Halloween to Cyber Monday sale: 30% off your first year of LMS Advisor, cloud or self-hosted.30% off your first year Ends Dec 1, 2026 Claim the offer
Online Proctoring

How to prevent AI cheating on certification exams in 2026

Invisible overlays, smart glasses and remote-control proxies slip past webcams. Here is the five-layer plan certification programs use to catch what proctoring misses.

AI cheating on certification exams cover illustration: a shield with an alert symbol over an online exam

To prevent AI cheating on certification exams, stop expecting a webcam to do the whole job. Build five layers: exam design that makes a chatbot or a leaked answer less useful, identity and device controls, in-session monitoring, data forensics after the exam, and a trained person who reviews the evidence and decides. Each layer covers blind spots the others have, and none of them works alone.

In LMS Advisor, the Test Center covers most of the software side in one place: question pools with difficulty levels, email one-time codes and ID photo capture, per-exam proctoring settings from fullscreen lockdown to virtual machine detection, and an evidence trail a reviewer uses to decide. The pressure is real and measurable. Meazure Learning reports that in 2025 its security team intervened in nearly 50,000 exam sessions and shut down 4,700 remote-control proxy sessions. Caveon's secret shopping from 2024-25 went further and found that more than 90 percent of attempted cheating or content theft went undetected by proctors. Below is the playbook we'd put in place for a certification program today, including the parts no vendor can see.

What does AI cheating on certification exams look like now?

"AI cheating" covers several different attacks, and they need different defenses. Open edX's March 2025 overview of modern cheating techniques lists AI answer generators, smart wearables, remote screen sharing, deepfake impersonation and browser manipulation. In practice, this is what exam owners run into:

ThreatHow it worksWhat can catch it
Chatbot on the same computerCandidate copies a question into ChatGPT or a similar tool in another tab or windowFullscreen lockdown, tab-switch and paste logs, screen snapshots
Chatbot on a second devicePhone or tablet beside the laptop, question photographed or typed inWebcam review, gaze and timing patterns, a second camera for high stakes
Invisible overlay appsAn assistant window that is drawn on the candidate's monitor but excluded from screen captureVery little on screen. Timing and response forensics, installed lockdown clients
Smart glasses and earpiecesA camera in the frame reads the question, answers arrive in the lens or by audioPolicy, eyewear checks, microphone clips, forensics. Webcams rarely see it
Remote-control proxyA paid expert takes over the machine through remote-access software while the candidate sits on cameraVirtual machine and environment checks, device fingerprinting, response patterns
Impersonation and deepfakesSomeone else sits the exam, sometimes behind a manipulated video feedID photo capture, snapshot comparison by a human, retesting on suspicion
AI-assisted item harvestingQuestions are captured and reconstructed into "practice" dumps, sometimes with AI filling the gapsItem exposure monitoring, large pools, candidate agreements

The overlay row is the one most buyers haven't heard of. As of September 2026, the meeting assistant Cluely says on its own homepage that it "never shows up in shared screens, recordings, or external meeting tools." It isn't unique to one product. Windows has supported a window setting called WDA_EXCLUDEFROMCAPTURE since Windows 10 version 2004, which shows a window on the monitor and keeps it out of screenshots and recordings. A proctoring tool that relies on screen capture will record a clean desktop.

Smart glasses are the other gap. The College Board banned smart glasses at the SAT starting in March 2026, prescription pairs included. A test center can enforce that at the door. A remote exam has to rely on policy, a camera check and whatever the data shows afterwards.

Why does proctoring alone miss most of it?

Proctoring watches the candidate. Most modern cheating happens where the proctor isn't looking: a device outside the frame, a window the operating system hides, a remote session, or a leak weeks earlier. Caveon's article makes the cost side plain too. It puts in-person proctoring at $30 to $100 per exam and live online proctoring at $15 to $50, plus fees, and it argues for secure exam design and data monitoring over more surveillance.

We agree, with one caveat. Proctoring still raises the effort needed for casual cheating, and it produces the evidence a reviewer needs when forensics raises a question. Treat it as one layer, not the plan. Our guide to how online proctoring works covers the controls in detail.

Layer 1: how can exam design make AI and leaked answers less useful?

A chatbot answers recall questions in a second, and so does a braindump. The design goal is to make each item harder to look up and each leaked item worth less.

  • Draw from pools larger than the exam. If a 60-item form draws from 200 or more items, two candidates comparing notes see mostly different questions.
  • Tag difficulty so random draws stay fair. Pull a fixed number of easy, medium and hard items per domain, or forms stop being comparable.
  • Shift weight from recall to application. Scenario items with an exhibit, a log file or a configuration screenshot take longer to paste into a chatbot and are harder to answer from a dump.
  • Use item types that require doing the task. A coding question that runs the candidate's code against test cases, or an image hotspot where they click the failing component on a wiring diagram, gives an AI less to work with than four options.
  • Consider negative marking on sections where guessing from a half-remembered dump is the pattern you see.
  • Retire and replace exposed items on a schedule, not only when someone reports a leak.

