Stanford's latest review of AI in education points to a distinction that schools cannot afford to overlook: producing better work with assistance is different from learning to do it independently. The implications reach beyond software choices to assessment, teaching and educational opportunity
A completed assignment is visible. The understanding developed while completing it is harder to see. Artificial intelligence makes that familiar educational problem more urgent because a fluent explanation, a polished paragraph or a correct answer can now be produced with little evidence of what the learner can do alone. A new Stanford SCALE review, published on 10 November, provides a timely basis for examining the distinction. Its causal evidence base has grown from 20 papers in March to 52. The update covers research published through June 2026, so its November release should not be confused with a test of every tool currently available.
The reviewers find a more promising, though still inconsistent, picture when students use AI to understand material rather than produce work for them. Evidence for broad conclusions about school-age learners remains limited: much of the research concerns postsecondary settings. The repository also includes preprints, and the review’s screening process has limitations. These qualifications are central to the news. They prevent a growing body of research from becoming an easy verdict that AI either improves education or inevitably damages it. The more useful question is what a particular arrangement asks learners and teachers to do, and whether it develops capacities that remain when the arrangement changes.
The difference between a useful answer and a learned skill
Consider an illustrative pupil faced with a difficult mathematical problem. An assistant could supply the solution, explain a relevant concept or ask the pupil to identify the next step. Each response could be helpful in a different sense, but the educational activity would not be the same. Receiving a finished solution might remove an immediate obstacle while also removing the opportunity to make and examine an attempt. A well-placed question might preserve that opportunity, although it could still be unhelpful if it assumes knowledge the pupil does not have. This is why the mere presence of conversational AI says little about the learning experience it creates.
The distinction also changes how success should be measured. If the objective is immediate completion, assistance can be judged alongside the work it helps produce. If the objective is learning, assessment must examine what the student can subsequently understand, explain or apply. These outcomes can diverge without any dishonesty by the learner. Someone may genuinely feel more capable while working with a responsive tool and still struggle when facing a new problem independently. A school therefore needs to decide which outcome an intervention is intended to improve. Otherwise, attractive results on assisted tasks can become evidence for a claim that those tasks were never designed to test.
An earlier experiment shows why the distinction matters
A 2025 field experiment by Hamsa Bastani and colleagues offers a concrete example. Nearly a thousand high-school mathematics students were provided with either a GPT-4 interface resembling a standard chatbot or a version designed with learning safeguards. Access improved performance during assisted problem-solving. When assistance was removed, the standard-interface group performed worse than students who had not received access; the learning-oriented version largely mitigated that negative effect. The experiment concerns particular tools, students and tasks. It should not be converted into a prediction that every chatbot will have the same effect in every classroom.
Its broader importance is the separation of two moments: supported performance and later independent performance. That separation gives schools a more informative way to investigate an AI intervention. An illustrative evaluation might ask students to work with support and then explain the underlying idea on a different task without it. The assessment would need to match what was taught and allow sufficient time for learning; simply making a test harder would not improve the evidence. The aim would be to discover whether assistance has helped the learner build something they can use, rather than merely helped them reach the end of one exercise.
The timing of help becomes a teaching decision
A possible response is to require an initial attempt before offering AI help. This is an instructional option to test, not a universal rule established by one experiment. Its rationale is that a first attempt can reveal a student’s interpretation of the problem and give subsequent feedback something specific to address. Yet insisting on prolonged unsupported effort could also leave a struggling learner confused. The relevant question is how to preserve meaningful participation while providing enough support to continue. A teacher might choose a small prompt, a worked example or a conversation instead of an automated response. The educational purpose should determine which form of help is appropriate.
This places design choices inside the teaching process. A system that immediately supplies a complete answer and one that asks a student to compare two explanations may use similar underlying technology while creating different tasks. Schools evaluating either system could observe how students actually use it, including whether they skip the intended learning steps. They could also ask what happens when the tool misunderstands the student’s difficulty. The point is not to imagine a perfect dialogue in advance. It is to investigate whether the real interaction supports the kind of thinking the lesson requires, and whether teachers can recognise when it fails to do so.
Teacher time is valuable, but it is a different outcome
The Stanford review also describes potential benefits for targeted educator tasks, including planning and feedback, alongside findings of unintended effects on students. Time saved by a teacher and learning gained by a pupil are related possibilities, not interchangeable measurements. An illustrative lesson plan drafted by AI might free time for a teacher to examine students’ misunderstandings. It might instead require extensive correction or make the lesson less responsive to the class. Those are possible mechanisms to investigate, rather than claims about the average effect of teachers using AI.
