Skip to content

AI & Assessment

The Two Numbers Every Assessment Leader Should Be Watching

By Veton KrasniqiAI Product Owner, EdTechAugust 3, 2026

An 85% rise in performance and a 25% decline under proctored conditions reveal the gap between AI-assisted output and retained student capability.

Two findings now belong at the center of higher education's assessment debate: performance rose on AI-susceptible work when assistance was available, while demonstrated retention fell when that assistance disappeared. The distance between those numbers is more important than either number alone.

There are two numbers that should be sitting in the middle of every higher education assessment conversation right now, and most institutions are only looking at one of them.

The first number is 85. Since ChatGPT's release, college student performance on AI-susceptible math problems has risen roughly 85% in non-proctored settings. The second number is 25. On the same problem types, under proctored conditions, performance has declined 25%.

The instinct is to treat those numbers as a contradiction, or worse, to pick the flattering one and move on. They are not a contradiction. They measure two different things. One measures what a student can produce when AI is sitting in the next tab. The other measures what the student retained when it is not. The distance between them is the story, and it is one most non-proctored grades are structurally incapable of telling.

Where these numbers come from

This is not a survey, and that distinction matters more than it sounds. Most of what we think we know about how students use AI comes from asking them, and students consistently under-report AI use because nobody enjoys admitting they skipped the hard part. Self-report is a soft foundation for a hard decision.

The evidence here is behavioral. Researchers from UC Irvine and McGraw Hill (Rismanchian et al., arXiv:2605.21629) analyzed 3.2 million real learning interactions on the ALEKS math platform, spanning a full decade from 2015 to 2025, along with 12.2 million placement-assessment response times. Instead of asking students what they did, they watched what students actually did, quarter after quarter, at population scale.

The design is the clever part. The researchers split problems into two categories. AI-susceptible problems are text-based word problems whose full text can be pasted straight into a chatbot. AI-resistant problems are graph-based and interactive, requiring manipulation of the platform itself, and cannot be transcribed into a prompt. Same students, same course, same quarter. The only difference is whether the problem can be handed to a machine.

After ChatGPT's release, college students spent 26.9% less time on the AI-susceptible problems relative to the AI-resistant ones, a decline that accumulated across eleven post-ChatGPT quarters. High school students showed an even sharper 31.3% drop. Grade 5 students, who have little autonomous access to these tools, showed no detectable change at all. The effect scales with the ability to reach for AI, which is exactly what you would expect if AI is the cause.

Why the proctoring result is the one that should keep you up at night

Here is the finding that turns an interesting behavioral study into an institutional problem. When these same students are later assessed on the same AI-susceptible material under proctored conditions, where AI is simply not available, their odds of answering correctly have fallen 25% cumulatively.

And the behavioral shift that shows up everywhere else? Under proctoring, it vanishes completely. When students are supervised, they engage with AI-susceptible problems at normal rates. The post-ChatGPT slope becomes statistically indistinguishable from zero.

That single fact rules out the comfortable explanations. Not a harder curriculum. Not students getting generally lazier. Not an efficiency gain where AI simply lets students learn the same thing faster. If any of those were true, the effect would not switch off the moment a proctor walks in. The effect is specific, it is conditional, and it points at one mechanism.

The researchers call that mechanism cognitive surrender. Students have not become more efficient. They have learned to route around the cognitive effort that actually builds durable knowledge. The problem is that the effort was never the obstacle to learning. The effort was the learning.

Think of it like physical training. If a machine lifts the weights for you, your numbers on paper look incredible. But the entire point of lifting was the strain, and the strain is the one thing you outsourced. You get the record without the muscle.

The validity problem hiding inside the integrity problem

For years, the AI conversation in higher education has been framed as an integrity question. Did the student cheat? Can we catch them? Detection tools, probability scores, flags. The framing assumes a discrete event, a moment where a specific student crossed a specific line.

This study reframes the problem, and the reframe is harder to solve. What the data describes is not a cheating event. It is an ambient, population-wide drift in how students allocate cognitive effort. There is no single line anyone crossed on a single night. There is just a slow, quarter-over-quarter shift that no detector was built to see.

That turns an integrity question into a validity question, and validity is the deeper one. When a non-proctored grade on AI-susceptible material can rise while the underlying knowledge falls, the grade stops measuring what we think it measures. GPA, course credit, and eventually degree attainment in AI-susceptible domains start to drift away from the thing they are supposed to certify.

This is the part people keep skipping. You can have a course with rising homework scores, satisfied students, and a widening hole in what those students can do without assistance. Everything on the dashboard looks fine. The dashboard is measuring the 85 and cannot see the 25. If most of your assessments are non-proctored, you are only seeing the 85. The 25 is still there. You are just not measuring it.

What this means for how we build assessment

The natural reaction is to reach for more surveillance, and that instinct misses the lesson. Notice what actually happened in the proctored data. Under proctoring, students did not just score differently. They behaved differently, engaging with the AI-susceptible problems at normal rates again. The secure condition did not merely catch bypass after the fact. It changed the behavior in the first place. The value of a secure environment is not that it polices students. It is that it restores the conditions under which learning behavior is real.

The implication for the people making the decisions

If you are a provost or a dean, the headline is not comfortable. Grade inflation by AI is now measurable at scale and concentrated in AI-susceptible domains. Students can earn higher non-proctored grades while retaining less. That is a validity issue that touches the credibility of the credential itself, and it will not fix itself by waiting.

If you are an assessment leader, the practical move is an audit. The AI-susceptible versus AI-resistant distinction is a usable lens today. Problems that require visual interpretation, multi-step manipulation, or context-specific reasoning are far harder to hand to a chatbot, and they were never designed as anti-cheating features. They were designed as learning features. The integrity benefit is a side effect of good pedagogy.

If you are faculty, the uncomfortable truth is that homework and unproctored practice are now inflated signals of understanding in AI-susceptible material. That does not make them useless. It makes them incomplete. The distance between what a student turns in and what a student can do unassisted is a 25% gap, and it is growing.

One caution, because credibility depends on it. This is a preprint, peer review is still in progress, and it was conducted on math content. Whether the distinction translates cleanly to writing, the lab sciences, or the humanities is an open question. The study validates the case for secure assessment. It does not validate any particular product. The direction of the finding, though, is consistent with earlier work at smaller scale, and the methodology is strong enough that the core result is unlikely to reverse.

The number that actually matters

Come back to the two numbers, because the whole thing lives in the space between them. The 85 is what AI can produce for a student. The 25 is what the student walks away with. When those two numbers move in opposite directions, the grade is telling you a story about the tool, not the person.

The question at the center of all of this is not whether students are using AI. That question is settled. The real question is whether we are assessing the AI or the student. For any domain where the answer is drifting toward the AI, the fix is not a better detector chasing a moment that already passed. The fix is to rebuild the conditions where effort still happens: secure the exam so it measures real knowledge, protect the productive friction during practice, and make the process visible enough that faculty can teach into the gap instead of grading over it.

The effort was never the problem to be optimized away. The effort was the point. Everything we build should be in service of putting it back.

What I would measure next

I would start by auditing which assignments are AI-susceptible, where independent capability still matters, and whether current grades can distinguish assisted output from retained knowledge. The aim is not to remove every tool. It is to preserve the productive effort that makes learning durable and to design assessments that still mean what institutions say they mean.