AI-Proof Assessment: A Guide to Testing Real Skills

Two-thirds of students across 20 US research universities used generative AI on their coursework. Nine percent used it to cheat. Among students who reach for AI every day, that number climbs to 26 percent.

Then the researchers who found all this wrote the sentence that should stop every training team cold: there is no single AI-proof assessment model.

So the quiz at the end of your compliance course is not measuring what you think it measures. Below is what the evidence actually says: why detection is a dead end, why the fix has almost nothing to do with catching cheaters, and how moving assessment inside the lesson makes it both harder to fake and better at teaching.

What the research on AI and assessment actually found

The largest study of student AI use to date found that misuse concentrates among the heaviest users and varies enormously by discipline. Igor Chirikov of UC Berkeley and René Kizilcec of Cornell surveyed more than 95,000 students across 20 US research universities during the 2023 to 2024 academic year, publishing in Science in May 2026.

The numbers, as reported by EurekAlert and Phys.org:

• About two-thirds of students used generative AI during the period, and 37 percent used it at least monthly.
• 9 percent admitted using it to cheat.
• Among daily AI users, the cheating rate hit 26 percent, against 7 percent for monthly users.
• Regular use ranged from 62 percent in computer science down to 24 percent in the arts.
• Estimated misuse ran from 17 percent in economics and 16 percent in journalism to just 5 percent in biology.

Kizilcec's summary was blunt: "Assessment reform is necessary and urgent."

The behavior is not slowing down. The UK's Higher Education Policy Institute found in its 2026 survey of 1,054 full-time undergraduates that 94 percent now use generative AI on assessed work, up from 89 percent a year earlier and just over half two years before that. The share pasting AI-generated text straight into submitted work reached 12 percent, up from 3 percent in 2024.

Is there such a thing as an AI-proof assessment?

Not as a single format you can buy, copy, or mandate. The Science authors say it plainly: "there is no single 'AI-proof' assessment model."

That conclusion follows directly from their data. If misuse runs at 17 percent in economics and 5 percent in biology, the exposure lives in the task, not in the technology. A take-home essay and a lab writeup fail in different ways, so a single institutional policy will be simultaneously too strict for one and useless for the other.

Which is why the three directions the researchers propose are worth reading closely:

1. Controlled testing environments, meaning pen, paper, and proctors.
2. Clearer guidelines about what counts as acceptable AI use.
3. Assessments that deliberately integrate AI, so learners demonstrate the professional capability of working with it.

Only the first is about catching people, and it is the one that scales worst. Proctoring a workforce is not an option, and detection tools invite a second problem the HEPI survey picked up: students already report anxiety about being falsely accused. An assessment regime that makes honest learners afraid has failed twice.

Why end-of-course quizzes stopped measuring skill

A multiple-choice quiz measures whether someone can produce the right answer, which is precisely the capability generative AI has commoditized.

The quiz was always a proxy. It worked because, for a century, producing the answer required knowing the answer. That link is now broken, and everything downstream of it, including the completion certificate, inherits the break.

Corporate training is quietly more exposed than higher education. A compliance module ends with ten questions, nobody is proctoring, the learner alt-tabs to a chatbot, and the dashboard reports 100 percent completion. The organization now holds documentation that proves nothing, which is worse than holding nothing, because it feels like assurance.

This is a different problem from proving your training program worked at the organizational level. This is about whether a single assessment is still a valid signal about a single person. Right now, for most formats, it is not.

Why assessment during learning beats testing after it

Testing is not only a measurement instrument. It is one of the most reliable learning interventions in the entire research literature, which means the fix for broken assessment is to stop treating it as the thing that happens after learning.

Retrieval practice, the act of pulling information out of memory rather than reading it again, produces more durable long-term retention than restudying. There is also a forward effect of testing: being tested on earlier material improves how well you learn new material presented afterward. In one classic experiment, learners quizzed on the first four word lists recalled roughly twice as much of the fifth list as learners who were never tested.

