AI Should Reduce Friction Around Learning, Not the Effort of Learning Itself
By Devon Wible, Vice President, Teaching and Learning, Catapult Learning
AI Should Reduce Friction Around Learning, Not the Effort of Learning Itself
AI can help educators plan, differentiate, analyze data, generate practice, cut documentation, and strengthen tutoring. But schools should not judge AI primarily by how much faster, easier, or more efficient it makes learning. Learning demands cognitive effort, and the central risk of AI in education is that it bypasses the very work students need to do to build knowledge, skills, independence, and judgment.
AI should reduce unnecessary friction around learning, not remove the cognitive effort of learning itself.
- Learning is effort. Students build lasting understanding through attention, retrieval, reasoning, explanation, feedback, revision, practice, and transfer. Tools that remove that effort may improve the product while weakening the learner.
- Younger and struggling learners carry the most risk. They need tailored support, strong relationships, explicit instruction, feedback, and productive struggle. AI should help teachers and tutors provide that, not replace them.
- The right near-term role for AI is adult-facing. Use it to help educators prepare, diagnose, scaffold, and document, so there is more time for instruction, not less.
- The model is not the product; the configuration is. The same AI model can be set up to hand over answers or to withhold them and coach. Buyers need to see the guardrails, not the demo.
Does this AI tool preserve, strengthen, or weaken the cognitive effort students need to learn?
What Is AI?
Artificial intelligence is a broad term for computer systems designed to perform tasks that typically require human intelligence: recognizing patterns, making predictions, generating language, summarizing information, recommending next steps, or responding to prompts.
The current debate in education is mostly about generative AI, tools that create text, explanations, questions, summaries, feedback, images, code, or lesson materials in response to a prompt. But AI in education is much broader than that single category. A general-purpose chatbot is different from an AI-powered instructional tool, which is different from a teacher-facing planning assistant, an assessment analysis tool, translation support, or an AI-powered platform that supports tutoring.
Educators need to distinguish among the AI tools in use and define clearly how each will be used. A recent paper from Instruction Partners shares learnings from a 2026 AI in Action Learning Tour, in which researchers studied 20 student-facing AI-powered products across 32 classrooms, 16 school systems, and seven states. Their findings reinforce that the instructional value of AI depends less on whether a product uses AI and more on what the product is designed to do, what role it gives the teacher, and what kind of thinking it asks students to do.21
So the first question for schools is not “Should we use AI?” The better questions are:
What work is the AI doing?
What work remains with the educator?
What cognitive effort is the student doing, and what cognitive effort does the tool remove?
Does the tool move students toward deeper learning and greater independence, or does it simply help them complete tasks more quickly?
How Generative AI Actually Works, and Why It Matters for Learning
To answer those questions well, leaders need a working understanding of what sits under the hood. Most of the tools now marketed to schools are built on large language models (LLMs), the same category of system behind ChatGPT, Gemini, Claude, and Copilot.
- An LLM is a prediction engine. It is trained on enormous quantities of text to estimate, one token (roughly, one word fragment) at a time, what is most likely to come next. A response is simply that prediction repeated until the model decides it is done.
- It is optimized for plausibility, not truth. The training objective rewards text that looks like what a person would write. Accuracy is a frequent by-product, not a guarantee. The model has no built-in check on whether a claim is correct.
- Confident errors are structural, not a bug that will be patched out. Researchers at OpenAI and Georgia Tech showed in 2025 that so-called hallucinations arise from how models are trained and evaluated: standard benchmarks reward a confident guess over an honest “I don’t know,” so models learn to guess.23
- Its behavior is shaped by instructions it is given, not just by the model. The same underlying model can be configured to hand a student the answer, or to withhold the answer and ask a question instead. That configuration layer, often called a system prompt or guardrails, is where most of the instructional design lives, and it is usually invisible to the buyer.
- It is non-deterministic. Ask the same question twice and you may get two different answers. That is a feature in creative work and a liability in instruction, where consistency and alignment to a scope and sequence matter.
