# What interview platforms can actually see when you use AI

Source: https://www.techinterview.org/post/3233476416/what-interview-platforms-detect-when-you-use-ai/
Updated: 2026-07-19 · techinterview.org

Sixty-one percent of companies now run software that watches for AI use during technical interviews, and 39% have moved more rounds back in person to sidestep the question. Both figures come from Greenhouse's 2026 hiring report, and together they show how far the ground has moved since 2023. The part worth understanding is what that detection software can and can't see, because most of the anxiety around it points at the wrong things.

Take an async assessment on HackerRank, CodeSignal, or Codility. The platform runs in your browser and logs a narrow set of events. It records when your tab loses focus, not which tab or app you moved to. It records paste events and their size. It samples the timing between your keystrokes. If a webcam step is on, it stores snapshots or a short recording. That is close to the entire surface. HackerRank added a browser-activity proctoring layer in mid-2025 that flags navigation away from the test window, and its competitors hook into the same signals.

Notice what that list leaves out. The browser can't see a phone propped next to your keyboard, a second laptop off camera, or a note taped to the wall behind the screen. CoderPad can't reliably tell whether you have a second monitor. Webcam checks catch you looking away, not what you're looking at. The most common way people get outside help in a proctored session is the one the software was never built to catch, which is worth remembering when a vendor sells you on how thorough their surveillance is.

## What actually trips a flag

The strongest tell is not the content of your code. It's the shape of how it arrived. A complete, correct solution that appears after two minutes of no typing, pasted in under two seconds, looks wrong to any keystroke-timing model, whether a human wrote it in another window or an assistant did. A cluster of tab-focus losses right before each function gets solved reads the same way. So does code whose variable naming, structure, and comment voice swing between problems in a way one person's habits usually don't.

None of these prove anything alone. A senior engineer who thinks for two minutes and then types fast will paste from a scratch file and lose focus to read the docs. That's why the better platforms treat a flag as a reason for a human to look, not an automatic reject. The number you should actually care about is the reviewer's read of your session replay, not the automated risk score the system attaches to it. A false positive that a careful reviewer clears costs you nothing. A false positive that a rushed reviewer rubber-stamps costs you the offer, which is the real reason to keep your working rhythm looking like your own.

Which raises the split that decides almost everything about your interview: is a human watching you in real time?

## Live interviews and async assessments are different games

In a CoderPad or Karat session, an interviewer is in the room with you, watching you type. There is barely any automated surveillance because there doesn't need to be. An experienced interviewer catches AI use the old way. They ask why you reached for a hash map instead of sorting first. They change a constraint halfway through and watch whether your explanation bends to fit it. Pasted or prompted code tends to fall apart under a follow-up the candidate didn't anticipate, and that collapse is far more legible to a person than a timing anomaly is to a model. This is also why Karat runs its screens with live human engineers rather than leaning on logging.

Async assessments are the opposite. No human is watching, so the platform fills the gap with instrumentation. That's where tab tracking, paste detection, and keystroke sampling earn their keep, and where the false-positive risk runs highest because a machine is inferring intent from timing alone. The rule of thumb: the less a human is present, the more the tooling watches, and the blunter that watching gets.

| Interview format | Human watching live? | What the platform monitors | What usually triggers a human review |
| --- | --- | --- | --- |
| Async assessment (HackerRank, CodeSignal, Codility) | No | Tab-focus loss, paste size and timing, keystroke cadence, optional webcam snapshots | Fully formed solution pasted after a silent gap; repeated focus loss before each solve |
| Live shared editor (CoderPad) | Yes | Minimal automated monitoring; interviewer sees typing in real time | Answers that don't survive a follow-up or a changed constraint |
| Live human screen (Karat) | Yes, a contracted engineer | Recorded session, human judgment of reasoning and process | Explanation that doesn't match the code, or reasoning that stalls under probing |
| Onsite / in-person | Yes, in the room | None needed | Direct observation of how you work |

## The companies that now hand you the AI on purpose

The louder story of 2026 is that a growing set of teams stopped fighting and rewrote the test. Meta reworked parts of its coding loop to let candidates use an AI assistant during the interview, and grades how you use it rather than whether you touched it. The framing sounds like this: use whatever tools you'd reach for on the job, and talk me through what you ask the model and how you check its answer. The skill on display is no longer recall of an algorithm you could have memorized. It's whether you can steer the model, catch it when it's confidently wrong, and turn a rough first draft into code you'd actually ship.

In that format the questions change shape. A few that show up in AI-allowed rounds:

- "Here's a function the assistant generated. It passes the happy path and breaks on empty input. Figure out why and fix it, out loud."

- "Prompt the model for a rate limiter, then tell me which part of what it gives you that you don't trust, and why."

- "You have 25 minutes and full AI access. Ship a working endpoint plus a test that proves it works."

These reward the person who reads generated code critically and expose the person who pastes and prays. Meta isn't alone here. The shift shows up anywhere the interview loop is trying to predict on-the-job performance rather than closed-book recall, which is most places that have thought hard about it.

## How to play it without getting burned

Read the rules they give you, and take them literally. If an assessment says no external tools, the safe read is to treat it as a closed-book test, because the cost of a flag is not a lower score. It's a withdrawn offer and a note in an applicant-tracking record you'll never get to see or rebut. If the instructions say AI is fine, use it well and say your reasoning out loud so the reviewer sees judgment rather than just output.

When AI is allowed, the behavior that reads as competence is checking the model's work in front of them. Run the code. Point at the line you're not sure about. Fix the bug the assistant introduced instead of hoping nobody notices it. When AI is not allowed and you're on an async platform, understand that pasting a full solution cold is the single most detectable thing you can do, more than any content-matching filter. If you're referencing your own notes, retype rather than paste, and keep your pace looking like a person solving a problem, not one transcribing an answer.

The deeper trap is outsourcing the thinking you'll later be asked to defend. In-person rounds came back for exactly this reason. If a company runs an easy remote screen and then a hard onsite where the same skills should reappear, an AI-assisted remote pass followed by an onsite that falls apart is a worse outcome than a clean remote miss. The gap between the two performances is the thing that gets remembered.

The detection arms race will keep escalating, and the invisible-copilot vendors will keep swearing they've beaten it this quarter. The more durable bet is the boring one: get good enough that you don't need to hide what you're doing, and aim for the interviews where using the tool is the job. Those are multiplying a lot faster than the ones built to catch you.