Save 30% with code HTAI25 at press.jhu.edu.

Buy the Book at Amazon
Buy the Book at JHUP

Design Principles for Teaching in the age of AI

Teachers should leave technology for what it does best. In the past, a technology might have removed some of the friction of access, first exposure and logistics (Bowen, 2012) to leave class time for the human relationships that drive effort and learning (Bowen 2021). AI inverts this problem. AI no longer frees time for friction, it dissolves the friction and effort of learning entirely. Students using AI often think they are learning, but with unguarded AI, they are mostly not. Teachers need to be more intentional about designing in the essential friction and desirable discomfort of learning. Research demonstrates how AI can both improve and impede learning. Combined with what we understand about the brain and human learning, we can propose a new set of design principles for teaching in the age of AI.

Slides and Citations for Re-Designing Learning in the Age of AI.

Research Findings So Far

What Students Are Doing

Findings 1-5 deal with adoption, perception, and behavior. They are context rather than evidence about learning, but they establish the scale of the change and why policy alone cannot manage it.

1. Student use of AI continues to climb

Multiple surveys (Digital Education Council (2024), Thomas et al. (2025), Student Voice Survey (2026) confirm steady increases in the use of AI.  The Higher Education Policy Institute in the UK reported an increase of 66% → 92% in one year (Freeman (2025). A four-semester study of roughly 300 students found AI use rising for both assessment preparation (57% → 83%) and studying (44% → 76%) (Parker et al., 2026).

2. Students say they use it carefully and for learning

3. Students’ use of AI to do academic work is also rising

Despite prohibitions, 21% of students say they use AI daily for coursework, with more than half saying they use it daily or weekly (Marken, S., 2026). A large 2025 study of actual student usage found a fairly even split between collaborative uses (guidance and refinement) and direct output creation, with slightly less direct output in STEM (Anthropic, 2025). Freeman (2025),  Student Voice Survey (2026), College Board (2025) (and most teachers) would confirm academic use of AI continues to rise. Policies are not enough.

Note that two of the best behavioral datasets here (Anthropic, 2025; Jones, 2026) are vendor telemetry: valuable because they measure actual use rather than self-report, but produced by companies with a commercial stake in student usage.

4. Students say AI improves their academic performance

Students report believing AI improves their academic work (Baltà-Salvador, R., et al. (2025) & Ravšelj, D., et al. (2025), and the more improvement they perceive, the more ethical they consider its use (Parker et al., 2026).  Kaya, M.H., and Adıgüzel, T. (2025) also find believing AI will improve academic performance is a driver of increased use. Marken, S., 2026 reports 35% of college students say this is extremely important in why they use AI, with 35% saying it is very important. 

5. Students who use AI find it more ethical than those who use it less

Students who used AI for studying (M = 3.66, SD = 0.95) rated using AI to complete academic work as more ethical than those who did not (M = 2.76, SD = 1.05), t(317) = −7.04, p < .001. Students who used AI for assignments (M = 3.55, SD = 1.00) likewise rated its ethicality higher than those who did not (M = 2.88, SD = 1.08), t(317) = −6.10, p < .001 (Parker et al., 2026). Uppal, K., & Hajian, S. (2025) also find that student perceptions of the utility of AI also reports higher grades. They also found a correlation between AI dependency and procrastination, but NOT with AI dependency and plagiarism. Students cite ethics as the top reason for not using AI (Marken, S., 2026).

What AI Use Does to Learning (Findings 6–13)

6. The popular use of AI to summarize is a learning trap.

Learning is in the synthesis. When AI performs the discovery and integration, the facts arrive but the schema-building that builds expertise (Oakley et al., 2025) does not. Note that summarizing and explaining are precisely the uses students cite most (finding 2): their favorite use may be a learning trap.

Seven preregistered experiments with roughly 10,000 participants randomly assigned people to learn a topic from the AI syntheses or from standard Google links containing the same core facts. AI learners spent less time engaging, reported shallower knowledge, felt less invested, and produced sparser, less original advice that others were less likely to adopt; the authors (Melumad & Yun, 2025) conclude LLMs are weakest for building procedural knowledge.

A similar randomized study of 91 university students researched a socio-scientific issue with either ChatGPT or a search engine. Again, the AI group experienced significantly lower cognitive load but produced lower-quality reasoning and argumentation (Stadler et al., 2024). And in a randomized experiment with 344 secondary students, those who offloaded note-taking to an LLM found the task easier, invested less effort, and retained less than students who generated their own notes—and more than a quarter leaned heavily on copy-paste when it was available (Kreijkes et al., 2026).

All of these conditions are more passive than the best learning designs would suggest, but AI summaries offer the temptation to reduce the essential friction of integration and schema-making.