Two implications follow. First, a student can only catch a fluent, wrong answer if they already know enough to check it. The learner who most needs support is the learner least equipped to evaluate what the tool produces, which is exactly why the knowledge-building work cannot be outsourced. Second, because the instructional behavior lives in the configuration rather than the model, “it uses AI” tells a leader almost nothing. What the tool is told to do, and what it is prevented from doing, is the whole question.
Not all “AI” is the same tool
The label covers at least five distinct categories, and they sit in very different places relative to the student’s cognitive work:
| Category | What it does | Where the cognitive work sits | Risk of shortcutting |
|---|---|---|---|
| General-purpose chatbot | Open-ended text generation on any topic (ChatGPT, Gemini, Copilot, Claude) | Wherever the student sends it. By default, the tool does the reasoning, drafting, and summarizing. | High |
| Student-facing AI tutor | An LLM wrapped in instructional prompts, content, and often voice or reading tools (e.g., Amira, Khanmigo) | Depends entirely on guardrail design. Can be Socratic or can be an answer machine. | Varies by design |
| Adaptive practice and recommendation engine | Predictive models that sequence items, adjust difficulty, and flag skills. Often not generative at all. | Student still does the item. The AI chooses what comes next. | Lower |
| Teacher- and tutor-facing copilot | Lesson planning, differentiation, scaffold generation, question banks, translation, progress notes | With the adult. The student never touches the tool. | Low |
| Assessment and data analysis | Pattern-finding across progress monitoring, diagnostics, and intervention data | With the adult. Informs instructional decisions. | Low |
The U.S. Department of Education’s developer guidance makes a related point: it asks product teams to design for teaching and learning, show evidence of impact, protect civil rights, ensure safety and security, and earn trust through transparency.25 Leaders can use that same list in reverse, as a set of demands.
The Purpose of School Is Learning, Not Task Completion
Schools exist to help students build the foundational knowledge, skills, strategies, and habits of mind needed to participate fully in life. That is why grade-level standards exist. Standards are not arbitrary checklists. They represent a cumulative vision of what students should know, understand, and be able to do over time.
Learning creates opportunity. Students learn to read with comprehension, write with clarity, reason mathematically, understand history and science, communicate, collaborate, question information, solve problems, and make sense of the world. These capacities open doors to college, careers, trade school, civic life, and personal independence.
AI undermines this purpose when it shortcuts students’ opportunities to develop these capacities independently. If a tool provides a summary, argument, answer, or explanation before the student has engaged in the necessary work, the student may finish the assignment without achieving meaningful learning.
This is not an argument against technology in education. Technology can expand access, support practice, improve feedback, enhance accessibility, and help educators better understand students’ needs. The concern is specific: certain types of AI, especially generative AI, can make it easier for students to bypass the cognitive work that school is designed to cultivate.
Supports learning
- A calculator, once students understand the operations it performs
- A spell-checker, once the writer can compose and revise
- A tool that helps a teacher spot a misconception, prepare a scaffold, or design better practice
Short-circuits learning
- A tool that produces the reasoning, structure, language, summary, or answer before the student has developed those capacities
- A tool that repeatedly supplies the writing, explanation, or answer students are supposed to develop
The issue is not whether a tool is digital, but whether it supports or short-circuits the work of learning. AI should not do the learning for students. It should help adults design better conditions for students to learn.
AI literacy matters, but it does not require integrating chatbots into every platform students use, or inserting AI into every interaction. Students can learn how AI works, where it fails, how bias appears, how privacy is affected, and why human judgment matters without outsourcing reading, writing, reasoning, questioning, or problem-solving to AI.
Learning Requires Productive Cognitive Effort
Learning is not the same as completing work. Students learn when they connect new information to what they already know, reflect on its meaning, practice over time, retrieve information from memory, receive feedback, correct errors, and apply their learning in new contexts.
