Anti-AI-cheating: how HireValid keeps hiring tests AI-proof

Why AI cheating needs more than a detector

A candidate who wants help from AI today has more options than ever. General chatbots answer a pasted question in seconds. Desktop assistants read the screen or listen to a call and suggest answers in a small window; one of them, Cluely, describes itself as "invisible to screen share". Some tools are marketed directly at interviews and hiring tests. Our employer's guide to AI interview cheating tools maps the main categories.

Tools like these change every month, so an approach that only tries to spot them is always one release behind. AI cheating prevention works better when it starts with the test itself. If the questions are different for every candidate, arrive one at a time under a tight time limit that a page reload cannot reset, and cannot be copied, outside help becomes slow, awkward and visible. That is the idea behind every HireValid test, and it is why we describe them as AI-proof.

Prevention is the first layer, not the only one. Anti-AI-cheating evidence shows the employer what happened during each attempt, and the final call always stays with a person. The rest of this page lists the real checks behind each layer.

Layer 1: tests that make AI help hard to use

A different test for every candidate. Each test draws its questions from large question pools, at least three times bigger than what a test uses and five times for cognitive tests, so two candidates rarely see the same set. How we build our tests explains how every pool is written and reviewed. Questions come in a different order for each attempt, and the answer options are shuffled too.

One question at a time, with nothing to steal. The browser receives one question at a time, rebuilt on the server from an allow-list of display fields. Answer keys, scoring keys, the hidden test cases of coding questions and the solutions of spreadsheet tasks never leave our servers, and an automated leakage suite checks every candidate response for them. Candidates cannot go back to an earlier question.

A clock the candidate cannot reset. Timing runs on the server. Reasoning questions carry their own limits, typically 45 to 90 seconds, which is long enough for someone with the skill and too short to paste a question into a chatbot and check the reply. Reloading the page brings back the same question with only the time that was left, and a question that runs out of time is recorded as timed out before the next one appears.

Copying is blocked, and attempts are noted. Question text cannot be copied, cut or selected, and right-click is switched off. A blocked copy is still recorded in the timeline. Full-screen mode is on by default and the employer can switch it off for a job.

Links that cannot be passed around. Each invitation link is personal, stored only as a secure hash, and expires; a finished test cannot be taken again. If the same link is started on a second device, only the newest session stays live.

Formats that resist AI. At least a third of the questions in every cognitive test are built on tables, figures or sequences rather than prose that pastes cleanly. Situational judgement tests turn on scenario details that generic AI advice misses, and spreadsheet tasks such as the Excel intermediate test ask candidates to do the work rather than describe it.

Layer 2: every check HireValid records

While a candidate works through the scored questions, HireValid records these checks. Anything that happens during the practice questions is shown but never counts.

Browser and page checks

• Time away from the test. Every time the test tab or window loses focus, and for how long.
• Full-screen exits, and how long the candidate stayed outside full screen.
• Pastes into answers, with the number of characters pasted.
• Text that appears without being typed: a block of text arriving at once, without the keystrokes that would produce it.
• Blocked copy attempts on question text.
• Reloads and reconnections, so a gap in the timeline has its explanation beside it.

Timing checks, worked out on the server

• Answers faster than anyone could read the question. Each test sets a minimum reading time, typically one and a half seconds, and every scored answer under it is noted.
• Unusually even timing. When answer times over a whole test are suspiciously regular, the pattern is flagged once.

Camera checks, with the candidate’s consent

• Webcam photos at random moments during the test, each checked automatically for no face, more than one face or a face turned well away from the screen.
• Camera turned off, with how long it stayed off, and camera permission withdrawn mid-test.
• An identity photo before the test starts, when the employer asks for one.

Device and network checks

• The same device as another candidate for the same job, and the same network as another candidate for the same job. Both candidates are flagged, whoever finished first.
• More than one screen connected, where the browser reports it.

Device and network details are never stored raw. Each is replaced by a fingerprint that only works within one job, so HireValid can tell that two attempts share a device without keeping the device’s details.

Layer 3: an Integrity Score you can explain

Every check feeds the Integrity Score, from 0 to 100, measured across all of a candidate’s tests in the job. Scores from 80 to 100 are High confidence, 50 to 79 Review suggested and 0 to 49 Serious concerns.

