How to prevent AI cheating in hiring assessments

AI cheating in hiring assessments happens when a candidate uses a tool like ChatGPT, Claude, or Gemini to generate answers instead of producing their own, usually by pasting a question into a second tab or screen and copying the response back. Preventing it isn't really a detection problem, even though detection tools exist; it's mostly a design problem, since a well-built assessment gives a chatbot far less to work with in the first place. This guide covers how candidates actually do it, what design changes cut it off, and where technical controls still earn their place. For the broader picture of running a fair, structured assessment program, see our complete guide to online assessment for recruitment.

How Candidates Actually Use AI to Cheat

Before fixing the problem, it helps to know what it actually looks like. The common patterns:

• Copy-paste to a chatbot

  - A candidate opens ChatGPT or a similar tool in another tab or on a second monitor, pastes the question, and copies the answer back - the simplest and by far most common method.

• Real-time overlay tools

  - Some browser extensions and desktop apps sit on top of a video call or test window and feed AI-generated answers as the candidate reads or speaks, harder to spot at a glance than a second tab.

• Pre-written responses

  - If a candidate has a rough idea what the assessment will cover, they generate answers in advance and paste them in once the test starts, which timers alone don't fully stop.

Different methods call for different countermeasures, which is why no single fix covers all of this.

Why This Matters More Than Traditional Cheating

Traditional cheating - googling an answer, having someone else help - usually still requires the candidate to understand enough to find or apply the right information. A capable chatbot can produce a fluent, plausible answer to almost any question with no understanding required on the candidate's part at all.

That changes what's at stake. A hiring decision based on an AI-generated response isn't measuring the candidate - it's measuring the chatbot, and the mismatch usually surfaces later, once the person is actually doing the job without a tool ghostwriting for them. This overlaps with, but isn't identical to, the broader cheating problem covered in our guide on how to reduce cheating in remote hiring tests - that piece covers proxy test-takers, collaboration, and search-engine cheating more generally; this one focuses specifically on AI-assisted answers.

Design-Based Prevention: Questions AI Can't Answer Well

The most durable fix isn't a monitoring tool - it's writing questions that a chatbot genuinely struggles with, regardless of whether the candidate has one open.

A few approaches that hold up:

• Base questions on internal, non-public context

  - A chatbot trained on public information can't meaningfully answer a question about your specific product, codebase, or case study, because it has nothing to draw on.

• Ask for reasoning, not just a final answer

  - "What would you do and why, given these three constraints?" is harder to lift from a generic AI response than "What is the correct answer?"

• Use realistic, slightly messy scenarios

  - AI tools tend to produce clean, textbook answers. A task built on an actual imperfect situation from the role often exposes a copy-pasted response by how poorly it fits the specifics.

• Add a short follow-up question tied to the candidate's own answer

  - Someone who wrote their own response can explain and extend it easily. Someone who pasted an AI answer often can't.

This is closely related to how a strong skills assessment gets built in the first place - our guide on how to build a skills assessment for hiring covers writing task-based questions that mirror real work, which is the same design discipline that happens to resist AI cheating as a side effect.

Technical Controls That Reduce AI Cheating

Design changes do most of the work, but a few technical controls still meaningfully raise the effort required to cheat, especially for objective or coding-style questions where "harder to phrase around" isn't really an option.

• Tab and focus alerts flag when a candidate navigates away from the test window, which is exactly what happens when someone switches to a chatbot tab. TunnelQuiz includes this as a built-in check during any test.
• Second-screen detection catches the setup where a candidate has the test open on one monitor and a chatbot open on another, a pattern that's become more common as AI tools have gotten easier to run alongside anything else on screen.
• Copy-paste blocking on the test interface makes it harder to move a question out to an AI tool and a generated answer back in, without needing to retype anything.
• Tight timers reduce the window for a candidate to switch away, get a response, and paste it back convincingly, especially on tasks designed to require more than a surface-level read.

None of these make AI-assisted cheating impossible - a determined candidate with a second device entirely outside the test environment can still get around browser-level controls. What they do is close off the easiest, most casual version of the problem, which is where most AI cheating actually happens.

What Not to Rely On

AI-content-detection tools that claim to identify "AI-written" text exist, but they carry a real limitation worth being honest about: they can misclassify genuine human writing as AI-generated, particularly from non-native English speakers or candidates with a formal writing style. Treating a detector's flag as proof of cheating, rather than one input worth a closer look, risks penalizing honest candidates.

The more reliable approach combines the design and technical layers above with human judgment on borderline cases, not a single tool's confidence score treated as a verdict.

Building an AI-Resistant Assessment Process

Putting this together, here's a practical sequence:

• Decide your policy first. Some roles genuinely benefit from testing how well someone directs an AI tool, rather than banning it outright - decide deliberately rather than defaulting to a blanket rule.
• Write questions around internal context and reasoning, not general knowledge a chatbot already has.
• Enable tab/focus alerts, second-screen detection, and copy-paste blocking for any test where unaided answers matter.
• Add a short live or written follow-up for high-stakes roles, tied to the candidate's own submitted answer.
• Review flagged sessions manually rather than auto-rejecting based on a detection tool alone.
• Revisit your questions periodically, since AI tools improve quickly and a task that resisted them a year ago may not today.

For roles where coding ability is the core signal, this same logic applies with a few extra considerations - see our guide on technical assessment for developers for the coding-specific version of this problem.

The Bottom Line

Preventing AI cheating in hiring assessments works best as a design problem first and a monitoring problem second - questions built around real, internal context give a chatbot far less to work with than generic knowledge checks ever will. Layer tab/focus alerts, second-screen detection, and copy-paste blocking on top, and you've closed off most of the casual version of the problem. TunnelQuiz includes these checks as part of its browser-based, no-install test environment.