Deans for Impact’s The Science of Learning summarizes these principles clearly. Students learn by connecting new knowledge to prior knowledge. Working memory is limited. Long-term learning requires meaningful processing. Practice should be effortful, spaced, varied, and paired with feedback. And students need to understand that learning requires effort and challenge.1
That does not mean school should be needlessly difficult. Poorly designed tasks, inaccessible materials, unclear directions, excessive cognitive load, or unsupported frustration do not improve learning. Cognitive load theory reminds us that working memory is limited and that instruction should minimize unnecessary or extraneous load.2
But productive cognitive effort is different. It is the mental work of making meaning, retrieving, explaining, organizing, revising, comparing, applying, and transferring. Learning science gives us a useful distinction: not all difficulty is beneficial, but some effort is necessary. Research on desirable difficulties shows that conditions that feel harder in the short term can strengthen long-term retention and transfer.3
Retrieval practice. Students often learn more when they actively retrieve information from memory than when they reread, rewatch, or receive the information again.4
Self-explanation. Students deepen understanding when they explain examples and reasoning to themselves rather than merely following worked steps.5
Productive failure. Under the right conditions, students benefit from attempting challenging problems before formal instruction, especially when that effort is followed by consolidation and explicit teaching.6
The implication for AI. Do not ask only whether AI makes a task easier. Ask whether it preserves the essential effort: retrieving, reasoning, generating, explaining, revising, and applying.
Scaffolding Should Build Independence, Not Dependence
Learning is not a binary choice between unsupported struggle and immediate assistance. Effective instruction falls between those extremes.
Vygotsky’s Zone of Proximal Development describes the space between what a learner can do independently and what they can do with support.7 This is where skilled teaching and tutoring matter most. The adult notices what the student can almost do, provides the appropriate scaffold, and helps the student move toward independence. Wood, Bruner, and Ross described scaffolding in tutoring as the support that enables learners to solve problems they could not yet solve on their own.8 The key is that scaffolding is temporary and purposeful. It supports the learner’s participation while preserving the learner’s role in the task.
The gradual release of responsibility model makes the same point in instructional terms. Pearson and Gallagher described the shift from teacher responsibility to student responsibility, and Fisher and Frey later helped popularize the model in classroom practice.910
The direction matters. Support should build capacity, not dependence. Tools that help a teacher identify a misconception, prepare a scaffold, or design better practice may strengthen instruction. A tool that repeatedly provides the reasoning, writing, explanation, or answer students are supposed to develop undermines the learning process and its goal.
AI should help educators scaffold learning, not short-circuit it.
The AI Risk Is Shortcutting
AI is powerful because it reduces effort, generating answers, explanations, summaries, drafts, translations, examples, and feedback almost instantly. That is useful in many adult workflows. For students, especially developing learners, the same capability can pose a risk.
Daniel Willingham has argued that students should use AI only for tasks they already know how to do well.11 That principle matters because many school tasks are not assigned solely for the final product. They are assigned because of the mental work students do while producing it. Natalie Wexler makes a similar point: educators need knowledge of cognitive science to make wise decisions about AI, because learning requires cognitive effort.12
A study of generative AI use among Chinese secondary students found that AI adoption raised homework scores and cut completion time, yet later exam performance declined.13 The context differs from U.S. classrooms, but the warning holds: short-term productivity gains can mask weaker long-term learning.
Robinson, Gormley, Ribeiro, and Loeb found access alone was insufficient. Human support increased engagement, but usage stayed low and reading achievement did not improve.16
Bastani and colleagues gave Turkish 9th–11th graders one of two tools built on the same model. “GPT Base” behaved like ChatGPT. “GPT Tutor” was prompted to give hints and withhold full solutions. Students with the unguarded tool did better on practice and worse on the exam, which the authors attribute to using the tool as a crutch. The guardrails largely erased the harm.22 The lesson for buyers: the model is not the product. The configuration is.
