AI Overreliance in Training: How to Keep Learners Sharp

Ninety-eight German ninth graders sat down with an AI math tutor and sent it 808 messages. They evaluated what it told them in 3.4 percent of those messages. Their test scores fell from 67.5 percent before the session to 56.9 percent after it.

That is AI overreliance, and it is the failure mode almost nobody is measuring. The learners were engaged. They were asking questions. The tool worked exactly as designed. And they came out knowing less than when they walked in, because asking an AI is not the same as learning from it.

Below is what the newest research says about why learners stop checking AI's answers, why this happens to experienced professionals just as readily as to teenagers, and the specific design choices that turn an answer machine back into a learning experience.

What is AI overreliance in learning?

AI overreliance is when a learner accepts AI output without evaluating it, offloading the thinking that the learning depended on. It is not laziness. It is a predictable response to a tool that removes friction from a process where friction was doing the work.

The University of Tübingen study makes the pattern painfully concrete. Researchers Rania Abdelghani, Peter Kaiser, and Kou Murayama gave a Mistral Large tutor to 98 students aged 14 to 15 across three public Gymnasium schools in Baden-Württemberg, then analyzed 1,616 chat turns during a curriculum-aligned math activity, as reported by EdTech Innovation Hub.

The behavior breakdown is the whole story:

• 3.4 percent of task-relevant messages evaluated the AI's response
• 5.7 percent monitored the student's own understanding
• 16.3 percent of requests asked the AI to verify anything

Here is the finding that should stop every L&D leader cold. 69.7 percent of these students said they wanted the AI to check their understanding. Their stated intentions did not significantly predict what they actually did. They knew what good learning looked like. Under the gentle pressure of a tool that would just tell them, they did not do it.

One caveat worth stating plainly: this is a work-in-progress paper accepted to the NextGen Learning Interfaces Workshop at AIED 2026, its behavioral coding is awaiting full human validation, and it has no control group. Treat the 3.4 percent as a vivid signal, not a settled law. What makes it credible is that it is not alone.

Why learners stop checking AI answers

Learners stop checking because ease feels like understanding, and our brains have no reliable way to tell the two apart. This is the single most important idea in learning science, and it was described decades before ChatGPT existed.

Robert and Elizabeth Bjork call it the distinction between learning and performance. In their work on desirable difficulties, they put it this way: conditions of learning that make performance improve rapidly often fail to support long-term retention and transfer, whereas conditions that create challenges and slow the rate of apparent learning often optimize long-term retention and transfer.

Rereading a chapter, they note, produces a "sense of familiarity or perceptual fluency that we interpret as understanding or comprehension," but which may be nothing more than low-level priming. An AI that answers instantly is a fluency machine. It manufactures the exact sensation of having learned, at scale, on demand.

The Bjorks also named the antidote. The generation effect is the long-term benefit of generating an answer rather than reading it, and it rivals the spacing effect in generality. Every time an AI hands over a solution, it deletes a generation opportunity. Do that a few hundred times and you have a learner who feels fluent and tests poorly, which is precisely what those German students did.

Does AI overreliance affect professionals too?

Yes, and the mechanism is confidence. Microsoft Research and Carnegie Mellon surveyed 319 knowledge workers who use generative AI at work weekly, collecting 936 first-hand examples of real tasks, in a CHI 2025 paper by Lee, Sarkar, Tankelevitch and colleagues.

Their central quantitative finding is one sentence long and worth memorizing: higher confidence in GenAI is associated with less critical thinking, while higher self-confidence is associated with more critical thinking.

Read that again, because it inverts the usual procurement logic. The better your people think your AI tool is, the less they will scrutinize it. Trust in the tool and scrutiny of the tool are in direct tension, and no amount of training on "AI literacy" changes the underlying arithmetic.

Two more findings from the same paper matter for anyone rolling out AI at work:

• Workers reported enacting critical thinking in 59 percent of the examples they shared, and reported that critical thinking took less effort with AI than without it.
• The paper flags what researchers call "mechanised convergence": teams with AI tools produce a less diverse set of outcomes for the same task than teams without them.

Note the honest hedge the authors themselves make. Reduced perceived effort could mean workers are offloading thinking, or it could mean they are doing the same thinking while feeling supported. Output diversity is a proxy for critical thinking, and, in their words, a flawed one. The direction is clear even where the magnitude is not.

How AI overreliance shows up in real performance data

The clearest evidence that overreliance costs real skill comes from watching what happened to math students when ChatGPT arrived. Sina Rismanchian, a doctoral researcher at UC Irvine, worked with McGraw Hill to analyze ALEKS, an online math platform used by more than four million students a year, comparing the period before ChatGPT's launch through the end of 2025.

