The reality of AI cheating in assessments: overblown or underestimated?

The impact of AI on test integrity has been an enormous talking point in occupational psychology circles since AI's proliferation.

Although test integrity has always been an important consideration, AI-based cheating represents a brand-new avenue for the unscrupulous that simply wasn't available before.

But how serious is the threat, and what, if anything, should test providers and employers do about it?

In this article, I will provide a realistic overview of the state of AI-based cheating in the selection and assessment space, and what people get right and wrong about this issue.

What AI-based cheating is and why it matters

AI-based cheating is a candidate using a generative AI tool to produce assessment responses on their behalf, so that the score reflects the tool's ability rather than their own.

Ultimately, AI-based cheating bypasses the whole reason we test candidates in the first place. The goal is to measure the candidate's specific skills and abilities, not the effectiveness of their chosen LLM.

If candidate scores no longer correlate with their actual abilities and skills, the assessment ceases to function as a useful selection tool. High scores mean nothing, low scores mean nothing, and from a predictive perspective the result becomes meaningless, making it unsuitable for high-stakes selection decision-making.

Realistically, it's comparable to getting a friend to complete the assessment for you, except that friend is accessible 24/7 and has no moral scruples about cheating.

Now, I have heard many employers use the excuse, "Well, they will be using AI in the job, so we don't care if they use it during screening." The problem with this line of thinking is that cheating on tests is not evidence of AI skill. Indeed, often you don't even need to write a prompt. You can simply upload a screenshot.

Consequently, one should always assume that AI-based cheating, when conducted successfully, fully nullifies the predictive validity of the assessment tool, reducing the expected ROI from selection and assessment.

How and why some providers understate the problem

The most common platitude used to dispel concerns about AI-based cheating is the argument that very few people actually cheat using AI.

Indeed, one study puts the rate of cheating at around 3% (Robie et al., 2026) when proper messaging against AI use is provided.

On the face of it, this number looks reassuring, but the proportion of cheaters in the applicant pool isn't even the most important variable. We also need to consider two other things: 1) how effective AI is at cheating, and 2) how selective you plan to be using the assessment.

Let's say you have 10,000 early careers applicants completing a test and 3% are cheaters. That gives us 300 cheaters in the applicant pool. If you plan on hiring 100 grads, and every single cheater makes it through the assessment stage, every single hire could end up being a cheater, reducing the assessment's predictive validity to 0.

The research does indeed suggest that AI cheating on occupational assessments is highly effective. Research on GPT-4 from 2024 already shows scores close to the 95th percentile (Hickman, Dunlop & Wolf, 2024), and as of 2026 more recent models have completely eclipsed their 2024 counterparts.

This is an uncomfortable reality for assessment providers. It suggests that the problem is serious enough to warrant fundamental changes in assessment design, which is time-consuming and expensive.

Consequently, it's no surprise that many providers simply rest on their laurels and downplay the severity of the issue, hoping that employers won't think too deeply about it.

How and why some providers overstate the problem

Although the threat is real, and greater than many seem to suggest, we have seen this kind of sudden integrity threat before.

When testing shifted from paper and pencil to online, the threat of item sharing and test exposure became a new and unique threat to assessment integrity. In much the same way, many people abandoned assessments altogether, believing that the answers would simply end up on forums and social media, nullifying their utility.

In response, the industry evolved, using item-banked assessments instead of fixed forms, mitigating the benefit of accessing exposed items and answers.

Naturally, you can't stop people from trying to cheat, but we do have a wide range of tools at our disposal. Candidate messaging, browser monitoring, in-person verification testing and AI resilient assessment design mean that AI cheating can be curbed significantly. We set these out as the ACT framework (assessment design, candidate messaging, technical controls) in our whitepaper, AI Resilience in Assessments.

The threat is also not uniform, as selection ratios largely determine the risk. Although early careers and high-volume hiring require a robust response, low-volume hiring is largely insulated. If you have five applicants and a 3% cheating rate, you simply don't need to worry about this issue as much.

But some providers are calling for a complete, ground-up rethink of assessments and present AI cheating as the final straw for psychometric testing as we know it.

These providers are quite likely using AI cheating as an excuse to bypass proper research and development, providing assessments that forgo the usual checks and balances that determine assessment quality by focusing on AI resilience as the only quality indicator.

Fixed-form assessments, theory-less constructs and black-box scoring are things that under normal circumstances would be easy to call out, but which get shoehorned in to quell fears about AI cheating.

So, it suits these providers to paint a picture of terminal obsolescence, making the issue out to be so catastrophic that all the rules need to change, especially the ones on validity and reliability.

Practical recommendations for employers

For employers, the key is to treat assessment design as another variable that can be controlled to improve AI resilience, without allowing it to override everything else that makes an assessment useful. Validity, reliability, fairness, underlying theory and candidate experience still matter as much as they ever have. An assessment that is extremely difficult to cheat on but measures nothing meaningful is not a good assessment. AI resilience should therefore be added to the list of quality criteria, rather than replacing the existing ones.

Employers should also think carefully about how they monitor the problem. Looking at year-on-year average scores won't tell you very much if only a small proportion of candidates are cheating. Instead, pay particular attention to the upper tail of the score distribution and, more importantly, the part of the distribution around the pass mark. So, if pass rates have increased by 50%, that's much better evidence of cheating than just looking at average scores.

The response should also be proportionate to the evidence. Organisations facing substantial AI-assisted cheating need stronger controls, changes to assessment design, verification testing or tighter monitoring. Organisations finding relatively little evidence can take a less disruptive approach. There is no reason to redesign an entire selection process purely because AI cheating is theoretically possible. Equally, don't simply ignore the problem when the data suggest it is happening. Your whole process is being undermined.

More importantly, employers should not abandon assessments altogether and retreat to CV screening. That is a massive step backwards, as CVs and application forms are substantially easier to outsource to AI. Removing structured assessment doesn't remove the AI problem, it just shifts more weight onto selection methods that are even more vulnerable to it.

But ultimately, employers must do something. The pressure on talent acquisition teams to demonstrate credibility in an AI-enabled environment isn't going away, it's only going to increase.

Final thoughts

The other issue here, particularly pertinent to early careers hiring, where the biggest risk lies, is the declining number of graduate-level positions.

Adzuna data reported in August 2026 shows UK graduate vacancies down 45% year on year, to the lowest level since its records began, no doubt in part because of AI itself.

Consequently, for those hiring grads, you can expect far more applicants than in previous years. Not only does that mean a more aggressive selection ratio, and thus a greater threat from AI cheating, but it also means that candidates are likely to be more desperate.

Additionally, candidates often suspect TA teams of using AI to screen them and thus see AI cheating as fair game.

As a result, I do think the actual proportion of AI cheaters is underestimated in the research, and more recent research is likely to uncover higher levels of cheating.

I therefore strongly believe that test publishers have an obligation to create new assessments that are innately more AI-resilient. Traditional aptitude tests with long passages of text and generous time limits are simply too big a risk in early careers hiring.

But even more importantly, those assessments must be item-banked, provide strong publicly available evidence of validity, reliability and fairness, and be well supported by underlying theory. AI cheating isn't an excuse to min-max one specific aspect of test integrity, as quality assessment hinges on multiple variables that all need to be met. That is the standard we hold our own MindmetriQ game-based assessments to, and the one I would encourage employers to hold every provider to.

Primary author

Ben Schwencke

Chief psychologist at Test Partnership. MSc in Organisational Psychology with over 10 years' experience in psychometric testing.

Contributed to by:

• 

Ben Hopgood