What should students learn when AI can help with almost everything?

Separate a better answer from better learning

An assisted assignment can improve while the learner becomes no more able to solve the next problem. Conversely, a well-designed tutor can help a student understand something they struggled with in class. The relevant outcome is what the student can explain, check and transfer, as well as what they can produce with a tool.

This question builds on my submitted support for practical learning and human judgement. I haven’t settled on a particular school curriculum or a blanket rule about assessment.

Two experiments pointing in different directions

Bastani and colleagues’ PNAS study randomised access to two GPT-based tools among nearly 1,000 students at one Turkish high school over four 90-minute maths sessions. Basic GPT improved assisted practice, but subsequent unassisted grades were 17% lower in relative terms, not 17 percentage points. Teacher-designed tutor guardrails removed the deficit without establishing a statistically significant exam gain. This is short-term evidence in one setting, not proof of universal skill decline.

Kestin and colleagues’ physics study used a randomised crossover design with 194 Harvard students across two lessons. A purpose-built tutor produced larger immediate learning gains in less time than active-learning classes. That is a reason to test thoughtful instructional design. It is not evidence that any chatbot replaces a teacher or guarantees long-term transfer.

Together, these studies challenge both a blanket ban and the assumption that completing more assisted work means learning more. The differences in subject, students and tools prevent a simple pooled verdict.

What expertise evidence can and cannot show

Anthropic’s June 2026 usage analysis associates greater demonstrated task expertise with session success. Expertise and success are inferred from conversations and telemetry, not independently assessed qualifications or final business outcomes. This observational association does not establish whether manual practice caused the expertise, or how people should acquire it.

The useful question is how students learn to specify a goal, spot an error, evaluate evidence and recognise when they need another person’s help. Subject knowledge remains relevant to those tasks. How much to teach through unaided work, guided AI use or other methods should be evaluated rather than assumed.

Implications to test

An assessment could combine an AI-supported project with an explanation of the reasoning, a critique of a flawed output and an independent transfer task. This is a proposed evaluation design, not a validated formula. It should fit the learner’s age, subject and accessibility needs.

Australia’s schools framework, whose 2024 review was endorsed in June 2025, provides responsible-use guidance. Policy guidance helps organise decisions; it does not establish causal learning effects.

What would change this answer

Longer studies measuring retention, transfer and independent judgement across Australian subjects and student groups would matter most. Evidence that a particular form of assistance reliably undermines these outcomes should narrow its use. Evidence of durable gains should support expanding it, with attention to access and teacher workload.

Limits of this edition

The evidence does not settle the full curriculum or the best balance of aided and unaided practice. Promising immediate results and worrying immediate results both require careful follow-up.

Related question

When AI makes requests cheaper, who pays to answer them? applies the distinction between more output and a better result to complaints, applications and appeals. It asks students and SMEs to check facts, include evidence and make the recipient’s task easier.