AI Can Help Students Perform Better. But Are They Actually Learning?
Key Points for Readers on the Go
AI-supported performance is not the same as durable learning. Students may produce stronger work while using AI without necessarily developing independent knowledge or reasoning.
The strongest evidence is still limited. The 2026 Stanford review found no high-quality causal studies of student-facing AI tools conducted in U.S. K–12 school settings.
General-purpose AI tools can create learning risks. When a tool provides answers or performs too much of the thinking, it may reduce productive struggle and opportunities to practice reasoning.
Scaffolding matters. AI tools that provide hints, guided questions, feedback, and graduated support may be more promising than tools that simply complete tasks.
Schools should measure more than output quality. A better-looking product does not necessarily demonstrate stronger learning, transfer, metacognition, or independence.
This post continues my series examining findings from Stanford’s 2026 review, The Evidence Base on AI in K–12.
The first post provided a broad overview of what the evidence currently says about AI in K–12 education. The next examined where teacher-facing AI may be most useful, particularly when it reduces routine workload, supports instructional feedback, and strengthens rather than replaces professional judgment.
This post turns to the student side of the issue and focuses on one of the review’s most important distinctions:
AI can help students perform better while they are using it. But that does not necessarily mean they are learning more.
That distinction matters because schools have traditionally treated the quality of student work as evidence of student learning. A stronger essay, a more complete answer, or a correctly solved problem may appear to demonstrate increased knowledge or skill.
With AI involved, that inference becomes more complicated.
A student may produce better work because the student learned more. The work may also be better because AI supplied language, organized ideas, suggested reasoning, corrected errors, or completed portions of the task.
Both outcomes may have value, but they are not the same outcome.
The Core Problem: Tool-Supported Work Can Look Better
A great deal of student AI use will look successful at first glance.
The essay may be cleaner. The answer may be more complete. The student may finish faster. The explanation may sound more sophisticated.
Those improvements are not meaningless. AI may help students access immediate feedback, organize their ideas, clarify confusing material, overcome language barriers, or begin tasks that might otherwise feel inaccessible.
However, the Stanford review highlights an important concern: improvements that occur while students are using AI do not always persist when the support is removed.
In other words, AI may help a student complete the task in front of them without ensuring that the student has internalized the knowledge or skill needed to perform independently.
For educators, this creates an immediate assessment question:
Are we measuring what the student can do, or what the student can do with the tool?
Both may matter. Schools may reasonably want students to learn how to use AI effectively, just as they learn to use calculators, search engines, word processors, and other tools.
But AI-supported performance should not automatically be interpreted as evidence of independent mastery.
Supported Performance and Durable Learning
The distinction between performance and learning is not unique to AI.
Students often perform better when they receive prompts, examples, corrective feedback, graphic organizers, adult assistance, or other forms of support. Educators routinely distinguish between what a student can accomplish independently and what the student can accomplish with scaffolding.
AI introduces a particularly powerful and flexible form of support.
It can explain, summarize, draft, suggest, revise, translate, organize, and generate. It can respond immediately and adapt its output to the user’s request. That makes it useful, but it can also make it difficult to determine how much of the final performance reflects the student’s own understanding.
Durable learning requires more than successful task completion.
It means that students can retain knowledge, apply skills in new situations, explain their reasoning, recognize errors, and continue performing when the original support is no longer available.
The Stanford review found mixed results when students were later assessed without AI access. In some studies, students performed better while using AI but did not demonstrate comparable benefits on subsequent independent assessments. Some uses of general-purpose AI may also have weakened recall or reduced the quality of students’ independent reasoning.
This does not mean AI necessarily prevents learning.
It means that performance during AI use is not sufficient evidence that learning has occurred.
Why Easier Is Not Always Better
Students generally prefer tools that make difficult work easier. Adults do too.
Reducing difficulty can be beneficial when the difficulty is unrelated to the actual learning objective. AI may help students understand confusing directions, retrieve background information, organize ideas, or receive feedback more quickly.
For students with disabilities, language-related needs, executive functioning difficulties, or limited access to individualized support, well-designed AI assistance may improve access to instruction.
However, some difficulty serves an important purpose.
Learning often requires students to make decisions, try strategies, experience uncertainty, recognize mistakes, revise their thinking, and persist when an answer is not immediately available. These experiences are sometimes described as productive struggle because the effort contributes to learning.
If AI removes too much of that effort, students may lose opportunities to develop the very skills the assignment was intended to teach.
This is especially important for higher-order tasks such as:
Writing and revising an argument
Solving unfamiliar problems
Conducting scientific inquiry
Evaluating the credibility of sources
Comparing competing explanations
Selecting and applying strategies
Explaining one’s reasoning
Monitoring comprehension
Recognizing and correcting errors
These are not simply answer-production tasks. The process of reaching, evaluating, and revising an answer is part of the learning.
Scaffolding Is Different From Substitution
The practical question is not simply whether students are using AI.
A more useful question is: What part of the learning process is the AI performing?
Consider two different AI tools.
One receives a question and generates a complete answer.
The other asks the student what they have already tried, identifies the point of confusion, offers a limited hint, and then asks the student to explain the next step.
Both systems use AI, but they are not instructionally equivalent.
The first may substitute for student thinking. The second may scaffold it.
Strong instructional scaffolding does not eliminate every challenge. It provides enough support for the learner to make progress while preserving meaningful responsibility for the task. Ideally, that support is gradually reduced as the student becomes more competent.