The design is elegant. Word problems are trivial to paste into a chatbot. Graphing problems require a screenshot and a manually recreated graph, so they resist offloading. If AI is the cause, only the word problems should move.

They did, according to reporting in the Hechinger Report:

• High schoolers spent 31 percent less time on word problems after ChatGPT, dropping from about four minutes to under three. College students spent 27 percent less. Fifth graders, who mostly were not using it, barely changed.
• On proctored placement tests, where no AI was available, performance on word problems fell from roughly 80 percent correct to about 60 percent, a 25 percent reduction in the odds of a correct answer.
• Graphing problems showed no decline at all.

Students got faster at the thing AI could do for them and worse at doing it themselves. Rismanchian's warning generalizes past math: "This cognitive surrender might be going on in writing, science, everything." This is a June 2026 preprint and has not yet been peer reviewed, so hold it loosely, but the control built into the design is hard to argue with.

The habit is already ambient. Pew Research Center found that 54 percent of US teens use chatbots for schoolwork help and 10 percent complete all or most of their schoolwork with one. These are the people your onboarding program will meet in eighteen months.

How to design training that prevents AI overreliance

Prevent overreliance by building interaction that makes the learner produce before the system reveals. The fix is structural, not motivational. Those German students already wanted to check their understanding, and wanting it changed nothing.

1. Ask before you answer. The tutor's default response to a question should be a question: a hint, a smaller sub-problem, or a check for understanding. Answer-on-first-ask is an assistant. Question-first is a tutor.
2. Force generation before reveal. Make the learner commit to an answer, a prediction, or an explanation before the correct one appears. This is the generation effect, and it is the cheapest intervention in learning science.
3. Make verification a step, not a virtue. Do not hope learners will evaluate output. Only 3.4 percent did. Build a beat into the flow where the learner has to check, correct, or defend what the system just said.
4. Calibrate trust deliberately. Since confidence in AI predicts less scrutiny, have the system occasionally be wrong on purpose, or ask the learner to rate their confidence before revealing whether they were right. Miscalibration is the disease. Feedback is the cure.
5. Protect the difficulty that is doing the work. When a learner struggles to retrieve something, that struggle is the product functioning. Smoothing it away feels like better UX and is worse learning.

This is exactly what interactive training videos are built to do. Rather than a lecture a learner half-watches while a chatbot does the thinking in another tab, an interactive video with an AI tutor stops to ask, waits for an answer, adapts to what the learner actually said, and responds in real time. The learner cannot coast, because the experience will not advance without them.

How to measure whether learning actually happened

Measure retrieval under conditions where the AI is not available, because that is the only number that survives contact with reality. Every metric in the studies above split cleanly into two categories, and most training dashboards only track the wrong one.

Stop trusting these:

• Completion rates
• Time in the platform
• Self-reported confidence
• Satisfaction scores

Start tracking these:

• Unassisted retrieval. Can the learner produce the answer later, without the tool? The ALEKS proctored tests caught the decline precisely because the AI was gone.
• Transfer. Can they apply it to a problem they have not seen?
• Verification behavior. What percentage of AI responses does the learner question, correct, or push back on? Tübingen's 3.4 percent is a benchmark, and a low bar to clear.
• Confidence calibration. Does learner certainty track learner accuracy, or has fluency decoupled them?

A training program that raises completion while lowering unassisted retrieval is not neutral. It is negative, and it is invisible on a standard dashboard.

FAQ

Is AI overreliance just a fancy word for cheating?

No, and conflating them will lead you to the wrong fix. Cheating is a choice to bypass learning; overreliance is what happens when a learner genuinely tries to learn and the tool quietly does the cognitive work anyway. The Tübingen students were not cheating. They wanted the AI to check their understanding, said so, and then accepted its answers without evaluation because that is what the interaction invited.

If AI tools reduce critical thinking, should we keep them out of training?

That is the wrong conclusion to draw. The Microsoft and Carnegie Mellon researchers found that AI shifts critical thinking rather than simply deleting it, moving effort toward verifying information and integrating responses. The problem is not the presence of AI, it is AI configured to supply answers instead of demanding thought. An AI tutor that asks questions produces the opposite effect of a chatbot that answers them.

How do I tell if my current training tool is creating overreliance?

Run one test: take the AI away and measure. Give learners a problem from last month's module with no tool access and compare their performance to their in-module scores. If completion was high and unassisted performance is low, you have been buying fluency and calling it learning, which is exactly the gap the proctored ALEKS tests exposed.

The uncomfortable lesson across all of this research is that engagement, speed, and learner satisfaction can all rise while learning falls. The only reliable defense is interaction that refuses to let the learner be passive: a system that asks before it tells, waits for a real answer, and adapts to what that answer reveals. Build training that makes people generate, verify, and struggle a little, and you get knowledge that outlives the session.