PISA 2025 found the same shape at population scale: across 91 systems, non-users outscored users in science on task-specific AI uses, though for summarising texts moderate users (once a month to twice a week) outperformed both rare and near-daily users (OECD, 2026). Summaries also change who does the connecting: AI-generated discussion summaries widened students’ reading while lowering betweenness centrality, so fewer students occupied the broker position, although the authors read this as exposure becoming less dependent on a few individuals rather than as a loss (Hao & Cukurova, 2026).

7. When student use is unguided, homework improves but exam scores fall

A preregistered experiment (N = 704) found that access to an AI assistant did not significantly lower unaided test scores when it gave hints by default and full answers only on request; brief feedback on each learner’s own AI use cut requests for full answers and raised scores (Maier et al., 2026).

Two 2026 experiments pull in opposite directions. In three randomized experiments with 1,060 US adults, 10 to 15 minutes with an AI assistant that answered on request lowered scores once the AI was removed and made people more likely to give up (G. Liu et al., 2026). A randomized experiment with 211 Middlebury undergraduates found the reverse: students who could use ChatGPT while learning an unfamiliar topic for an hour scored 0.28 standard deviations higher on unaided tests, and a smaller gain (0.21 SD, significant only at the 10% level) remained a week later (Contractor & Reyes, 2026). In that study, study time was fixed and cheating on the AI-free tests rose. Both are working papers, not yet peer reviewed. Outside the classroom, a preregistered three-month field experiment with 133 patent lawyers found that AI drafting help improved everyone’s drafts, but only senior lawyers did better on a later task without AI; junior lawyers, who gained most while assisted, showed no average gain (Autor et al., 2026, working paper). In a Carnegie Mellon proof course, students randomly given early chatbot access scored higher on homework but not on the next AI-free exam, and heavier chatbot use went with lower exam scores, while a feedback tool that required a proof attempt first did not (E. Chen et al., 2026, preprint).

8. AI disrupts self-assessment: students overestimate what they learn

In two studies (N = 246; replication N = 452), participants solved LSAT logical-reasoning problems with ChatGPT. Performance improved by about three points against a norm population, but participants overestimated their own performance by about four points (and the usual Dunning–Kruger pattern disappeared). With AI, everyone overestimated. Most strikingly, higher AI literacy predicted worse metacognitive accuracy: the most technically knowledgeable users were the most confident and least precise about their own performance (Fernandes et al., 2026).

A new mechanism study suggests that anxiety, not utility, drives dependence. In a stratified survey of 400 undergraduates using validated scales, AI anxiety strongly predicted AI dependence (β = 0.53), where dependence means shallow reliance and offloading, not frequent use. Anxiety eroded students’ confidence in their own capabilities (β = –0.48), and lower confidence predicted more dependence (Wu et al., 2026). AI literacy did not weaken the anxiety-to-dependence link: the more a student knows about AI, the more acutely they feel the gap between their capabilities and the machine’s, and that felt gap feeds dependence. (Cross-sectional, single-institution data—a mechanism to design around, not yet a causal claim.)

A contrasting null tempers this: in 467 Spanish undergraduates, fear of failure and self-esteem did not predict ChatGPT dependence, and neither peer nor teacher support moderated it; the authors suggest reliance may track the functional pull of quick answers rather than emotional vulnerability (Galindo-Domínguez et al., 2026). (Also cross-sectional and single-institution.)

A general pattern is emerging: AI can manufacture fluency which then masquerades as learning and damages self-perception: the Turkish students in Bastani et al. (2025) did not perceive the damage; the students in Kreijkes et al. (2026) preferred the condition that taught them less because it felt easier; and knowledge workers with higher confidence in AI do less critical thinking (Lee et al., 2025).

The design implication is that self-assessment cannot be left to the student (see principle 7 below). There is also an uncomfortable corollary for AI-literacy programs: AI literacy might disrupt other learning outcomes. If literacy is knowledge about AI rather than practiced metacognition with AI, it may inflate confidence faster than competence.

In five experiments where the AI advice was built to be wrong, people with AI answered more questions, were right about a third as often, and nearly doubled their confidence (Marcoccia et al., 2026). On a rule-discovery task, default chatbot feedback inflated confidence about as much as deliberately flattering feedback (Batista & Griffiths, 2026). After short chats with a flattering chatbot, people rated themselves more knowledgeable while a quiz showed no change in what they knew (Rathje et al., 2026). Across 11 models, AI affirmed users’ actions 49% more often than people did, and users trusted the flattering models more (Cheng et al., 2026). Warning people about flattering AI lowered their trust in it but did not reduce its influence on their views (Ye et al., 2026). All but Cheng et al. are preprints.