A well-designed student-facing AI tool might ask:
What have you tried so far?
Which part of the problem is confusing?
What evidence supports your conclusion?
Can you explain why you selected that strategy?
Does the answer make sense in context?
How could you check your work?
What would you do differently next time?
These prompts preserve student participation in the reasoning process.
The AI is helping the student think rather than performing all of the thinking for the student.
The Risk of Metacognitive Outsourcing
One of the less visible risks of AI is that students may begin to outsource metacognition.
Metacognition includes recognizing what one understands, identifying gaps in knowledge, selecting an appropriate strategy, monitoring progress, evaluating whether a strategy worked, and changing approaches when necessary.
These skills are central to independent learning.
They are also closely connected to executive functioning, self-regulation, persistence, and academic problem-solving.
When AI immediately identifies the important information, chooses a strategy, evaluates the quality of an answer, and proposes the next step, the student may have fewer opportunities to practice those processes.
The resulting work may still look strong. The student may even feel that the material makes sense because the AI has produced a fluent and coherent explanation.
But fluency can create an illusion of understanding.
Students still need opportunities to determine whether they genuinely understand the content, whether the answer is accurate, and whether they could reconstruct the reasoning independently.
Student Work May No Longer Show What We Think It Shows
AI also changes how educators should interpret work samples.
Historically, a polished written product might provide information about a student’s vocabulary, organization, grammar, planning, written expression, content knowledge, and reasoning.
When AI contributes to the product, the work sample may reflect a combination of:
The student’s independent skill
The quality of the student’s prompts
AI-generated language or ideas
Automated editing and revision
Feedback supplied by the tool
The student’s ability to evaluate and integrate suggestions
That does not make the product invalid. It changes what the product can reasonably be said to measure.
For school psychologists and other educational professionals, this issue may affect the interpretation of classroom work, intervention data, curriculum-based products, and referral information.
A polished assignment may provide less evidence of independent writing ability than it once did. A completed task may reveal less about whether the student can plan, organize, persist, self-monitor, or solve problems without assistance.
Schools will need assessment practices that distinguish among:
What students can do independently
What students can do with ordinary instructional support
What students can do with AI assistance
How effectively students can evaluate and use AI-generated information
Each of these may be educationally relevant, but they should not be treated as interchangeable.
What Schools Should Ask Before Using AI With Students - The practical question is not whether students should ever use AI. The better question is how the tool is being used and what learning objective it is intended to support.
What skill is the student supposed to develop? - If the goal is to teach writing, reasoning, problem-solving, source evaluation, or self-monitoring, the AI should not perform all of those processes for the student.
Does the tool provide answers or scaffolding? - Tools that ask questions, provide hints, and offer graduated support may preserve more student thinking than tools that generate complete responses.
What happens when the tool is removed? - Students should have opportunities to demonstrate what they can do without AI assistance.
Is the AI reducing unnecessary barriers or productive struggle? - Reducing irrelevant confusion may improve access. Removing all cognitive effort may undermine learning.
How will educators know whether learning transferred? - Schools need assessment routines that examine retention, application, explanation, and independent performance—not merely the quality of the AI-supported product.
Does the student understand the AI’s role? - Students should be able to identify which parts of the work reflect their own reasoning, which parts were supported by AI, and how they evaluated the tool’s suggestions.
Is support being gradually reduced? - AI scaffolding should ideally help students become more independent rather than creating permanent dependence on the tool.
What This Means for School Psychologists
This distinction is particularly relevant to school psychologists because it connects directly to learning, assessment, intervention, disability identification, and equity.
AI may become a valuable scaffold for some students. Learners with writing difficulties, executive functioning challenges, language-related needs, or limited access to individualized support may benefit from tools that provide clarification, structure, examples, or immediate feedback.
But may benefit is not the same as will benefit.
Schools still need to evaluate whether the support improves access while preserving the targeted skill. They also need to determine whether students are becoming more independent over time or increasingly reliant on the tool.
School psychologists may need to help teams ask questions such as:
What ability are we attempting to assess?
What assistance was available when the student produced this work?
Can the student demonstrate the same skill without AI?
Is the tool functioning as an accommodation, an instructional scaffold, or a substitute for the targeted skill?
Does AI use change how progress should be monitored?
Are students receiving equitable access to high-quality, privacy-protective tools?
Could differences in AI access distort comparisons among students?
These questions do not require schools to reject student-facing AI.
They require schools to be more precise about what AI-supported performance means.
Bottom Line
The earlier posts in this series emphasized two broad conclusions from the Stanford review: the evidence base remains limited, and the effects of AI depend heavily on the tool, task, context, and way the technology is implemented.
The student-facing evidence adds another essential point.
AI can help students produce stronger work while they are using it. That can be useful, especially when the technology improves access, provides feedback, or offers carefully designed instructional scaffolding.
But the deeper question is whether students are developing knowledge and skills that last.
For schools, the goal should not be better-looking work alone. It should be stronger reasoning, improved transfer, greater independence, and a clearer understanding of what students can do both with and without technological support.
Student-facing AI should therefore be judged not only by what it helps students produce, but by what it helps them develop.
Source
Fesler, L., Martinez, J., Agnew, C., & Loeb, S. (2026). The Evidence Base on AI in K–12: A 2026 Review. AI Hub for Education of the SCALE Initiative, Stanford University.
AI Use Disclosure - I used AI to help draft and revise this blog post based on the source report. I reviewed and edited the content before posting.