Be honest about the limits. Current models can reason through many scenario items, so design reduces the advantage rather than removing it. Large testing programs go further with linear-on-the-fly or computer-adaptive testing and equated forms, which need psychometric support and a platform built for it. If you're rebuilding the bank anyway, start from a job task analysis and exam blueprint so the new items measure what the credential claims.

Layer 2: which identity and device controls stop proxies and remote control?

Proxy testing is organized business. Meazure's report describes finding 14 professional cheating rings within the first 2,000 exams for one client. The controls that help most happen before the first question:

  1. Tie access to a person. Invite candidates by email and require a one-time code sent to that address. It's weak identity alone, but it stops casual link sharing and gives you a record.
  2. Capture an ID photo at the start and keep it with the attempt, so a reviewer can compare it with webcam snapshots later.
  3. Lock the attempt to one session and one device. A second browser tab or a device switch mid-exam should be caught and logged, not quietly allowed.
  4. Check the environment. Virtual machine detection matters because a candidate can run the exam inside a VM while the host machine runs anything. External monitor and camera checks catch the second screen a helper is watching.

Know the trade-off. Browser-based proctoring runs inside the browser's sandbox, so it can't list other programs on the computer or see a remote-access tool directly. Installed lockdown clients can inspect more of the machine, at the cost of an install step, admin rights on managed laptops and more support tickets on exam day. Ask any vendor, in writing, which remote-access tools they detect and how.

Layer 3: what should in-session monitoring actually watch for?

Each control should answer a specific threat from the table above. A starting setup for an external certification:

  • Fullscreen lockdown with logged exits. Leaving fullscreen or switching tabs is logged every time, with a violation limit that ends the attempt after repeated events.
  • Blocked copy, paste and right-click, with attempts logged. A candidate who repeatedly tries to copy question text is telling you what they planned to do with it.
  • Screen snapshots. They won't show a capture-excluded overlay, but they do catch the ordinary chatbot tab and a candidate who drags a window over the exam.
  • Webcam snapshots with face monitoring. Flags for no face or more than one face in the frame. Reviewers also spot the classic second-device posture: eyes repeatedly dropping to the same spot below the camera.
  • Microphone monitoring, with consent. Short clips captured when speech is detected. Some candidates read questions aloud to a voice assistant, a pair of glasses or a helper on a call, and a transcript of those clips is quick to review.
  • Timing anomaly detection. Long idle gaps followed by bursts of fast, correct answers are a pattern worth flagging on their own.

For smart glasses, add a pre-exam step where the candidate holds their eyewear up to the webcam, backed by a policy that bans any glasses with a camera or display. It won't stop a determined cheater, but it removes the "I didn't know" defense.

One setting deserves thought before launch: which events end an attempt. A dropped Wi-Fi connection or a face briefly leaving the frame happens to honest candidates. Repeated tab switching or paste attempts rarely do. Put the second group toward the violation limit and send the first group to review.

Layer 4: what is exam data forensics, and what should you run?

Exam data forensics is the statistical analysis of responses, timings and session data to find cheating that nobody saw live. It's the layer that catches the overlay user, the glasses wearer and the proxy ring, because their answers still have to go through the exam. Meazure's April 2026 write-up on data analytics in exam security describes the main techniques:

TechniqueWhat it looks forCan a small program do it?
Response-time analysisAnswer speeds that suggest prior access to items, or effort that doesn't match resultsYes, if your platform exports attempt timings
Similarity analysisCandidates sharing unusual runs of identical wrong answersPartly, with a spreadsheet and a clear threshold
Device and network fingerprintingShared devices, system traits or VPN infrastructure across different candidatesHard without vendor tooling
Item exposure monitoringItems performing easier over time, or watermarked content turning up onlineYes, track item difficulty over time
Cross-session analyticsBehavior patterns that repeat across many sessions, a sign of coordinated ringsMostly vendor or consultant work

If you run a few hundred attempts a year, start with two checks. Track each item's percent-correct by month, because an item that suddenly gets easier is probably circulating. Then sort attempts by time against score and look at the fastest top scorers. Neither is proof, but both justify pulling the evidence.

Layer 5: who reviews the flags, decides and documents?

Meazure sums up where the industry has landed: "AI assists, humans decide." Automated flags are signals. Poor lighting, glasses glare or a child walking through the room all trip monitors, and none of it is cheating.

Write the review process down before the first exam goes live:

  • Who reviews flagged attempts, and who makes the final call on high-stakes cases (ideally not the same person)
  • What evidence the reviewer checks: event log, snapshots, audio transcripts, timing, forensics results
  • The outcomes available: accept, require a retest under stricter conditions, invalidate the score, or refer for sanctions
  • How a candidate appeals, and how long they have
  • Where the decision and the reason are recorded, so the next reviewer sees the same standard applied

A retest is often the fairest tool you have. A flagged candidate who passes again under full proctoring has shown what they know.

Which candidate policies close the gaps?

Technology can't enforce a rule you never wrote. CompTIA is a useful model. Its unauthorized training materials policy says candidates must not use AI or LLM tools to try to recreate, predict or obtain exam questions, and that AI-generated content containing CompTIA exam content may be treated as cheating regardless of intent. Candidates caught using unauthorized materials lose their certification and are banned from CompTIA exams for at least 12 months.

