How Detectable Is LastRound AI, Really? We Checked the Mechanism

You’re two minutes from hitting “Share Screen” on a Google Meet coding round. LastRound’s overlay is sitting on the right third of your monitor, already showing a cleaned-up version of the question your interviewer just read out loud. Whether the tool works isn’t really in question at that point. What you actually want to know, cursor hovering over the share button, is whether the person on the other end of the call sees that panel too.

We built the product, so “is LastRound AI detectable” isn’t a question we get to answer from some neutral corner of the internet. What we can do instead is show the actual mechanism: the specific line of code that hides the window, the operating-system quirk that can undo it, and the much larger set of detection methods that have nothing to do with your screen at all. Nothing below is a lab score we ran and are reporting back to you. It’s the plumbing, with links to primary sources, so you can check the claim yourself instead of taking a product page’s word for it.

What “Invisible” Actually Means for This Product

LastRound ships its AI Interview Copilot two ways: as a browser tab at app.lastroundai.com, and as a downloadable desktop app for macOS and Windows. Only one of those two actually disappears from a shared screen. Our own help center says so directly. The web-based version, in its own words, “is visible when you share your screen,” while the desktop app is described as “completely invisible during screen sharing.” The product page’s headline claim, “100% Undetectable,” describes the desktop build specifically. That split isn’t always obvious on a first read of the marketing copy, and that’s fair to hold against the copy as much as the code underneath it.

The distinction matters more than it sounds. If you’re doing a quick behavioral round from a work laptop where IT won’t let you install anything, you’re almost certainly on the web app, and the honest answer to “is LastRound AI detectable” in that setup is: yes, trivially, the moment you share your whole screen. The desktop app is the one actually built to solve the problem this post is about.

How Does the Overlay Actually Stay Off the Shared Screen?

It comes down to one function call, not some proprietary trick. The desktop app calls Electron’s setContentProtection(true) on its own window, and the operating system takes it from there, excluding that specific window from anything reading the screen: a video call’s capture engine, a browser’s screen-share API, even the OS’s own screenshot tool. Electron’s own documentation for that method is more specific about what happens on each platform than most product pages bother to be.

On Windows, the call maps to SetWindowDisplayAffinity with the value WDA_EXCLUDEFROMCAPTURE. Microsoft’s own API reference describes the intended use case almost too plainly: “One use for this affinity is for windows that show video recording controls, so that the controls are not included in the capture.” That value only fully works on Windows 10 version 2004 and later. Older builds fall back to painting a black rectangle where the window would be. Which is worse, oddly. A black box announces itself.

On macOS, the same call sets the window’s sharing type to NSWindowSharingNone, an older API that predates the current wave of “invisible” interview tools by years. LastRound didn’t invent it, and neither did Cluely. Every Electron-based overlay in this category is standing on the same two OS primitives, dressed up in different marketing language.

The macOS Wrinkle Worth Knowing About

Here’s the part most product pages leave out entirely. Electron’s own documentation carries a specific warning: because of an intentional change Apple made to macOS, newer Mac apps built on Apple’s ScreenCaptureKit framework will capture a content-protected window anyway, despite setContentProtection(true) being turned on. In plain terms, the exact protection this whole category relies on has a known, documented soft spot on the Mac side, one Apple built on purpose, not a bug waiting on a patch.

We haven’t found a mainstream interview or conferencing platform that has actually moved its screen-share path onto ScreenCaptureKit in a way that exploits this. The major video call apps and browser-based assessment tools still capture through the older pipeline this protection was designed for. So today it’s closer to a fault line than an active leak, but it’s exactly the kind of thing one macOS point release could change, which is a strange thing to build a stealth feature on top of. We’re saying that plainly instead of burying it in a footnote.

What Actually Makes an AI Copilot Detectable (It’s Rarely the Screen)

Assume, for a second, that the window-hiding trick above works perfectly, on every platform, every OS version, forever. That still leaves most of how people actually get caught untouched, because most of it was never about what’s on your screen.

CNBC’s reporting on the Interview Coder story (the tool that later rebranded as Cluely, after founder Chungin “Roy” Lee raised $5.3 million in seed funding from Abstract Ventures and Susa Ventures) is full of interviewers describing exactly this. Anna Spearman, founder of the recruiting agency Techie Staffing, told CNBC what tips her off: “I’ll hear a pause, then ‘Hmm,’ and all of a sudden, it’s the perfect answer.” Henry Kirk, a software developer and co-founder of Studio.init, said the old tell (eyes darting to a second screen) isn’t reliable anymore because the tools got better at putting answers somewhere a candidate doesn’t have to look away to read. He ran a coding challenge that June with 700 applicants and recorded the sessions to check.

“More than 50% of them cheated,” Kirk told CNBC.

