How Candidates Cheat AI Interviews in 2026, and How to Detect Each Method

How do candidates cheat AI interviews in 2026?

Four ways: an answer overlay that screen sharing cannot see, a second device or an off-camera helper, a proxy or deepfake sitting the interview for someone else, and an AI copilot that hears the question and drafts the answer. Most proctoring still watches the browser tab and the webcam, which misses three of the four.

The numbers are large enough to plan for. In Gartner's 2025 candidate surveys, 6% of candidates admitted to interview fraud, either posing as someone else or having someone pose as them, and Gartner predicts one in four candidate profiles worldwide will be fake by 2028. The tools are sold openly: Cluely's homepage says it 'never shows up in shared screens, recordings, or external meeting tools'.

The proxy problem has moved from theory to law enforcement. The FBI warned in 2022 of deepfakes and stolen identities used to apply for remote jobs, and in June 2025 the Justice Department announced coordinated action against North Korean IT worker schemes that had obtained work at more than 100 US companies using stolen and fake identities.

Each method needs a different detector, so the sections below take them one at a time. The signal streams and the questions to ask a vendor are in the AI interview proctoring guide.

How do answer overlays hide from screen sharing, and how do you detect them?

On Windows, an app can mark its window as excluded from capture, so it shows on the candidate's monitor and nowhere in a screen share or recording. A desktop proctoring app can list every window and read that flag. A browser-based proctor cannot, which is why overlays defeat most web tools.

The mechanism is a documented Windows API. SetWindowDisplayAffinity with WDA_EXCLUDEFROMCAPTURE, available since Windows 10 version 2004, shows a window on the monitor while it does not appear at all anywhere else. Microsoft's own example is keeping recording controls out of recordings; the same call hides an answer panel. On macOS the picture is less clean: Apple now calls the equivalent window setting a legacy constant that macOS no longer uses and tells developers not to rely on it to hide content, so results vary with the capture method.

Detection is the mirror image. The companion call, GetWindowDisplayAffinity, reads that setting for a window from any process, so a desktop app can enumerate visible windows and flag any that are excluded from capture, with the process that owns each one. Legitimate software can set the same flag, so a hit is evidence for review, not a verdict.

Behaviour backs the check up. Reading from an overlay tends to show as a pause, then fluent and evenly paced recitation, then trouble when the interviewer asks one level deeper about the answer just given. That is why conversation design, covered below, is the layer that still works when the environment check has been bypassed.

# Windows only: list visible windows that are hidden from screen capture.
import ctypes
from ctypes import wintypes

user32 = ctypes.WinDLL("user32", use_last_error=True)
WDA_MONITOR, WDA_EXCLUDEFROMCAPTURE = 0x01, 0x11

EnumWindowsProc = ctypes.WINFUNCTYPE(wintypes.BOOL, wintypes.HWND, wintypes.LPARAM)
user32.EnumWindows.argtypes = [EnumWindowsProc, wintypes.LPARAM]
user32.IsWindowVisible.argtypes = [wintypes.HWND]
user32.GetWindowDisplayAffinity.argtypes = [wintypes.HWND, ctypes.POINTER(wintypes.DWORD)]
user32.GetWindowThreadProcessId.argtypes = [wintypes.HWND, ctypes.POINTER(wintypes.DWORD)]
user32.GetWindowTextW.argtypes = [wintypes.HWND, wintypes.LPWSTR, ctypes.c_int]


def windows_hidden_from_capture():
    flagged = []

    def check(hwnd, _lparam):
        affinity = wintypes.DWORD()
        if user32.IsWindowVisible(hwnd) and user32.GetWindowDisplayAffinity(
            hwnd, ctypes.byref(affinity)
        ):
            if affinity.value in (WDA_MONITOR, WDA_EXCLUDEFROMCAPTURE):
                pid = wintypes.DWORD()
                user32.GetWindowThreadProcessId(hwnd, ctypes.byref(pid))
                title = ctypes.create_unicode_buffer(256)
                user32.GetWindowTextW(hwnd, title, 256)
                flagged.append(
                    {"pid": pid.value, "title": title.value, "affinity": hex(affinity.value)}
                )
        return True

    user32.EnumWindows(EnumWindowsProc(check), 0)
    return flagged  # evidence for a reviewer, never an automatic rejection

How do you catch a second device or an off-camera helper?

Through audio, not video. Every relay needs to hear the question, so deliver questions only through headphones and verify them with an echo test: play a tone and check whether the microphone picks it up. A phone on the desk or a helper across the room then hears half a conversation, which is useless to them.

Make the check continuous. Enumerate audio outputs at the start, run the echo test, then repeat it quietly mid-session with a near-inaudible tone, so a candidate cannot pass the check and then switch to speakers. Add diarization to flag a second voice or whispering in the room.

The one workaround is to repeat the question aloud so the relay can hear it, which is itself a signal: flag verbatim restatement of a question before an answer. Second monitors are an environment problem. A desktop app can count displays, and in browsers that support the Window Management API, Screen.isExtended reports multiple screens, though it is experimental, missing from several major browsers and reads false when the window-management permission is blocked.

Treat gaze as weak evidence. Looking away while thinking is normal, and candidates with a disclosed accommodation should have visual signals switched off entirely. A glance pattern is a reason to watch the clip, not a finding.

How do you spot a proxy or deepfake interviewee?

Verify identity at every stage, not once. Check a government ID with a liveness test before the interview, match the same face at assessment, interview and onboarding, and detect virtual cameras on the device. The FBI's own advice is identity verification during interviewing, onboarding and throughout employment.

