Using AI tools likely erodes critical thinking skills when they are used to bypass effortful thinking, a mechanism called cognitive offloading, though current evidence shows a correlation rather than a definitive cause. The headlines are running with a simple story, but the reality for leaders is more nuanced: the problem isn’t the tool, it’s how we design the work around it.
A Correlation Is Not a Cause, But It’s a Clue
#A 2025 study in the journal Societies by Gerlich is making the rounds, reporting a strong negative correlation between frequent AI tool use and critical-thinking scores. Before we declare a cognitive apocalypse, let’s look under the hood. The author is commendably direct about the limits here: this is a correlation, not a proven cause. It is just as possible that people who already have weaker critical thinking skills are more likely to lean heavily on AI. That’s reverse causation, and it’s a classic trap.
The study also relies in part on self-reported assessments, which are notoriously tricky. So, as direct proof that your team’s use of a large language model is actively diminishing their intellect, this paper is shaky. It’s a single study in a field the researchers themselves call nascent. But dismissing it entirely would be a mistake. A correlation, even a fuzzy one, can point toward a real mechanism. And this one points at something we already know is real.
The Offloading Penalty: When ‘Helpful’ Becomes a Hindrance
#The plausible mechanism here is cognitive offloading. I’d define The Offloading Penalty as the measurable decay in a person’s retained skill and independent judgment that occurs when they consistently use a tool to get an answer without doing the underlying work to derive it. You’ve felt this yourself if you’ve ever outsourced your sense of direction to a GPS. After a decade of turn-by-turn directions, can you still navigate by landmark or dead reckoning? For most of us, that skill has atrophied.
This isn’t a new phenomenon, but generative AI puts it on an industrial scale. The tool is so fast, so fluent, that it becomes frictionless to offload the entire process of thinking, not just the tedious part. The result is an answer, but not the capability that comes from arriving at that answer yourself. For any leader who cares about the long-term skill of their team, this accumulation of capability debt is the real risk, far more than any single bad output from a model.
Why Effortless Answers Don’t Stick
#The Offloading Penalty isn’t just a folk theory. It’s the flip side of a mountain of research on how people actually learn and retain complex skills. Work by researchers like Roediger & Karpicke (2006) has shown for years that the act of effortful retrieval, pulling information out of your own head, is what makes memories durable. Simply rereading or being given the answer is dramatically less effective. This is called the testing effect, and it’s one of the most robust findings in cognitive psychology.
More recent work confirms the risk when it comes to AI. A 2025 study by Bastani et al. found that how an AI tutor is used matters immensely; learners who leaned on it passively performed worse once the tool was taken away. The AI helped them complete the assignment, but it hindered their learning. It removed the struggle, and the struggle is where the learning happens. Capability comes from the reps, the practice, and the desirable difficulty of doing the work. Smooth, frictionless answers produce smooth, frictionless forgetting.
How to Use AI Without Paying the Penalty
#For a manager or L&D leader, the takeaway from the Gerlich study isn’t to ban the tool. That’s a losing battle and a missed opportunity. The job is to design work and training that forces the right kind of effort.
Instead of asking your team, “Use the AI to find the answer to X,” try reframing the task:
- Draft, then refine: “Draft your own answer to X first. Then, use the AI as a sparring partner to critique and improve your draft.”
- Generate practice, not answers: “Use the AI to generate a dozen realistic practice questions about our new process. We’ll use them to quiz ourselves tomorrow.”
- Check, don’t create: “The AI has produced this summary. Your job is to fact-check it against the source documents and find the three places where it’s subtly wrong.”
Each of these approaches keeps the human in the loop, doing the effortful cognitive work that builds and maintains real skill. The AI becomes a tool for practice and metacognition, not a crutch that lets the underlying mental muscle atrophy. The Offloading Penalty is real, but it’s a tax you only have to pay if you choose to.
Sources
#- Bastani, H., et al. (2025). Generative AI can harm learning. PNAS.
- Gerlich, R. (2025). Cognitive Offloading in the Age of AI. Societies.
- Roediger, H. L., & Karpicke, J. D. (2006). Test-Enhanced Learning. Psychological Science.