AI Interview Cheating: What Each Method Looks Like and How to Detect It

Candidates use AI in AI interviews in three ways: they open it in another tab, they run an invisible overlay on the same screen, or they ask a second phone. Standard browser proctoring catches only the first. The other two need a camera that sits outside the computer. This post walks through what each kind of AI interview cheating looks like, which proctoring layer catches it, and how to read the evidence without rejecting honest candidates.

The pressure is real. In a 2025 Gartner survey of 3,000 job candidates, 6% admitted to interview fraud, meaning they posed as someone else or had someone pose as them. A separate Gartner survey found 39% of candidates had used AI somewhere in the application process. Gartner expects one in four candidate profiles worldwide to be fake by 2028. AI interviews are a first-round screen, so they absorb the bulk of this.

Three ways candidates use AI in an AI interview

Each method leaves a different trace, and that trace decides what you can detect.

Two of the three methods are invisible from inside the browser. That is the core problem with AI interview cheating, and the reason the rest of this post splits detection into two layers.

Layer one: what standard proctoring catches

In-browser proctoring handles method one well. The settings worth turning on for every AI interview:

• Detect switched tab, with a screenshot. When the candidate leaves the interview window, the platform records the switch and captures what they switched to. On Equip, that screenshot is the single most useful piece of evidence, because it shows the ChatGPT window, not just a timestamp.
• Enforce full screen. The interview occupies the whole display, so leaving it is a visible, logged event rather than a glance at a second window.
• Disable copy and paste. The candidate cannot lift the question text out of the interview, which removes the fastest route into an AI tool.
• Face and audio detection. Flags a second face in frame, or voices other than the candidate's.

These measures are standard and most platforms, including Equip, offer them without extra cost. Candidates see which measures are active before they start, which is itself a deterrent. The limitation is architectural: everything above runs inside the browser, so it can only see what the browser can see.

Layer two: the overlay problem

Cluely grew out of Interview Coder, a tool its founders built to pass coding interviews at large tech companies. The pitch is an AI that watches the screen, listens to the audio, and answers in real time through a window that does not appear in screen-share or screen recordings. The company raised $5.3 million on that premise in 2025, and several competitors now sell the same "undetectable" feature.

From inside the browser, a candidate running one of these tools looks clean. No tab switch, no full-screen break, no paste event, one face on camera. The overlay reads the question from the screen and the candidate never has to touch it. A behavioral review can raise suspicion (unusually uniform pauses, eyes tracking across a line of text while speaking), but suspicion is not evidence you can act on.

The only reliable way to see an overlay is to film the screen from outside the computer. That is what Auxiliary Device proctoring does.

Catching what the browser cannot: Auxiliary Device

Auxiliary Device turns the candidate's own phone into a second proctoring camera. The flow, as it runs on Equip:

1. The candidate opens the interview on their computer and scans a QR code with their phone. The two devices pair.
2. They record a short scan of their surroundings, which is uploaded for your review.
3. They position the phone so its camera shows the keyboard, their hands and the screen. Equip captures a frame from that feed every few seconds for the length of the interview.

One camera placed this way closes both invisible methods at once. An overlay window shows up on the screen in the phone's frame, because the phone is photographing the physical display rather than a software capture of it. And the phone that would normally be used to ask ChatGPT is now the proctoring device, pointed at the desk. If it locks or the camera feed stops, the candidate sees a warning on their computer and must fix it before continuing. Pairing is mandatory: the questions do not load until a phone is connected.

On Equip it is free, disabled by default, and switched on per interview under Proctoring Settings. The Enhanced proctoring level turns it on along with the other measures when you set up an interview. The help article on AI usage detection has the exact steps.

Read that second number as the share of cheating you would have missed with standard settings alone. Across proctored AI interviews on Equip, 63% raised at least one flag, and 78% of those flags came only from the Auxiliary Device feed. Everything in that second group looked clean to the browser.

Reading the evidence: the Trust Score

Detection produces evidence. It does not produce a verdict, and treating it as one is the most common way teams get this wrong.

Every proctored AI interview on Equip gets a Trust Score, a 0 to 100% rating in the report header that condenses all proctoring events into one number. Violations lower it; a clean session keeps it high. The practical rule:

A low score is a signal, not proof. Background noise at home, a notification that briefly pulled focus, or a family member walking past the camera all register as violations. So does an accidental tab switch. The score exists to tell you where to spend review time, not to make the decision for you.

When you open a low-scoring report, work through it in this order:

1. Switched-tab screenshots. These settle most cases on their own. A ChatGPT window with the interview question pasted in needs no further interpretation. A calendar notification does not.
2. Auxiliary Device evidence. The surroundings video and the periodic keyboard screenshots. Look for a second device on the desk, another person, or a window on the screen that the browser recording never showed.
3. Session recording. If something looks suspicious but inconclusive, play back the candidate's screen and weigh it against how they performed.

A setup that fits the stakes

Not every interview needs every layer. Candidates tolerate proctoring when it matches the seriousness of the stage, and over-proctoring a casual first screen costs you completions. A reasonable default:

• High-volume first screen (one-way interview): tab-switch detection with screenshots, full screen, copy-paste disabled. Review anything under 80%.
• Shortlist or final-round AI interview (conversational): all of the above plus Auxiliary Device. This is where an offer is on the line and where overlay tools are most likely to appear.
• Technical roles, any stage: Auxiliary Device on.

Tell candidates what is on before they start. Equip shows the active measures on the first screen and asks for consent. Candidates who know the camera sees their desk mostly do not bother. The deterrent effect is cheaper than the detection.

For the broader setup, the guide on how AI interviews work covers the full flow, and the post on one-way vs conversational AI interviews helps decide which format each stage needs. The original reasoning behind the two-camera approach is in why traditional proctoring fails against ChatGPT.

When AI use is the skill, measure it instead

Everything above assumes AI use in the interview is cheating. For a growing number of roles, that assumption is wrong. If the job involves drafting with an LLM, debugging with a coding assistant, or reviewing AI output for errors, then how well a candidate works with AI is the thing you want to see, not the thing you want to block.

The fix is to separate the two deliberately. Proctor the stage where you need the candidate's unaided thinking. Then add a stage where AI use is expected and scored: how they prompt, whether they catch the model's mistakes, whether they know when to stop trusting it. Equip builds custom workflows for exactly this, so you can run an AI-permitted task alongside the proctored interview and score each on its own terms. Ask the Equip team about it if the role calls for it.

Either way, the principle is the same. Decide in advance what you are measuring, make the rules visible to the candidate, and read the evidence before you read the score.