The FBI has described the tells: lip movements that do not match the audio, and coughs or sneezes heard but not seen. In January 2025 it reported North Korean IT workers using AI and face-swapping technology during video interviews, and recommended cross-checking applicants who share resume content or contact details, and completing as much of hiring as possible in person.

Liveness checks ask for spontaneous actions, such as turning the head or holding the ID beside the face, because a pre-recorded or synthesised feed has to produce them live. Run them with consent, as a stated step, not as a hidden test a candidate can fail without knowing.

Placement matters more than cleverness. Identity checks belong before the interview and again before an offer, as their own step with their own consent screen. A harder interview does not catch a stand-in who knows the material, and a laptop shipped to an unverified address undoes every check before it.

How do you detect AI copilots that listen and answer in real time?

Mostly by conversation design, because a copilot on the same machine can capture system audio even when the candidate wears headphones. It still needs time to transcribe, generate and be read, so questions that pivot on the candidate's own last answer, interruptions and requests to explain their own code expose the lag.

Design for latency. Ask for answers to begin within a few seconds, interrupt mid-answer with 'why not the other approach?', and ask the candidate to reconcile something they said two turns earlier. A relay is always one turn behind and cannot follow a conversation that keeps pivoting on its own content.

In coding rounds, have the candidate change their own code under a new constraint and explain why. Pasted or dictated code often comes with a confident account of the general approach and a weak account of the specific line the interviewer points at, which is exactly the gap to probe.

AI interviewers add one more attack: text aimed at the scoring model. Greenhouse's November 2025 survey found 41% of job seekers admit using prompt injections. Treat every transcript and document as data, never as instructions, and keep the rubric out of the candidate's reach. The architecture is in how to build an AI interviewer.

What should you never do when detecting cheating?

Never infer emotion, never auto-reject, and never collect biometrics without consent. The EU AI Act has banned emotion recognition in workplace settings, which includes hiring, since 2 February 2025; GDPR restricts decisions based solely on automated processing; and Illinois BIPA requires a written release before collecting face geometry or voiceprints.

Emotion inference fails on science as well as law. A 2019 review in Psychological Science in the Public Interest found that the way people express emotion varies substantially across cultures, situations and individuals, and that a scowl often communicates something other than an emotional state. Nervousness is not a cheating signal, and in the EU, inferring it from a face or voice at work is prohibited under Article 5(1)(f).

Identity checks are biometric processing. Under Illinois BIPA, a face scan or voiceprint needs written notice of the purpose and retention period and a written release, with liquidated damages of $1,000 per negligent and $5,000 per intentional or reckless violation; a 2024 amendment limits repeated identical violations to one recovery per person. Under GDPR Article 9, biometric data used to identify someone is a special category that generally needs explicit consent.

Keep a person on every decision. GDPR Article 22 restricts solely automated decisions with significant effects, NYC Local Law 144 requires an independent bias audit before an automated employment decision tool is used, and the EU AI Act lists recruitment AI as high-risk under Annex III, with those obligations applying from 2 December 2027 under the Digital Omnibus.

Should you go back to in-person interviews to stop cheating?

For the final round of a remote role with access to sensitive systems, often yes; for every round, no. The FBI recommends completing as much of hiring as possible in person, and in Gartner's surveys 62% of candidates said they were more likely to apply where in-person interviews are required. The cost is reach, travel and time.

The practical split is by risk. Screening rounds stay remote and layered, because that is where the volume sits and where an AI interviewer saves the most time. The last conversation before an offer moves in person, or to a supervised location, for roles that touch production systems, customer data or money, which are the kinds of roles the FBI's 2022 warning named.

Where in person is impossible, a verified video call with the identity checks above, a live change to the candidate's own earlier work and a second interviewer gets most of the benefit. What does not work is a longer remote interview with the same controls: a relay does not get tired.

What does a layered defence look like in practice?

Five layers, each closing a class of attack the others miss: identity before and during the process, a desktop environment check, headphone-only audio verified by an echo test, a conversation designed to punish latency, and a person reviewing timestamped evidence. Skip any one and a known method walks through the gap.

The layers are not equally expensive. Conversation design costs nothing but interviewer discipline and catches the widest range of methods. Headphone enforcement is a small piece of client code. The desktop app is the heaviest lift and asks the most of candidates, so reserve it for high-stakes rounds. Identity checks belong at the edges of the process, where a stand-in is cheapest to stop.

I built AccioMatrix, an AI assessment and interview platform that now serves 20+ enterprise clients, so this is the problem I work on. The integrity result I can point to is published: a Retell AI case study reports 70% fewer false-positive assessments, from 50% to 15%.

Price the layers before you build them. The desktop app, the identity step and the review tooling are separate lines, and the AI product cost estimator shows what each adds to the build and to the monthly bill.

How do you keep false positives down?

Score evidence, not suspicion. Every signal becomes a timestamped clip with a stated reason, several weak signals must agree before a flag, visual signals are switched off for disclosed accommodations, and a person decides. A system that flags a large share of honest candidates is not rigorous; it is unusable, and it trains recruiters to ignore it.

Calibrate on real sessions before launch. Have two reviewers label a sample blind, measure how often the system agrees with them, and raise thresholds wherever it flags innocent behaviour. Then tell candidates what is checked and why, before the interview starts; in Illinois and the EU that disclosure is a legal duty, not a courtesy.

If you want the layered checks built into your own interview flow, an AI interviewer is built at $0: the work is split into checkpoints with acceptance criteria agreed before it starts, and each one is invoiced only after you have seen it and accepted it. If you would rather hand over the whole hiring drive, that is engineering hiring.