Process data points the same way. A 2025 study in the British Journal of Educational Technology compared students writing with ChatGPT, with a human expert, or with a checklist. The ChatGPT group showed fewer of the metacognitive moves that drive learning, such as planning and evaluating their own work. The authors call the pattern “metacognitive laziness”: learners offload the self-monitoring that the task was designed to build.24
This is not only a concern for struggling learners. For students who are behind, the risk is that AI becomes a thin substitute for human instruction, scaffolding, feedback, and relationships. For high-performing students, the risk looks different: AI becomes a tool for keeping up, polishing work, producing more, and appearing successful while bypassing the slower cognitive work that builds depth. In Never Enough, Jennifer Breheny Wallace describes the pressure many high-achieving students feel to perform, optimize, and appear exceptional.14 In that context, AI may become another way to meet relentless demands without developing the understanding, resilience, and judgment those assignments were meant to cultivate.
Students who are behind often need more than explanations. They need diagnosis, explicit instruction, feedback, encouragement, language support, confidence-building, and a trusted adult who can interpret hesitation, avoidance, frustration, and partial understanding. They need someone who can decide when to reteach, scaffold, press, or pause.
Research on AI tutoring suggests caution. Stanford’s National Student Support Accelerator has emphasized that “AI tutoring” is not a single model: human-led tutoring supported by AI differs from AI-led or AI-only tutoring. NSSA’s 2026 brief concludes that current evidence supports AI tools that enhance tutor effectiveness and educator capacity, rather than AI as a replacement for high-impact tutoring.15 Recent reporting from The 74 makes the issue concrete: states are investing in AI tutoring tools such as Amira and Khanmigo, while researchers caution that evidence at scale remains limited and student uptake remains a major challenge.17
This is an equity issue. If well-supported students receive human-rich instruction while struggling learners receive only automated support, technology will not close the gaps. It will reinforce them.
A Coherent Vision Must Come Before Scale
The field currently lacks a coherent, evidence-based vision for AI in schools. Much of what appears inevitable is actually market momentum. Education leaders should be cautious about adopting AI simply because vendors have made it ubiquitous. The question is not “How much AI can we integrate?” but “What learning purpose does this serve, and what cognitive work does it protect?”
Some school systems are beginning to draw clearer lines around student-facing AI:
NYC Public Schools announced a one-year moratorium on student-facing generative AI tools for grades 2K–8 for the 2026–27 school year.18
LAUSD, the nation’s second-largest district, restricted student access to generative AI platforms for 2026–27 while it develops policy recommendations and guardrails.19
These moves by the nation’s two largest districts are policy signals, not blanket rejections of AI. Both are slowing student-facing use while clarifying how AI should operate within their systems.
The developmental boundary extends beyond the earliest grades. The concern is strongest in elementary and middle school, where students are still building foundational knowledge, language, attention, reasoning, and academic habits. But it does not disappear in high school. Older students also need protected opportunities to learn the processes of writing, research, scientific investigation, mathematical reasoning, revision, collaboration, and complex problem-solving. When the learning goal is the process itself, AI should not bypass it. A high school student learning to write a research paper, conduct a science lab, analyze evidence, or build an argument needs to practice the thinking, not simply produce the artifact.
Teaching students about AI is different from putting AI directly in front of developing learners as an instructional tool. Schools can build AI literacy without making AI the default mediator of learning.
The Right Role for AI Is to Support Educators
Language has shifted toward viewing teachers as guides, facilitators, or motivators while technology delivers instruction. That framing is too thin.
Teachers are trained educators. They understand standards and content, child development, classroom culture, assessment, motivation, misconceptions, scaffolding, feedback, and the science of learning. Their work is not to keep students engaged while a platform delivers content. They design and guide a learning process that helps students build knowledge, develop skills, ask better questions, reason more clearly, and become more independent over time.
Students also learn powerfully with and from one another when an educator designs and oversees the work. Discussion, collaboration, peer explanation, debate, shared problem-solving, and feedback all require human judgment and social interaction. These are not secondary to learning. They are integral to it. Teachers do this more effectively than computers. Students, guided by skilled educators and working with other students, do this better than technology.
The most responsible near-term role for AI is adult-facing and instruction-supporting. AI can help teachers and tutors:
These are meaningful efficiencies. Tutoring and intervention at scale are challenging. Programs must train tutors, align instruction, monitor progress, communicate with schools and families, and continuously improve. AI can reduce friction in that system.
