Not another blog post about AI and education… No, I’m not trying to guess your thoughts, but it is what I sometimes think myself. Sometimes I feel like everything has already been said and written. Whether it be by AI or not (I wanted to insert an em dash here just for fun). But there is a new systematic review by Li and colleagues, published in Computers & Education: Artificial Intelligence. With this study, they attempt to move beyond those slogans by bringing together 67 studies from higher education. Their conclusion is both nuanced and actually not that surprising: ChatGPT does not work “by itself.” The effect depends mainly on how you use it. But read on anyway.

What makes this review interesting is that the authors explicitly distinguish between critical thinking and creative thinking. That sounds logical, but it happens remarkably rarely in AI discussions. Critical thinking is about analysing, evaluating, and arguing. Creative thinking is about generating ideas, combining them, and developing new perspectives. In theory, these can reinforce each other. In practice, this sometimes proves to be the case. But not always.

The review shows three recurring patterns.

First, critical and creative thinking can reinforce each other. This happens especially when teachers embed ChatGPT in well-designed learning activities: assignments that require reflection, comparing AI answers, argumentation, feedback loops, and explicit guidance. In those situations, students use AI less as an “answer machine” and more as a conversational partner or thinking tool.

But a second pattern also appears. In quite a few studies, ChatGPT supports creativity more easily than critical thinking. Students brainstorm faster, generate more ideas, and experience fewer creative blocks. At the same time, they often evaluate information less critically. They adopt answers more quickly, check less, and reason less deeply.

And then there is a third pattern: both forms of thinking decline. This happens mainly in weakly supervised contexts where students use AI primarily to finish tasks quickly. Researchers describe this as cognitive offloading: the thinking, and therefore part of the learning, shifts from the student to the system.

In itself, that aligns quite well with older insights from educational and cognitive science. Technology does not automatically make learning better or worse. The way tasks are designed remains crucial. A calculator can be a tool for complex thinking, but also a way to stop thinking altogether. The same apparently applies to generative AI.

It is also interesting that the review itself remains reasonably cautious. We know that many studies on AI are small-scale, short-term, and often rely on self-reports. This is the case here as well. Students, for example, say they work more creatively or reflect more effectively, but that does not automatically mean their skills improve sustainably. Strong causal evidence is still limited for the time being.

That makes this review precisely more credible than many spectacular AI claims. The authors do not act as if the debate is already settled. They primarily show patterns. And those patterns point in the same direction remarkably often, which largely aligns with what we already know about education in general. AI appears to be primarily an amplifier of existing teaching methods. In strong learning environments, it can help make thinking visible and iterative. In weak learning environments, it sometimes simply makes superficiality more efficient.

Or, more briefly summarised, ChatGPT probably changes less in what students think than in how and when they think.