zoom, Teams, Meet, and OBS record only what's behind it.
• macOS has a per-window content-protection setting. The window is composited to your physical display but withheld from capture buffers.

So the window is genuinely on your screen and genuinely absent from the shared feed. That part isn't snake oil. But "the OS hides the window" is a long way from "nobody can ever tell."

Where the invisibility breaks

The capture-exclusion trick is conditional, not absolute. It can fail in ways most tools don't advertise:

• Newer macOS capture paths. Apple's modern ScreenCaptureKit framework composites all visible windows into a single frame before capture, which can bypass the older content-protection flags. Invisibility on macOS is not a "set it once and forget it" guarantee — it depends on how the tool is built and how well it's kept current with each OS release.
• Sharing the whole screen vs. a window. Exclusion behaves differently depending on whether someone captures an entire display or a single application window, and on the version of the conferencing app.
• Sharing the wrong thing. The most common way people get caught has nothing to do with detection: they share the wrong window, or their audio leaks.

The honest takeaway: invisibility is a property of a well-engineered, well-maintained overlay, not a magic constant. When you evaluate any tool, that engineering — and whether the team keeps up with OS changes — is the whole ballgame.

What video calls can and can't detect

Plain video calls — Zoom, Google Meet, Microsoft Teams — cannot detect a third-party overlay or screen recorder. They run as ordinary user-space apps with no system-level process monitoring. Zoom's own documentation notes it can't even detect third-party recording of a Zoom call. There is no secret "screen-recording-detection API" these platforms call on a participant.

That means in an ordinary conversational interview, a sales call, or a meeting, the technical detection surface is close to zero. The risk is behavioral, not technical.

The behavioral tells

This is what actually gives people away, and detection vendors and interviewers describe the same patterns:

• Flat answer latency. A consistent three-to-five-second pause before every answer, regardless of difficulty, because a model takes roughly the same time on an easy question as a hard one. Humans vary.
• "Reading" eyes. Horizontal, left-to-right eye movement while you're supposedly speaking from memory.
• Answers that sound like documentation. Over-structured, textbook phrasing with none of the natural "yeah, kind of, I mean" texture of real speech.
• Collapsing under follow-ups. The single biggest tell: an answer that outruns your actual ability, so you can't defend it when the interviewer digs in.

Note these are heuristics interviewers use, not a proven detector. But they're why reading polished AI output verbatim, with lag, is the fastest way to fail.

Proctored coding platforms are a different game

Conversational calls are one thing; proctored assessments are stronger. HackerRank, CodeSignal, and lockdown browsers like Proctorio run in-browser or via an installed agent and can flag much more:

• tab and focus switching, copy-paste (with content), and webcam snapshots;
• "unusually linear" typing — complex solutions appearing with no pauses or debugging, a classic paste/AI tell;
• physically connected second monitors (HDMI/extended displays), and, with an installed companion app, virtual machines.

CodeSignal reported that cheating attempt rates on proctored assessments more than doubled to roughly 35% in 2025 — so the platforms are watching closely and investing in catching it. A browser-only proctor still can't see a genuinely separate device, but on a monitored coding test the safe assumption is that copy-paste, tab-switching, and off-screen glancing are all being scored.

The detection layer most tools ignore: the process list

Screen-capture exclusion hides the window. It does nothing about the process. Some proctored platforms (HackerRank and Codility among them) and dedicated monitoring tools scan the running process list — the same thing you see in Activity Monitor on macOS or Task Manager on Windows — against a blocklist of known assistant names. If your tool runs as "LockedIn AI" or "Cluely," a name-based scan can flag it even though the overlay never appeared on screen.

This is exactly why some tools quietly use randomized process names. It's also where Assistly takes a different, simpler approach: you can rename the app and swap its icon to whatever you like. A name-based scan has nothing recognizable to match, because there's no "Assistly" in the list — just whatever you called it. Combined with OS-level screen-capture exclusion, that closes the two most common technical detection paths at once. (It can't change behavioral tells — those are still on you.)

The one case everyone cites

The famous "caught" story is Roy Lee, the Columbia student behind Interview Coder (later Cluely). The detail people skip: he wasn't caught by any detection technology. Amazon's systems didn't flag the tool. He posted a YouTube video of himself using it in an Amazon interview, and two days later someone sent a tip to Columbia. The consequence — a one-year suspension — came from self-publicity and a human report, not anti-cheat software.

That's the pattern across every credible case: exposure comes from human observation, self-disclosure, or failing on the job afterward — not from a real-time overlay detector in the call.

What this means for choosing a tool

If you're going to use real-time AI in your conversations, the things that actually matter are:

1. A real, OS-level invisibility mechanism that's actively maintained across OS updates — not a browser extension that shows up the moment someone asks you to share your full screen.
2. Low latency, so you're not sitting in that tell-tale uniform pause.
3. Guidance you can speak in your own voice — structure and talking points, not a wall of text to read aloud.

This is the bar we built Assistly to clear. The overlay is excluded from screen capture at the OS level on macOS and Windows, guidance streams in real time so you're not waiting on it, and personas tuned to your own background and résumé keep the suggestions sounding like you rather than generic model output. Fully configurable hotkeys let you show, hide, and drive the overlay without ever reaching for the mouse.

A tool can hand you the right point at the right moment. It can't make you sound like yourself or defend a follow-up — that part is still on you. Use it to be prepared and present, within the rules of whatever meeting, call, or interview you're in.

See how Assistly compares to the popular options: vs Cluely, vs Interview Coder, vs Final Round AI, and vs LockedIn AI.