The professional work lies partly in deciding what to retain, change and discard. A teacher needs to connect a suggested activity to the curriculum, the pupils’ prior knowledge and the constraints of the classroom. This means evaluating AI support as part of a working arrangement, including the effort required to review it. If a tool shifts effort from drafting to verification, the change may still be worthwhile, but its value needs to be understood honestly. A useful school evaluation could examine both workload and instructional consequences. A reduction in preparation time should neither be dismissed nor allowed to stand as proof that students have learned more.
Assessment has to make thinking visible
Generative AI raises a practical problem for assessment because a finished product offers fewer clues about how it was produced. One response is to create more opportunities for students to explain decisions, compare approaches and revise work after feedback. These are proposals for making learning visible, not a claim that every assignment should become an oral examination. Different subjects and learners require different forms of evidence. A short explanation of why a method was chosen might reveal something that a polished final answer cannot. A sequence of drafts could show how a student’s reasoning changed, provided the process itself remains meaningful rather than a requirement to manufacture extra paperwork.
Such changes need care. Assessment can become more burdensome or intrusive if institutions respond to uncertainty by demanding constant proof of authorship. A better starting point is the learning objective: what should the student be able to do, and what evidence would reasonably show it? Some tasks may appropriately include AI because the objective concerns evaluating assisted work. Others may require independent performance because foundational knowledge is the point. Making those purposes explicit could reduce confusion for pupils and families. It would also help teachers explain why assistance is welcomed in one activity and restricted in another, without implying that the rules are arbitrary.
Educational opportunity depends on the support around the tool
The social stakes extend beyond whether a pupil has access to a chatbot. An illustrative pair of students could use the same application yet receive very different guidance about checking explanations, recognising uncertainty and deciding when to stop. One might have an adult available to discuss the response; the other might rely on it as the only accessible source of help. These examples are hypothetical, not reported cases from the review. They show why equal access to a product cannot by itself establish equal educational benefit. The surrounding support, the task and the learner’s existing understanding all deserve attention when institutions consider who might gain.
Schools could therefore examine the distribution of outcomes rather than rely solely on an average improvement. A pilot might ask whether an intervention helps pupils with different starting points, whether it creates new dependence on paid access or whether its language demands exclude some learners. Those questions would need suitable evidence and should not be answered by assumption. A promising result in one setting can justify further investigation while leaving transfer to another setting uncertain. The institutional challenge is to learn from research without allowing its most convenient headline to stand in for the experiences of the particular children a school serves.
Confidence and understanding need room to come apart
A responsive assistant may make a learning activity feel easier and more welcoming. That experience can matter even when it is not evidence of greater subject knowledge. A learner who feels able to ask a question might participate more readily; a learner who accepts a fluent answer too quickly might miss an important uncertainty. These are possible consequences to observe, not an assertion that confidence is inherently misleading. Education needs both the willingness to engage and the ability to judge an explanation. An evaluation that records only satisfaction would leave the second question unanswered, while an evaluation that ignores experience could miss why a tool is used or abandoned.
This suggests a more careful conversation with students about what assistance is doing. They could be invited to identify which part of an explanation they understand, what they would check and how they would approach a similar task without help. Such activities are illustrative ways of making self-assessment more concrete; they require adaptation to age and subject. The aim is to preserve the student’s role as an active participant in learning. A tool should not become the only judge of whether its own explanation has succeeded. Teachers and learners need opportunities to examine that success through independent evidence and discussion.
A growing evidence base should sharpen the questions
The 13 November Chalkbeat report on the review brings these findings into a live public debate about educational AI. The appropriate response is neither to wait for perfect certainty nor to treat adoption as proof of progress. Schools can choose bounded uses, explain their objectives and review the consequences. They can also preserve alternatives when an intervention is unsuitable. Research helps most when it improves those decisions: distinguishing the task a tool assists, the capacity it is intended to develop and the evidence needed to know whether that development has occurred.
The central social question is what education asks a person to become able to do. AI can change the resources available during that process, and it may support useful forms of practice or reduce some of teachers’ workload. The new synthesis shows why the effects must be examined in context rather than inferred from polished output. A finished assignment is an achievement of a kind, but it is not the whole purpose of learning. The more durable test is whether students gain understanding they can explain, question and carry into another situation, with the judgement to use assistance when it helps and recognise what they still need to learn.
Image credit: A Sideview of an Empty Classroom by Paul The Writer, CC0 1.0. Illustrative photograph unrelated to the studies discussed. Resized for this site; WordPress may crop display versions.