Your candidate agreement should cover, by name:

  • AI assistants on any device, including overlay and "meeting assistant" apps
  • Smart glasses, smartwatches and earpieces, prescription or not
  • Remote-access and screen-sharing software
  • Braindumps and AI-generated "practice exams" built from real items
  • Your right to require a retest, invalidate a score and revoke a credential after it's issued
  • What you capture (camera, screen, audio), who reviews it and how long you keep it

Repeat the key rules as an honesty pledge on the exam's first screen. It feels like a formality, but in an appeal it's the page you point to. Plan accommodations early too, since some assistive technology conflicts with lockdown controls.

How to run the five layers in LMS Advisor

We built the Test Center in LMS Advisor for certification programs, and it covers layers 1 to 3 plus the evidence for layer 5. Here's how the plan above maps to it:

LayerIn LMS Advisor
1. Exam designA question bank with pools, difficulty levels and negative marking. The 14 question types include coding items that run the candidate's code and image hotspots. AI question generation drafts new items for review, and a bank audit helps you find weak ones
2. Identity and deviceEmail invites with one-time-code access, ID photo capture, single-session and device lock, virtual machine detection, external monitor and camera checks, heartbeat monitoring
3. In-session monitoringNative proctoring settings you switch on per exam: fullscreen lockdown, blocked and logged copy, paste, right-click and tab switching, violation limits, screen and webcam snapshots, AI face monitoring, consented microphone clips transcribed with OpenAI Whisper, idle and timing-anomaly detection
4. Data forensicsNo built-in forensics suite. Result exports to CSV or Excel and the REST API feed the item-difficulty and time-versus-score checks in your own spreadsheet
5. Human decisionAn evidence trail of logged events, snapshots and transcripts for reviews and disputes, an honesty pledge on the first screen, and in-exam issue reporting so an honest candidate can flag a dead camera

What it won't do: it can't see a capture-excluded overlay or detect smart glasses (we don't think any browser-based tool can promise that), and it has no LOFT or adaptive testing engine or live human proctors. Programs that need those should look at the specialist options we compared in built-in versus bolt-on proctoring. For how exams, certificates and public verification fit together for a credentialing body, see our page for certification bodies. To test these settings against your own exam, start an LMS Advisor trial and run a practice attempt.

Frequently asked questions

Can proctoring software detect ChatGPT during an exam?

It can detect the usual ways ChatGPT gets used on the same computer, such as a new tab, leaving fullscreen or pasting question text. It can't see a phone outside the camera's view or an overlay hidden from screen capture, which only webcam review and data forensics might catch.

Can proctoring software detect AI smart glasses?

Not reliably. A webcam image rarely shows enough detail to tell smart glasses from ordinary frames. The practical defense is a written ban, a pre-exam eyewear check on camera, microphone monitoring with consent and forensics that flag answer patterns out of line with the candidate's history.

How do proxy test takers cheat on online certification exams?

The common method is remote control: the candidate sits on camera while a paid expert operates the computer through remote-access software. Others use a virtual machine, a second hidden screen or outright impersonation. ID capture, device lock, VM and monitor checks and device fingerprinting in forensics are the main countermeasures.

What is data forensics in exam security?

Data forensics is the statistical review of answers, timings and session data to find cheating that live monitoring missed. Typical checks include response-time analysis, similarity between candidates' wrong answers, device fingerprinting and tracking items that get easier over time. Results justify a closer look or a retest, not an automatic penalty.

Is using AI to study for a certification exam cheating?

Studying with AI is generally allowed, but using it to recreate or obtain real exam questions usually isn't. CompTIA, for example, treats AI-generated content that contains its exam items as a policy violation even when the candidate didn't intend to cheat. Check each certifying body's candidate agreement.

Can LMS Advisor help prevent AI cheating on certification exams?

Yes, for the layers software can cover. LMS Advisor combines large question pools and hands-on item types with per-exam proctoring (lockdown, ID photo capture, device lock, VM detection, face and consented audio monitoring) and an evidence trail for human review. It can't see capture-excluded overlays or smart glasses, so pair it with a candidate policy and your own data forensics.

Should certification programs move back to test centers?

For the highest-stakes exams, a test center is still the strongest control against wearables and remote-control proxies, because staff can inspect people and hardware. It costs more and limits access, so many programs keep remote delivery and use test-center retests when doubts arise.

Add LMS Advisor to your preferred sources

Add lmsadvisor.com as a preferred source on Google, or open this article in your AI assistant to use it as a source.

Swati Priyadarshani
· Founder & CEO at LMS Advisor

I'm Swati Priyadarshani, Founder & CEO of LMS Advisor and Co-Founder of WorldWin Coder Pvt. Ltd. (est. 2019). Over the last 10+ years, I've helped enterprises build learning platforms that actually get used - not just deployed.

Keep reading

See LMS Advisor with your own use case

A product specialist walks you through authoring, a proctored exam, certificates and the AI tools, using the programs you actually run.