Do candidates cheat on online skills assessments, and does proctoring / anti-cheat actually stop it (including AI/ChatGPT)?

Cheating on unsupervised online assessments happens often enough that every major vendor sells detection features, though honest prevalence numbers are harder to pin down than the marketing suggests. CodeSignal announced in February 2026 that flagged fraud attempts across its platform rose from 16 percent of assessments in 2024 to 35 percent in 2025, reaching 40 percent at entry level [6]. HackerRank has separately stated that 25 percent of technical assessments show signs of plagiarism [4]. Both come from companies selling anti-cheat features, neither publishes a sample size or a method, and both count flags rather than confirmed cheating. Other widely repeated survey statistics were left out because no openable primary source stood behind them.

What generative AI changed

A 2024 study tested ChatGPT-3.5, Gemini 1.0 Pro, and Meta AI against archived programming problems. ChatGPT solved 71.43 percent of 98 archived LeetCode problems, Gemini 68.37 percent, and Meta AI 58.16 percent, and all three obtained HackerRank problem-solving certificates at basic and intermediate levels [7]. Difficulty mattered, since the same models cleared only 8 of 49 problems across nine live contests [7]. Those models are two generations old now, so current systems clear the easier tiers more comfortably.

What proctoring changes in measured scores

Proctoring moves scores less than most buyers expect. The meta-analysis of 49 studies and 100,434 test takers cited in a 2025 review found unproctored scores higher by a standardized mean difference of 0.20 [8]. The accompanying experiment with 226 participants found no significant main effect of proctoring condition and observable cheating in only 3 of 74 recorded participants [8]. Both settings were low stakes, and a job offer raises the incentive considerably.

Candidates use four broad methods, and only some of them leave any trace inside the assessment software.

A second physical device running a chatbot

A phone or second laptop beside the test window is the hardest method to detect, because nothing on that device touches the assessment session. TestGorilla, describing the limits of its own detection, notes that candidates can read from secondary devices or monitors the software does not track [1]. Webcam proctoring sees a face and a rough gaze direction, not what sits two feet to the left.

An invisible overlay on the test machine

A second class of tool renders answers in a layer that never appears in a shared screen or recorded video feed. Cluely launched in 2025 around a hidden in-browser window the company called undetectable, prompting two startups to ship counter-detection products within weeks, and its chief executive answered by floating smart glasses or a recording necklace as ways past detection [11].

Someone else takes the test

Proxy test-taking and identity fraud are two of the four violation categories CodeSignal says its systems target [5]. The variant that defeats a webcam is a coached candidate, where the right person sits on camera while a helper watches a shared screen and feeds answers by phone.

Circulated answers from leaked question banks

CodeSignal runs a service called LeakSweep that monitors sites including Stack Overflow and LeetCode for leaked questions, and HackerRank sends leaked-question alerts [5][2]. Any static item pool used at volume degrades on a schedule you do not control. TestGorilla separately lists VPNs, notes outside the webcam view, and deepfaked video or voice among the ways its own controls get bypassed [1].

Session monitoring watches the candidate and the test machine while the assessment runs, which makes it good at catching activity on that machine and blind to everything else.

Recorded and live proctoring

Recorded proctoring covers the candidate's face, the room within the camera's field of view, and screen activity for the whole session [5]. Reviewing footage is real work, so set the review threshold before you turn it on, because a recording nobody watches is only a deterrent. A live proctor can ask a candidate to pan the room and stop a session in progress, which makes it the only control here that adapts to something unexpected [5], and the most expensive to run.

Lockdown browsers and tab tracking

Locking the browser and logging tab or focus changes closes off the easiest route, which is opening a chatbot in the next tab [1][2]. Neither reaches a second device, and overlays that draw on top of the screen without taking window focus pass through both [11]. Candidates on personal machines also decline lockdown installs more often than teams expect.

Keystroke and paste detection

Paste events, unusual typing rhythm, and large code resets separate original work from transcription, and HackerRank names code resetting and external copy-paste among the patterns its model watches [3]. A candidate who reads an answer off a phone and types it by hand produces a clean keystroke record, which is why HackerRank's own testing showed reworked model output passing similarity checks [4].

Content and identity controls work on the test material and on who is sitting for it, so they hold up better against a second device than session monitoring does.

Question randomization and large item pools

Rotating variants and retirement thresholds limit the damage when items leak, and CodeSignal rotates thousands of question variations so candidates do not receive identical assessments [5]. That solves the circulated-answer-key problem and does nothing about a model solving a fresh variant in under a minute [7]. Ask any vendor how many live items exist per skill and how often they retire.

