TL;DR
Define permitted tools, design useful tasks and investigate integrity concerns without overclaiming.
Who this guide is for
This guide is for employers who run online assessments and want the results to reflect each candidate's own skill. It focuses on decisions you control: what each stage should measure, how questions are written, which signals you collect, how you review a concern and how you confirm skills later. It does not describe how any assistance tool works in detail, and it is never a guide to using or concealing one.
Work through it with a real assessment open beside you. For every task in that assessment, you will write down three things: what the task is meant to show, which tools the candidate may use, and what you would do if something looked unusual. Those three notes are the backbone of an integrity approach that candidates can understand and reviewers can apply.
Every scenario below is illustrative. The aim throughout is a layered defence: tests that make outside help hard to use, evidence that shows what happened, and a fair review when something looks unusual.
Why has cheating in online tests become harder to judge?
For years, the main worries about unsupervised tests were a friend sitting nearby or answers shared between applicants. General-purpose AI assistants have changed the picture. A candidate can now paste a question into a chat tool, use an app that reads the screen and suggests answers, or keep a second device out of view. Some of these tools are marketed specifically at interviews and assessments.
Two consequences follow for employers. First, polished written answers are no longer strong evidence on their own, because fluent text is cheap to produce. Second, any single technical signal has become less decisive, because the same event can come from an innocent cause or from assistance. The response is not to monitor everything, but to design stages where outside help is less useful and to treat signals as prompts for review.
For an overview of the tool categories and how they affect interviews as well as tests, read the employer's guide to AI interview cheating tools. If a specific product name keeps coming up, the pages on whether Cluely is detectable and Interview Coder for employers give an employer-side view.
Decide what each stage must show on its own
Integrity starts with a definition. If you have not decided what independent work means for a given stage, neither you nor the candidate can tell whether a rule was broken. Many hiring processes have several stages, and the right answer can differ between them.
A short screening test usually exists to show unaided reasoning, judgement or core skill. A later work sample may be closer to the real job, where colleagues and tools are normally available. Writing the rule per stage avoids a blanket ban that is unrealistic for modern work and a blanket permission that makes early screening meaningless.
A permitted-tools plan by stage
The table below is a hypothetical plan for a junior analyst vacancy. Your own plan will differ, but the structure is worth copying.
| Stage | What it should show | Tools allowed | Tools not allowed | | --- | --- | --- | --- | | Online screening test | Reasoning and SQL fundamentals | Built-in calculator, assistive technology | AI assistants, other people, search | | Take-home task | Cleaning and summarising a small dataset | Spreadsheet, documentation, AI with disclosure | Another person doing the work | | Structured interview | Explaining the take-home choices | The candidate's own submitted work | Real-time prompting tools |
Notice that assistive technology is listed as allowed. A rule that accidentally forbids a screen reader or dictation software turns an integrity policy into an accessibility barrier.
Write the rule the candidate will read
Translate the plan into one or two sentences per stage, in plain language, and put them in the invitation as well as on the first screen of the assessment. A hypothetical example: "For this 25-minute test, please work alone and do not use AI tools, search engines or help from other people. Screen readers, magnifiers and other assistive technology are welcome."
A clear rule does two jobs. It sets an expectation most candidates will follow, and it gives you a fair basis for a follow-up question if something looks unusual. Without it, a candidate who used a tool in good faith has done nothing wrong.
Design questions that are less useful to outsource
The strongest integrity measure is usually the question itself. Generic prompts invite generic, easily generated answers. Specific tasks built around supplied details, realistic constraints and a time limit that matches the work make outside help slower and less helpful, while still giving honest candidates a fair chance.
Build tasks around supplied details
Give the candidate the facts they need inside the question: a short customer message, a small table of numbers, a policy excerpt or a fragment of code. Then ask for a decision that depends on those details. A general-knowledge question can be answered by anything that knows the topic; a question about this table, these constraints and this deadline requires reading carefully and choosing.
Situational judgement scenarios work well here when the options differ in subtle but meaningful ways. Our article on situational judgement examples explains how to write options that reveal a real difference in approach and how to record a scoring rationale rather than just an answer key.
