Updated 4 November 2026
A high score in an online campus assessment does not always tell you whether a candidate can do the work. Your team sees the gap later, when a candidate reaches the technical interview and cannot explain the code, design choice or answer they submitted. That does not automatically mean the candidate cheated. But it does mean the test score needs to be checked against what the candidate can explain.
For campus recruitment leaders running large technical hiring drives, there are two separate questions: did the right person take the assessment, and did the candidate answer it without unauthorised help? Online proctoring may help with parts of the first question. It is much less useful if the second question is whether the candidate genuinely understands the answer.
Start by separating impersonation from AI assistance
Impersonation and outside assistance are different problems. If someone else sits the test, the control has to establish who is taking it. If the candidate is getting answers from an AI assistant or another person, the control has to examine how the answers are being produced and whether the candidate can explain them.
The distinction matters. A webcam, ID check or test-centre process may help establish identity, but it does not tell you whether a candidate is reading answers supplied by someone else. Browser lockdown and screen recording may catch some activity, but they do not necessarily reveal an invisible assistant running elsewhere on the device.
Why proctoring alone is not a reliable answer
A University of Twente experiment published in 2021 illustrates the limitation. Researchers tested Proctorio with 30 students: six were instructed to cheat and five were asked to act nervous. The system's automatic analysis did not flag any of the six students who were told to cheat. The researchers then created profiles designed specifically to identify the known cheaters, but three of the six still remained undetected. This was one controlled experiment with one system and a small sample. It is not proof that every proctoring product performs the same way. It is a reason not to treat automated proctoring as conclusive evidence.
The paper is available from the University of Twente research repository. The practical lesson for a hiring team is straightforward: monitoring what happens on a test screen is not the same as establishing that a candidate understands the answer.
AI interview assistants have made the problem harder
Candidates can now use tools designed to listen to an interview, read a question or coding prompt, and generate an answer while the conversation is still happening. Some are built to stay out of a screen share, which makes traditional tab-switch or screen-recording checks less useful.
Cluely, Parakeet AI, and Final Round AI all promote real-time interview assistance. Their features differ, but the core risk for a campus recruiter is similar: a candidate may receive suggested answers while appearing to participate in a normal live interview.
This is not limited to one product or one browser extension. Assistants can run as desktop apps, overlays or phone-based tools. Trying to maintain a list of software names and block each one is unlikely to be a durable approach. The more useful question is whether the interview process can recognise patterns associated with outside assistance and then test the candidate's understanding.
How Fabric detects AI-assisted answers
Fabric analyses more than 20 integrity signals across each interview. Rather than relying only on whether a particular app is installed or visible on screen, it looks for patterns in the interview itself, including response timing, phrasing, audio and consistency across answers. These signals are considered together; one pause, one unusual phrase or one behavioural cue is not enough to establish that someone has cheated. Fabric's published technical explanation says the interview is processed after it ends, with the detection result available within about five minutes.
Fabric's technical deep dive reports a 4/4 result against four interview-assistance tools: Cluely, Interview Coder, Parakeet AI and Final Round AI. Fabric says its system detected all four in that test set. The public site provides a detailed test walkthrough for Cluely, but not equivalent tool-by-tool reports for the other three. The result is useful evidence that Fabric's approach identified signals associated with these tools in its tests; it is not a promise that every attempt, product version, device setup or candidate will always be detected. The system looks for the footprint of AI assistance, not a guaranteed software fingerprint.
You can read the Fabric Cluely detection test and report for the documented Cluely example, and the technical explanation of Fabric's cheating-detection approach for more on the signals.
Detection should lead to a better question, not an automatic rejection
Even a well-supported integrity flag is a reason to review the interview, not a verdict. A candidate may pause because they are thinking, struggle because the connection is poor, or phrase an answer differently because they are nervous. Your team should be able to inspect the relevant evidence and compare it with the candidate's answers and performance.
The strongest second check is a follow-up that makes the candidate reason about the answer they just gave. For a coding assessment, that might mean asking them to explain a design choice, change a requirement or describe an approach they rejected. Someone who understands the solution should be able to work through the reasoning, even if they need time to think.
Three follow-up questions to use in your next campus drive
You can try this with a small group before changing your whole assessment process. Select a sample of high-scoring candidates and ask each person the same questions about their submitted work:
- Walk me through the part of your solution that you found hardest to get right.
- If I change this requirement, which part of your solution would need to change first?
- What approach did you consider and reject, and why?
Look for whether the candidate can explain the decisions in their own work and adapt when the problem changes. Do not score confidence, accent or speed of speech as a proxy for honesty. If a candidate struggles, consider nervousness, language, accessibility and interview conditions before drawing a conclusion. A follow-up is evidence about understanding, not a standalone cheating detector.
Run consistent interviews across a large placement season
A follow-up process only helps if you can run it consistently across the number of candidates in your drive. Fabric's Campus Hiring Mode is designed for large campus workflows and supports up to 500 concurrent interviews, subject to the configuration agreed for your deployment. Candidates can join through a phone browser without installing an app. Discuss your expected volumes, college schedule, language requirements and connectivity conditions with the Fabric team before the drive.
See the Fabric campus hiring workflow to understand how interviews, placement-cell coordination and candidate access fit together.
Fabric combines live, role-specific interviews with follow-up questions and integrity signals. Recruiters can review the candidate's answers alongside the signals raised during the interview, rather than relying on an assessment score or a single flag in isolation. Your hiring team remains responsible for reviewing the evidence and making the selection decision.
Frequently asked questions
Can online proctoring stop candidates from using AI tools?
Not reliably on its own. Some AI assistants are designed to remain invisible during screen sharing or to run outside the browser. Proctoring can still be one layer in your process, but it should not be your only check. Combine it with live questions that test whether candidates understand and can adapt their submitted work.
Can Fabric detect Cluely, Parakeet AI and Final Round AI?
Fabric's technical deep dive reports that its system detected Cluely, Interview Coder, Parakeet AI and Final Round AI in a 4/4 test set. The public site includes a detailed test report for Cluely, but not separate reports for the other three tools. Fabric analyses more than 20 signals associated with AI-assisted answers rather than depending only on identifying a named application. This does not guarantee detection in every interview or setup. Flags should be reviewed with the underlying interview evidence.
Does Fabric verify that the person taking the interview is the candidate?
Fabric's integrity signals can identify a second voice or patterns associated with outside assistance, but that is not the same as proving a candidate's identity. Employers should handle identity checks through their own assessment and recruitment process.
Should we reject every candidate flagged for cheating?
No. A flag is a prompt for human review, not proof. Review the relevant answer and timestamp, consider the candidate's explanation, and apply the same process to everyone. A single signal or a nervous response should not trigger automatic rejection.