None of that is about screen capture. It’s about pacing, phrasing, and a recorded video an employer can rewatch later at half speed. A perfectly invisible overlay does nothing to fix a two-second lag between question and answer that sounds rehearsed.

The tooling on the other side has caught up too, and most of it skips the screen entirely. CodeSignal’s own writeup of its assessment platform describes a “Suspicion Score.” It is built from “telemetry from our advanced work simulations to reliably identify activities correlated with use of AI assistance, such as unusual typing or speaking patterns.” The same system also tracks “when candidates copy and paste text from another window, and what was copied.” CoderPad, used by companies like Karat for live pairing rounds, runs a similar playbook: paste tracking, full code playback so a reviewer can watch how a solution actually got typed, and alerts when a candidate exits and re-enters the test window. Neither system cares whether an extra window exists on your monitor. Both care about your keystrokes and your clipboard.

Then there’s the newest layer, aimed specifically at tools in LastRound’s own category. Validia built a free product called Truely in direct response to Cluely: the interviewer asks the candidate to download and run it, on a Mac running macOS 10.15 or later, during the call itself, and it watches the candidate’s own machine, not the shared screen, for signs of a hidden assistant. If an employer sends you a Truely link before your interview, whatever the desktop app is or isn’t hiding from Zoom stops being the relevant question.

What LastRound Itself Won’t Promise You

Our own desktop setup guide carries a warning most vendors would quietly cut from the copy: “This application has not been tested on all coding platforms. Some platforms may detect screenshot activity or external tools and could flag your session or deduct credits during assessments. Always verify that your interview platform permits the use of external assistance tools before starting your session.” That’s not a competitor’s claim about us. It’s our own help center, written by the team that ships the product.

We think that line actually undersells the point. Read as written, it sounds like a compatibility footnote, the kind of thing you’d see under “browser support.” What it’s really describing is that this whole category, LastRound included, is running an arms race against every video platform, proctoring vendor, and now dedicated detection startup that ships an update. Nobody in this space gets to call that race finished, and any product page that implies otherwise (ours has, at times) is overselling.

How This Actually Gets Checked, Platform by Platform

LastRound’s engineering team maintains a per-platform verification process instead of a single “works everywhere” claim: an automated daily check that drives a real, unmodified system Chrome through the exact screen-capture path that Google Meet, HackerRank, LeetCode, and CodeSignal all use, plus manual checks on native apps like Zoom, Microsoft Teams, and Webex, which ship their own capture engines and can’t be tested the same automated way. Current status per platform is published on a status page rather than baked into one static claim on the marketing site.

We’re not going to print a “verified on X of Y platforms” number in this post. That number moves every time a browser updates or a platform changes its capture pipeline, and the point of publishing it separately, at a URL you can check yourself, is so you get today’s answer instead of trusting whatever was true on the day this went live.

None of this applies to practice interviews, worth saying plainly: a mock session isn’t shared with anyone, so there’s nothing to hide from in the first place. Stealth only becomes a question the moment a real interviewer, or a proctoring tool, is watching the other end of the call. If you want a closer look at how the wider “invisible AI tool” category is racing against its own detectors, we went deeper on that arms race, Cluely included, in our Cluely vs. LastRound AI comparison.

How long a session actually lasts, which matters here

Detection risk scales with exposure, so the length of a real session is a relevant number. Ours is public now.

Across 509 LastRound sessions between 21 March and 30 July 2026, the median ran 3 minutes 9 seconds. The mean ran 8 minutes 28 seconds. The average session covered 6.8 question-and-answer exchanges.

That’s a much smaller surface than the “AI copilot whispering through a 45-minute interview” picture the detection debate usually assumes. Most sessions are somebody checking two or three answers.

None of that makes the tool undetectable, and this post has been careful not to claim that. It does mean the exposure window in typical use is minutes, not the length of the interview.

FAQ

Does LastRound AI show up if I share my whole screen instead of just one window?

Yes, on the web app. No, on the desktop app, assuming a Windows build from 2020 or later or a Mac that hasn’t hit the ScreenCaptureKit edge case described above. Entire-screen share and single-window share both route through the same OS capture pipeline that setContentProtection excludes the desktop app’s window from, so the sharing mode itself doesn’t change the answer the way most people assume it does.

Can a proctoring tool like CodeSignal or HackerRank tell I’m using LastRound AI?

Not from the screen capture itself, based on how those platforms describe their own systems. But both track typing patterns, pasted text, and tab or window switches, none of which content protection touches at all. If your answers arrive at a suspiciously even pace or land word for word after a pause, that’s a separate risk the window-hiding technology was never built to solve.

So, is LastRound AI detectable? On a plain screen share, through the actual capture pipeline, the honest answer today is: not on the desktop app, not by the mechanism most people picture when they ask the question. On your pacing, your clipboard, or a dedicated tool like Truely running on your own machine, invisibility on screen was never the thing keeping you safe, and no vendor, us included, should tell you otherwise.