Size time limits to the task
A time limit is not meant to create panic. It is meant to make the test about what a person can do with their own knowledge in a realistic period. Try each section yourself and with a colleague, then set the limit so a competent person can finish without rushing. A limit that only the fastest readers can meet measures reading speed and disadvantages candidates who need extra time.
Draw items from pools and vary the order
When every candidate sees the same questions in the same order, answers can be shared between applicants and looked up in advance. Questions drawn from a larger item pool, with options shuffled, reduce that risk. HireValid tests draw questions from randomised pools, so two candidates for the same role rarely see an identical sequence.
Ask for reasoning in a later stage
A short follow-up in which the candidate explains one or two of their own answers is one of the most informative checks available. It is not a trap: someone who reasoned their way to an answer can usually describe the route, while someone who relied on outside help may struggle to explain choices they did not make. Plan this follow-up for everyone on the shortlist, not only for people with signals.
For reasoning tests in particular, our article on what cognitive ability tests measure covers how to separate reasoning from job knowledge and how language demands can affect a result.
What can a browser-based assessment observe?
A well-built assessment records a lot. Depending on the settings you switch on, HireValid records the candidate leaving the assessment window or full-screen mode and for how long, pasting text into an answer, text that appears without being typed and answers that arrive faster than anyone could read the question. With the candidate's consent, webcam photos at random moments are checked for no face, more than one face or a face turned away, and an identity photo is taken before the test starts. It also notes when another candidate for the same job used the same device or network.
Read every signal in context. Normal life creates some of them too: notifications, unstable connections, password managers and assistive software can all look like unusual events, which is why a person, not the software, makes the call.
HireValid combines the signals it records into an Integrity Score for each attempt, with the underlying events listed so a reviewer can see what produced it. The integrity page explains the signals in more detail, and the help article on understanding integrity signals shows how to read them in a result.
How should you interpret an Integrity Score?
Read an Integrity Score as a summary of how many observed events need context, not as a verdict on honesty. A high score means fewer events worth asking about. A lower score means a reviewer should look at the timeline and think about plausible explanations before forming any view. It is never proof that a candidate cheated.
A hypothetical review
Suppose two candidates for a support role have similar test results. One has a clean timeline. The other left the window twice for about ten seconds each, early in the test, and has no paste events or timing anomalies. A reasonable reviewer would note the events, consider that a notification or a quick check of the deadline is a likely explanation, and progress the candidate on the strength of their result. If doubt remained, the follow-up interview could include a short live scenario.
Now suppose a third candidate shows several long absences followed by complete, polished answers pasted in, on questions where most people type. That pattern deserves a conversation. It still does not justify a conclusion without one, because the candidate may have drafted in another application for accessibility reasons, or misunderstood the rules.
Review a concern through a fair procedure
Agree the procedure before you need it. When a concern arises under time pressure, a written process keeps reviewers consistent and protects candidates from snap judgements.
Steps for a proportionate review
- Check the facts: which events were recorded, when and on which questions.
- Consider innocent explanations, including accessibility tools and technical problems.
- Compare the pattern with the candidate's other evidence, such as written answers and interview performance.
- If doubt remains, ask a neutral question, such as "We noticed the test window lost focus several times; was anything interrupting you?"
- Offer a fair way to show the skill again, such as a short supervised task or live discussion.
- Record the decision and the reasons, separating what was observed from what was concluded.
The words you use matter. "The system shows you cheated" asserts something the evidence cannot support. "We would like to understand a few events in the record" invites an explanation and keeps the decision with a person. Our article on stopping AI cheating in assessments discusses how to treat browser signals as context and why AI-text likelihood estimates should be used conservatively, if at all.
Confirm skills in later stages
No single stage should carry the whole integrity burden. A screening test that is reasonably resistant to assistance, followed by a structured interview and a short practical task, produces a picture that is much harder to distort than any one component.
In the interview, return to the evidence. Ask the candidate to talk through a scenario they answered, or give them a variation of a question with one detail changed and ask how their answer would change. Our guide to structured interviews shows how to write these questions and score answers consistently across interviewers.