But the purpose of this efficiency must be clear: more time for instruction, not less. Better human preparation, not a weaker human presence. Stronger scaffolding, not automated substitution.
Seven Adoption Principles
Questions to put to any AI tool before it enters your schools:
- Center cognitive effort
Ask whether the tool increases or decreases the cognitive effort required to achieve the learning goal.
- Reduce unnecessary friction, not crucial thinking
Use AI to remove administrative burdens, accessibility barriers, and poor workflow design for adults. Do not use it to remove retrieval, reasoning, explanation, revision, or productive struggle for students.
- Preserve gradual release through human-centered instruction
Teachers and tutors should own the instructional arc from modeling to guided practice to collaboration to independent application. AI may help design scaffolds, examples, practice, and checks for understanding, but it should not determine the release of responsibility or provide the reasoning students are meant to develop.
- Keep educators in the instructional loop
Teachers and tutors should interpret AI outputs, validate recommendations, and make final instructional decisions.
- Measure learning, not just completion
Track durable outcomes: retention, transfer, explanation quality, independent performance, confidence, and growth. Do not rely only on usage, speed, engagement, or task completion.
- Protect foundational learning, deep process, and developing independence
Use student-facing AI with particular caution in elementary and middle school; with students receiving intervention, special education, or foundational skill support; and in any task where the goal is to develop the process of thinking, writing, researching, investigating, revising, reasoning, collaborating, or problem-solving. This includes advanced students, who may be especially tempted to use AI to optimize output without engaging in the deeper work.
- Interrogate vendor incentives
As Natasha Singer has reported, major technology companies are shaping school AI adoption through training, curriculum, advocacy, philanthropy, and platform ecosystems.20 Schools need a clear instructional north star before vendors define it for them.
- Which model, and whose? Name the underlying model and provider. What happens to the product when that provider changes or deprecates it?
- Show us the guardrails. What is the tool instructed to do, and what is it prevented from doing? Can it give a full answer or final draft? Can the district configure that by grade, subject, or task?
- Where does the student’s thinking happen? Walk through one assignment end to end. At each step, who is retrieving, reasoning, drafting, and revising: the student or the model?
- Can the teacher see everything? Full transcripts, flagged misconceptions, and time-on-task should be visible to the educator, not just an engagement dashboard.
- What happens to student data? Is any student input used to train or fine-tune models? What is retained, for how long, and who can delete it? How does the product meet FERPA and COPPA obligations?
- How are wrong answers handled? What is the measured error rate on your content, and what does the student see when the model is uncertain?
- What is the evidence? Ask for outcomes on retention and independent performance, not usage or satisfaction, and ask who conducted the study.
AI may become a valuable part of education, but education leaders need to be precise about the kind of value they seek. The promise of AI should not be that it makes learning easier for students. It should be that it helps adults design and safeguard the conditions for deep learning.
Learning requires effort. Not pointless struggle. Not confusion for its own sake. But the productive cognitive effort of retrieving, reasoning, explaining, practicing, revising, and applying. That effort is not a flaw in the learning process. It is the learning process.
AI should not do the learning for students. It should help adults design better conditions for students to learn. That is the line schools should take.