Code similarity and AI plagiarism models

Similarity detection compares submissions against other candidates and known public solutions, and MOSS-style comparison survives cosmetic edits such as renamed variables. HackerRank's own 2023 testing showed that reprompting ChatGPT produced solutions that passed those checks cleanly, and the company put MOSS false positive rates as high as 70 percent [4]. Its replacement model combines code and behavioral signals at a documented 85 percent precision, with a caveat that detection is limited on very short answers [3].

Identity verification

Photo ID matching and multiple-face detection address the proxy test taker directly, which no other control here does [2][5]. It cannot see a helper sitting off camera, and it carries the most legal weight on the list, since face matching produces exactly the kind of data Illinois regulates [13].

No control on the market reliably stops a candidate consulting a chatbot on a device the assessment software cannot see. Detection raises the effort and the risk, because a candidate who pastes model output gets caught by paste logging and similarity checks while a candidate who retypes it usually does not [1][3][4].

Detection of AI-assisted work rests on behavioral inference rather than on identifying machine-written text. HackerRank's model combines copy-paste activity, code similarity, time to solve, tab-switching, and iteration patterns, then returns a high or medium suspicion level at a stated 85 percent precision [3]. CodeSignal's Suspicion Score works the same way, turning solution similarity, telemetry, and copy-paste activity into a trust level per submission [5]. Both descriptions are the vendors' own, and neither has been audited publicly.

The setting that helps most removes the ambiguity, and HackerRank lets you declare and enforce whether candidates may use AI tools during an assessment [2]. Once AI use is either permitted and scored or banned and enforced, the assessment measures something you can name.

What anti-cheat technology is not

It is not proof. A flag is a probability, and at 85 percent precision roughly one flagged session in seven belongs to someone who did nothing wrong [3]. Rejecting on a flag alone, with no human review, converts a detection tool into a defensibility problem. It is also not a validity fix, since a test a model can solve measures the model.

Assessments built on your own material and followed by a live conversation are the hardest to outsource, and better evidence sits behind them than behind any monitoring stack. In the 2022 re-analysis of personnel selection validity, which corrected overcorrected range-restriction estimates and cut most validity coefficients by .10 to .20, structured interviews came out as the top-ranked selection procedure [10].

Work samples tied to something only you have

A task drawn from your actual codebase or a scenario your team faced last quarter has no public solution for a model to retrieve and no answer key to circulate. HackerRank's multi-file project questions move in this direction, placing a candidate inside a realistic codebase instead of an isolated puzzle [2]. The cost is authoring effort, so most teams keep one or two per role family.

A live follow-up on submitted work

A fifteen-minute conversation about a take-home separates the candidate who wrote it from the candidate who received it, with no surveillance at all. Ask why an approach was chosen and what would break at ten times the data volume. Someone who cannot narrate their own decisions has told you what you needed to know.

Tasks that require judgment about ambiguous context

The 2024 platform study shows where models weaken, since they were strong on archived, well-specified problems and much weaker on live contests [7]. Assessments built on incomplete requirements put the work back where a model gives least help.

Treat the assessment as a signal, not a verdict

Using a score to rank and route rather than to reject outright limits the damage a cheated result can do, and it softens the false positive problem, since a flag becomes a question for the next conversation rather than a silent rejection.

Proctoring costs completions and creates data you then have to protect and justify. Research on test takers found participants concerned about both the volume and the personal nature of information shared with proctoring companies [9]. A candidate with options who is asked to install monitoring software and record their room for a first-round screen has an easy alternative.

Biometric identity checks and Illinois BIPA

Face matching produces a scan of face geometry, which the Illinois Biometric Information Privacy Act defines as a biometric identifier [13]. Before collecting one, a private entity must inform the subject in writing, state the purpose and length of storage, obtain a written release, and publish a retention schedule with destruction within three years of last interaction [13]. Section 20 gives any aggrieved person a private right of action with liquidated damages of $1,000 per negligent violation and $5,000 per intentional or reckless violation plus fees, which turns this into a litigation risk rather than paperwork [13].

Recorded video and the Illinois AI Video Interview Act

An employer that asks an applicant to record a video interview and uses AI to analyze it for an Illinois-based position must notify the applicant beforehand, explain how the AI works and what characteristics it evaluates, and obtain consent; on request the videos must be deleted within 30 days including backups [14]. An AI-scored recorded assessment with a video component sits close enough that counsel should decide, not the vendor.

Accommodation duties under the ADA

An employer must make reasonable accommodations, including in the administration of tests, to the known limitations of an otherwise qualified applicant with a disability, unless doing so is an undue hardship [12]. Proctoring rules collide with this regularly, since a screen reader triggers focus events and a movement disorder triggers gaze flags. Confirm before you buy that extended time and assistive technology run without generating integrity failures. This is general information about US federal and Illinois law rather than legal advice.