Each kind of signal is weighed by what matters about it: time away is weighed by how long the candidate was away rather than how many times they switched, pastes and text without typing by how often they happened, and webcam checks by the share of photos affected. Every signal also has a cap, set so that one signal alone, or any two together, cannot push a candidate into Serious concerns. The single exception is a pair that points to someone else taking the test: no face in the photos together with the same device as another candidate. It takes several independent signals to reach Serious concerns, which keeps honest candidates with a noisy room or a shared office network out of it.

The weights live in a versioned ruleset, and every score records the version that produced it, so a result can still be explained a year later. Free shows the score and band; Starter and Growth add every signal with its weight, and an evidence timeline that lists each event in plain words, in order, with the photos. The integrity summary also appears in every candidate PDF report.

How to read a result like this

Eight kinds of signal, none conclusive on its own, add up to Review suggested: enough for a person to look closely, not enough to call it serious. Open the timeline and read it in order. A 24-second tab switch next to a calendar notification is not the same as 300 characters arriving at once in a written answer.

HireValid never rejects, hides or ranks a candidate automatically because of a signal. A notification, a slow connection, assistive technology or a family member walking past can each produce an event, so a low score tells you where to look first, not what happened. Ask the candidate a neutral question about anything unusual, or confirm the skill in a short structured interview, and decide on all the evidence. The help article on understanding integrity signals shows how to read a result step by step.

Fair to honest candidates

Anti-cheating should never mean treating every applicant as a suspect. Honest candidates are the majority, and they deserve a calm test and a clear explanation.

Candidates are told first. The welcome page lists what is recorded and why, and candidates agree before any photo is taken. They see the time limit, the tools they may use and a few practice questions before scored questions begin, and nothing from the practice questions counts. The candidate guide explains the experience from their side.

The test never stops for a check. If the camera turns off, the test carries on and the timeline simply notes it. The camera is optional: an employer can run any job with webcam photos off, and the other checks still apply.

Photos and devices are protected. Webcam and identity photos are encrypted, stored under the hiring company and candidate, visible only to that company through short-lived links, and deleted automatically after the company’s retention period, 90 days at most. Device and network details are kept only as fingerprints that work within one job. The Biometric Data Policy sets out the details.

Accommodations are built in. Employers can give a candidate extra time, and assistive technology is never treated as cheating. A candidate can request an accommodation before starting. Our fairness approach explains how we think about the rest.

How this compares with other anti-cheating software

Most anti-cheating software for hiring takes one of three routes. Live or recorded proctoring watches the candidate through the camera and microphone for the whole test, which is thorough but costly, stressful for candidates and heavy on personal data. Lockdown browsers control the testing computer, but need a download and are more common in education than in hiring. AI-text detectors try to judge whether writing came from a chatbot, and their results are hard to explain to a candidate.

HireValid takes a lighter, layered route built for small teams: AI-proof exams by design, a focused set of anti-cheating checks and a score a person can explain line by line. Nothing needs installing, the camera is optional and the evidence is specific. If you are comparing tools, ask each vendor three questions: does every candidate get a different test, are answer keys ever sent to the browser, and can you see the evidence and the weight behind every flag? The comparison hub shows how several assessment vendors describe their integrity features, with sources.

What this looks like on a real hiring round

Imagine a small company hiring a customer support agent, with 60 applicants in a week. It sends every applicant one link to the customer support agent bundle: customer service situational judgement, English proficiency, attention to detail and typing, about 33 minutes in total.

Each applicant gets a different set of questions. Most finish with a High confidence band. Two show Review suggested. The first timeline reads "Left the test tab", then "Returned to the test tab after 61 seconds", twice. The second shows "Pasted 412 characters" into a typed answer and several answers faster than the question could be read. The hiring manager asks both about it in a short, structured interview. One had a family interruption and solves a similar scenario live with ease; the other cannot explain their own answer. The decision is easy, fair and documented.

That is the point of a layered, AI-resistant assessment: most candidates sail through, the few results worth a second look stand out, and the evidence makes the conversation simple. To see the same approach as a product page, visit AI-proof hiring tests, or read how to stop AI interview cheating before candidates reach your shortlist.