Sources
- Deans for Impact, The Science of Learning, Second Edition. deansforimpact.org
- Sweller, J. (1988). Cognitive Load During Problem Solving: Effects on Learning. Cognitive Science, 12, 257–285. doi.org/10.1207/s15516709cog1202_4
- Bjork, E. L., & Bjork, R. A. (2011). Making Things Hard on Yourself, But in a Good Way: Creating Desirable Difficulties to Enhance Learning. bjorklab.psych.ucla.edu
- Roediger, H. L., III, & Karpicke, J. D. (2006). Test-Enhanced Learning: Taking Memory Tests Improves Long-Term Retention. Psychological Science, 17(3), 249–255. doi.org/10.1111/j.1467-9280.2006.01693.x
- Chi, M. T. H., Bassok, M., Lewis, M. W., Reimann, P., & Glaser, R. (1989). Self-Explanations: How Students Study and Use Examples in Learning to Solve Problems. Cognitive Science, 13(2), 145–182. doi.org/10.1207/s15516709cog1302_1
- Kapur, M. (2008). Productive Failure. Cognition and Instruction, 26(3), 379–424. doi.org/10.1080/07370000802212669
- Vygotsky, L. S. (1978). Mind in Society: The Development of Higher Psychological Processes. Harvard University Press. hup.harvard.edu
- Wood, D., Bruner, J. S., & Ross, G. (1976). The Role of Tutoring in Problem Solving. Journal of Child Psychology and Psychiatry, 17(2), 89–100. doi.org/10.1111/j.1469-7610.1976.tb00381.x
- Pearson, P. D., & Gallagher, M. C. (1983). The Instruction of Reading Comprehension. Contemporary Educational Psychology, 8(3), 317–344. doi.org/10.1016/0361-476X(83)90019-X
- Fisher, D., & Frey, N. (2007). Checking for Understanding: Formative Assessment Techniques for Your Classroom. ASCD. ascd.org
- Daniel T. Willingham, Students Should Only Use AI For Things They Already Know How To Do Well. dtwuva.substack.com
- Natalie Wexler, How Should AI Be Used in Schools? nataliewexler.substack.com
- David Stromberg, Victor Lei, and Yanhui Wu, The Generative AI Learning Penalty: Evidence from Chinese Secondary Education, CEPR Discussion Paper No. 21577, June 2, 2026. cepr.org
- Jennifer Breheny Wallace, Never Enough: When Achievement Culture Becomes Toxic and What We Can Do About It. Portfolio/Penguin, 2023. penguinrandomhouse.com
- National Student Support Accelerator, AI Tutoring is Not a Monolith: What We Actually Know, August 20, 2026. nssa.stanford.edu
- Carly D. Robinson, David Gormley, Ana Trindade Ribeiro, and Susanna Loeb, Access is Not Enough: Human Support Improves Engagement with AI Tutoring, EdWorkingPaper No. 26-1451, June 2026. edworkingpapers.com
- Linda Jacobson, AI Tutors Not Yet a Replacement for Humans, Research Says, The 74, September 1, 2026. the74million.org
- NYC Public Schools, Guidance on Artificial Intelligence. schools.nyc.gov
- Anna Merod, LAUSD Restricts All Students From Using AI Tools, K-12 Dive, September 3, 2026. k12dive.com
- Natasha Singer, How Big Tech Captured American Schools, The New York Times, August 23, 2026. nytimes.com
- Instruction Partners, AI in Action Learning Tour: A Practical Look at AI’s Role in Instruction, September 2026. ailearningtour.instructionpartners.org
- Bastani, H., Bastani, O., Sungu, A., Ge, H., Kabakcı, Ö., & Mariman, R. (2025). Generative AI Without Guardrails Can Harm Learning: Evidence from High School Mathematics. Proceedings of the National Academy of Sciences, 122(26), e2422633122. papers.ssrn.com
- Kalai, A. T., Nachum, O., Vempala, S. S., & Zhang, E. (2025). Why Language Models Hallucinate. arXiv:2509.04664. arxiv.org/abs/2509.04664
- Fan, Y., Tang, L., Le, H., et al. (2025). Beware of Metacognitive Laziness: Effects of Generative Artificial Intelligence on Learning Motivation, Processes, and Performance. British Journal of Educational Technology, 56(2), 489–530. arxiv.org/abs/2412.09315
- U.S. Department of Education, Office of Educational Technology (2024). Designing for Education with Artificial Intelligence: An Essential Guide for Developers. files.eric.ed.gov/fulltext/ED661949.pdf
Ready to Learn More?
Catapult Learning is proud to offer an ESSA tier-2 validated tutoring program. Through our commitment to strong partnerships and seamless implementation, you can rest assured that your student’s will receive the support you need. It is why we see results such as 100% of struggling readers improving at least one grade level after our implementation.
Take the next step, and visit our High-Impact Tutoring page to see our model in action, read more success stories, and to contact us for a free consultation. Together, we can achieve brighter outcomes.
