AI learning and critical thinking

Does using AI help people learn?

The answer depends on what the system does and what the study measures. A tool can improve the work produced during a lesson without improving the learner’s later performance. It can also give someone patient, well-timed help they would otherwise struggle to obtain. Treating every chatbot as the same kind of tutor hides those differences.

In a high-school mathematics experiment, Bastani and colleagues found that unrestricted generative-AI assistance improved practice performance but could leave students performing worse when the tool was removed. A version designed to offer learning safeguards largely mitigated that harm. The study concerns a particular educational setting and tool design; it does not establish that AI makes a generation less intelligent. Bastani et al., PNAS, June 25, 2025.

A separate randomized study in undergraduate physics found greater learning gains with a carefully designed AI tutor than with the in-class active-learning lessons used for comparison. Its tutor incorporated teaching practices deliberately; its result is not a verdict on every chatbot, subject or classroom. Read alongside the mathematics study, it makes the design of assistance more interesting than a simple choice for or against AI. Kestin et al., Scientific Reports, June 3, 2025.

What is cognitive offloading?

Here, cognitive offloading means handing part of a thinking task to something outside ourselves: notes, a calculator, a colleague or an AI system. The useful question is which part has moved. Asking for a worked example, checking a calculation and allowing a system to choose the next task are different arrangements. Each can change what a learner practices and what remains their decision.

In this reading path, watch for the moment help becomes an assumption built into the day. If a school expects everyone to use a tutor, declining its help may carry a cost. If an employer treats faster understanding as a reason to assign more work, the benefit and the burden may reach different people. Those are questions the fiction explores, rather than outcomes proved by the studies above.

Read the human consequences

Questions to carry into the stories

• What can the person now do that they could not do before?
• Which choices still require their attention, and which arrive already made?
• What would count as evidence of learning after the assistance stops?
• Who gets the time or money saved, and who absorbs the extra work?

For another angle, The Last Original Thought follows a novelist confronting AI-generated work that resembles her own. Its questions concern authorship and meaning; they should not be mistaken for a finding about students’ learning.