The most direct evidence for video-based learning comes from Szpunar, Khan, and Schacter, whose PNAS study interpolated short quizzes into an online lecture. They split a 21-minute statistics lecture into four segments. One group got a brief quiz after each segment. Others restudied the material or simply continued.

Results for the quizzed group:

• Mind wandering at 19 percent of probe points, versus 39 percent for restudy and 41 percent for controls.
• 90 percent on the final cumulative test, versus 68 to 76 percent for the comparison groups.
• Additional notes taken on 17 to 24 percent of slides, against 6 to 9 percent.
• Lower test anxiety, and lower perceived cognitive demand.

That last finding deserves a second read. The group that got tested more often found the experience less mentally taxing, not more. Frequent low-stakes retrieval does not add burden, it replaces the burden of trying to stay awake through a monologue.

The broader pattern holds well beyond video. A meta-analysis of 225 studies comparing active learning to traditional lecturing found exam scores rose about 6 percent, while failure rates fell from 33.8 percent under lecturing to 21.8 percent under active learning.

How to design AI-proof assessment that measures real skill

Design for observation of the process, not collection of the product. A finished artifact tells you nothing about who made it. A live exchange tells you almost everything.

Five principles, in the order you should apply them:

1. Move the questions inside the content. Checkpoint questions every few minutes, in the flow of the lesson, not a gauntlet at the end. This is the core mechanic behind interactive training videos, and it is the same mechanic Szpunar tested.
2. Ask for reasoning, not recall. "Why would this approach fail for an enterprise customer?" survives contact with AI far better than "Which of the following is a benefit of X?"
3. Make it adaptive. A fixed question bank can be pre-solved once and shared. A follow-up question generated from the learner's actual answer cannot be anticipated, because it did not exist until they answered.
4. Assess in scenario. Situational judgment and roleplay force application, and application is where borrowed answers come apart.
5. Let AI in where the job allows it. If the role involves working with AI, assess that. This is the third path the Science authors point to, and for most corporate roles it is the honest one.

Notice that none of these require surveillance. They work by making the assessment something that has to happen to the learner, in real time, rather than something they hand in.

What this means for onboarding and compliance training

Replace the terminal quiz with continuous comprehension checks, and record what the learner struggled with rather than only what they scored.

• Onboarding: ask new hires to explain the process back, mid-lesson, in their own words. Ramp time becomes visible weeks before the first performance review.
• Compliance: present the ambiguous scenario, not the definition. Nobody violates a policy because they could not recall its title.
• Skills training: have the learner teach the concept to the tutor. Explanation exposes understanding that recognition hides.
• Reporting: a completion percentage is an attendance record. The questions a learner missed, the concepts that needed re-explaining, and the point at which they finally got it are the actual data.

FAQ

Can AI detection tools solve the assessment problem?

No. Detection is an arms race against tools that improve faster than the detectors, and false positives carry real costs: the HEPI survey found students already anxious about being wrongly accused. Redesigning the assessment removes the incentive entirely, which is a more durable fix than trying to police it.

Should employees be allowed to use AI during an assessment?

Usually yes, if they will use it on the job. The Science researchers explicitly recommend assessments that integrate AI so learners can demonstrate professional capability with it. The question shifts from "did you use AI" to "did you use it well, and can you defend the result."

How do you assess soft skills when AI can write the answers?

Put the learner in the interaction rather than asking them to describe it. A written answer about handling an upset customer is easy to generate, while a live roleplay that adapts to what you just said is not. Real-time conversation is the assessment.

The takeaway

The reason there is no AI-proof assessment format is that assessment was never supposed to be a format. It was supposed to be a moment of genuine retrieval, and generative AI has simply exposed how few of our moments qualified. Build the questions into the learning, make them adapt to the person answering, and you get an assessment that resists faking for the same reason it teaches: because the learner has to actually think, right now, in front of you.