A work sample completed with the shortlist, live or within a short window, is the closest you can get to the real job. If the job allows AI tools, consider allowing them in the work sample too, with disclosure, and asking the candidate to explain what they accepted, changed or rejected. That shows judgement in a way a ban cannot.
Should you ever allow AI in an assessment?
Sometimes, yes. If the role will use AI tools every day, a stage that forbids them may test a version of the job that no longer exists. What matters is that the rule is explicit and the task is designed for it. A task that allows AI should ask for something the tool cannot supply on its own: a choice between options using supplied context, a critique of a generated draft, or a decision that depends on knowledge of your customers.
Keep at least one stage, usually the initial screen, where independent work is required. That gives you a baseline that is comparable across candidates, and it makes later, tool-assisted stages easier to interpret.
Keep monitoring proportionate
Every signal you collect has a cost in candidate trust, privacy and, sometimes, accessibility. Before switching on a setting, ask whether it answers a question you actually have for this role. Full-screen enforcement and paste restrictions may suit a short screening test; webcam snapshots may be unnecessary for many roles.
Tell candidates which settings are on, what is recorded, who can see it and how long it is kept. The candidate privacy notice and the biometric data policy describe how HireValid handles this information, and candidates can learn about the experience on the candidate information page.
Check each setting against your accommodation route. A candidate using dictation software may trigger typing-burst signals; a candidate who needs breaks may leave the window. Agree in advance that approved adjustments are attached to each result, which stops reviewers from mistaking them for concerns. Candidates find the adjustment route in our accommodation request help page.
Assessment integrity in the wider process
Assessment integrity is one part of a fair hiring process, not a separate project. The same principles that make a test resistant to assistance, such as relevance, clear rules and consistent review, also make it easier to defend and more respectful to candidates. If you are setting up testing for the first time, start with the practical guide to pre-employment tests and return here once you know which tests you will use.
The AI-proof hiring tests page describes how HireValid's design choices fit together. Treat any provider's claims of complete protection with caution, ours included: the realistic goal is AI resistant assessments and careful review, not an assurance that assistance is impossible.
Your integrity review sheet
For each task in your assessment, fill in a row with these columns. The exercise usually reveals at least one task that would be better redesigned than monitored more closely.
- Task and what it should show without assistance.
- Tools permitted, written exactly as the candidate will read them.
- Signals switched on for this task, and the reason for each.
- Likely innocent explanations for each signal.
- The neutral follow-up question a reviewer would ask.
- The alternative evidence a candidate could provide, such as a live task.
- Who makes the decision and where it is recorded.
Repeat the sheet after each hiring round. Note which signals led to useful conversations and which only created noise, and remove settings that did not help.
Questions to answer before switching on a setting
- Which stages require independent work, and is that written in the invitation?
- Does each question depend on details that make outside help less useful?
- Is every signal described to reviewers as an observation rather than a conclusion?
- Do approved adjustments appear on the result so they are not mistaken for concerns?
- Is there a named person who reviews concerns and talks to candidates?
- Will important results be confirmed in an interview or work sample?
When you can answer each of these, you have a practical plan to prevent cheating in online assessments without treating honest candidates as suspects. Try the candidate view yourself with the free sample test before sending your first invitation.
What is the most effective way to prevent cheating in online tests?+
Design, more than monitoring. Short time limits, questions drawn from randomised pools and tasks that need the candidate’s own reasoning make outside help less useful. Add clear rules, integrity signals for review and a structured follow-up, and keep a person in charge of decisions.
Are AI-resistant assessments the same as AI-proof?+
They describe the same goal. An AI-proof assessment is built so AI help is hard to use, with timed, randomised questions served one at a time and copying blocked, and it records anti-cheating evidence when someone tries. Look for both layers, then confirm important results in a structured interview or work sample.
Should I ban AI tools in every assessment?+
Not necessarily. Many employers keep timed screening tests for independent work and allow AI, with disclosure, in a later work sample when the job uses it. Whatever you decide, state the rule for each stage before candidates start.
What should I do if an integrity signal suggests AI help?+
Treat it as a reason to look closer, not as proof. Review the signal in context, then give the candidate a fair chance to explain or to complete a short follow-up task under clear rules. Record what you observed and what you decided, and keep a person responsible for the final decision.