How to detect AI-coached candidates (2026)

What is interview anti-cheat (and why 2026 is different)?

Interview anti-cheat is the set of signals + interview-design choices used to detect AI-coached candidates without sliding into invasive proctoring. The 2025 threat model exploded: ChatGPT-on-second-monitor, AirPods-fed answers, deepfake video, pre-recorded prompts. The 2026 answer isn't more lockdown software — those get bypassed in days. The answer is layered detection (response-latency patterns, linguistic fingerprinting, CV cross-reference, retake fingerprinting) paired with interview questions that require lived experience the candidate can't fake in real time. This guide is the operator playbook for both halves: which detection signals actually work, which are theater, and how to redesign interview questions to make cheating not worth the effort.

The state of cheating in 2026

2025 was the year cheating became trivial. Second monitor with ChatGPT open. AirPods with a friend feeding answers. Pre-recorded video for async assessments. By Q4 2025, recruiters at Anthropic, Google, and Stripe were all reporting double-digit-percent cheating rates on technical screens.

The vendors responded. Some well, some badly. This is a quick guide to the signals that actually work.

Signals that work

Response latency patterns. A clear question, a 6-second pause, and then a perfect, structured 60-second answer is a fingerprint. Real humans answer with hesitations, false starts, "uhh let me think about that for a second." The pause-then-perfect pattern is coaching or AI assistance. Score it.

Linguistic fingerprinting. LLMs have characteristic patterns: certain transition phrases, certain sentence structures, a tendency toward 3-bullet-list answers. Calibrated detection works — false positive rates under 5% at well-tuned thresholds. The Stanford CRFM 2024 study on LLM-text detection found per-task accuracy of 89-94% on standardized prompts when the detector is tuned to the domain.

CV cross-reference. The candidate claims 5 years of Python — but in the actual interview, when asked about a specific scenario, they describe a Python pattern that didn't exist in the language until 2 years ago. That's a tell. Cross-referencing answers against the CV catches inflated claims.

Eye-movement and gaze. Frequent off-camera glances at consistent timing (right before an answer) suggest reading from a second screen or another person. This is the highest false-positive risk signal — neurodivergent candidates and candidates with anxiety have different eye-movement patterns. Use it as supplementary, not load-bearing. The EEOC's 2024 ADA guidance on AI hiring is explicit that disability-correlated behavioral signals cannot be the basis of rejection without accommodation.

Retake fingerprinting. If you allow retakes, fingerprint the audio across attempts. If retake 3 is significantly more polished than retakes 1-2, the candidate workshopped the answer with help.

Signals that look impressive but don't work

Browser tab detection. Vendors that promise "we detect when the candidate switches tabs" sound reassuring. They're easily fooled (use a second device entirely). And they're hostile to candidate experience — most candidates legitimately have a CV open in another tab.

Webcam-only proctoring. Forcing the candidate to film their workspace is wildly invasive, hurts completion rates, and doesn't catch the candidate who set up the cheating in advance. Greenhouse's 2026 candidate report found webcam proctoring is the single largest driver of abandonment in async-video interviews.

Hard "AI-detected" plagiarism scores. AI detection on prose has near-zero accuracy. Don't use it as a primary signal. The HBR primer on AI hiring ethics explicitly warns against single-signal automation in adverse decisions.

How to design interviews AI can't fake answers to

The most important shift: ask questions where the answer requires real lived experience.

Bad question (easy to fake): "How would you handle a difficult customer?"

Better question (harder to fake): "Walk me through the last difficult customer call you actually took. What was the customer's name or company, what was the issue, what did you say, and what happened next?"

The "actually took" + "what was the name" + "what happened next" turns this from an answer-generation problem (easy for ChatGPT) into a memory-retrieval problem (impossible to fake without a real story).

Apply this pattern across the whole interview. Move from hypothetical to specific. From general to particular. From "tell me about a time" to "tell me about the last time."

What to tell candidates

If you're using anti-cheat scoring, disclose it. "This interview is monitored for response patterns to ensure interview integrity. We do not use webcam proctoring or browser monitoring." This is honest, calibrates expectations, signals you take this seriously, and helps satisfy NYC Local Law 144's notice requirements if you hire in NYC.

When the score is high — what then?

Don't auto-reject. Anti-cheat scoring is informational. A high score means:

1. You probably want to do a live human follow-up
2. You ask follow-up questions specifically designed to surface real experience
3. If the candidate aces the follow-up, you've found a great candidate (and the original score was a false positive)
4. If the candidate falls apart on the follow-up, you've saved your team from a bad hire

The point of the score is to flag where to dig deeper, not to make the decision for you.

Try it

Read Anti-cheat scoring for how Raffi specifically implements this. Or try a demo interview yourself — 5 minutes, no signup, see exactly what the scoring looks like.