9. Pre-AI tutoring systems were effective when well-designed

Robust, replicated studies found that pre-AI systems designed to tutor (and not just to give answers) were often as effective as human tutoring — but context and design mattered (Kulik & Fletcher, 2016; VanLehn, 2011).

10. Purpose-built AI tools that withhold answers have produced learning gains

The largest trial so far points to when the help arrives: in a randomized study of 6,997 middle-school students, AI raised the chance the next attempt after a mistake was correct by 8.5 percentage points, but one week later produced only a 3.1-point gain on the exact material practiced and none on anything else, and only when paired with mastery progression (Oreopoulos et al., 2026).

In a randomized study in an introductory programming course, students rated a Socratic AI tutor that could also see their code as the least supportive of four tutor designs (Eastwood et al., 2026, preprint).

In a preregistered randomized study in an online MBA finance course, a tutor built around the course raised post-test scores by 6.6 of 55 points and final exam scores by 2.6 of 100 compared with students free to use consumer AI; a voice version cost 2.8 times as much and added no measurable learning (Yang et al., 2026, preprint). In 900 real student conversations with 35 teacher-configured bots, the bots skipped steps their instructors had specified 54% of the time (Z. Liu et al., 2026, preprint).

Two randomized trials qualify this finding. In a programming course, a hint-first tutor and unrestricted ChatGPT both raised exercise scores, but neither produced greater knowledge gains than working without AI (Bassner et al., 2026). In a preregistered trial with 1,179 UK secondary students, AI tutoring matched expert human tutors inside the learning platform but did not raise scores on a standardised maths test after 10 weeks (LearnLM Team, Google & Eedi, 2026).

11. Good AI-human workflow might support focus by reducing extraneous cognitive load

12. Access is not enough

13. AI can effectively support the humans around students to deliver learning gains

The same Stanford group behind finding 12 ran the first preregistered RCT of a human-AI system in live tutoring: 900 tutors and 1,800 K-12 students from historically underserved communities. Students whose tutors had real-time AI coaching (Tutor CoPilot) were 4 percentage points more likely to master math topics, and students of the lowest-rated tutors gained 9 points—at a cost of roughly $20 per tutor per year. Analysis of over 550,000 messages showed the AI shifted tutors toward probing questions and away from giving answers (Wang et al., 2024). ProjectCafe is an AI-powered support tool for teachers that reviews class video for key moments that are then presented to the teacher (or a human tutor) for analysis (Urban Assembly, 2026). In a randomized trial in a Michigan economics course (354 students, 11 TAs), TAs who wrote feedback with AI suggestions got better student revisions (d = 0.5), though a later post-test showed no difference and grading took just as long (Lu et al., 2026).

AI aimed at students must fight the offloading problem, but AI aimed at the teacher might offload the extraneous tasks. Which tasks always needs to be carefully designed, but supporting teachers with AI seems a positive for both students and teachers.

KEY FINDING: Who does the cognitive work matters.

The person who does the work gets the benefit. This was true before AI, but since AI makes it easier to offload thinking, AI policies will not be enough (findings 1 and 3, above).

Cognitive offloading is often a choice (Dunn & Risko, 2016). Humans routinely run a cost–benefit analysis (often unconsciously) and then decide whether offloading serves our immediate needs: our working memory is limited, so making a grocery list is a great idea. Learning makes this choice more critical.

Educators conceptualize this in two related ways. Cognitive Load Theory (CLT; Sweller et al., 2019) holds that when we exceed our working memory, we have too much cognitive load and cannot learn. Kalyuga and Plass (2025) reconceptualize that load in two parts: intrinsic load and extraneous load. Teachers have long worked to minimize harmful extraneous load (too many words on slides, or a confusing assignment). Students say they mostly use AI to reduce extraneous load (finding 2) and making instructions clearer should free more cognitive capacity for the learning itself.

But important cognitive work can feel extraneous, and learning also requires engaging with a challenge that is neither too hard nor too easy. Vygotsky (1978) called this the zone of proximal development; Csikszentmihalyi (1990) called it flow; Bjork and Bjork (2011) called it “desirable difficulty.” This is the intrinsic load of the content that constitutes the learning (de Bruin et al., 2023). Without AI guardrails, students will often offload the most desirable difficulties (finding 6). In other words, AI lets students make the wrong offloading choice (finding 4).

Your cognitive-load choices depend on your goals. If you want to understand a concept, you ask AI to explain it (finding 2). If you want an A, you ask AI to do your homework (finding 4). And students often think efficiency is the goal (Zhai et al., 2024). The real problem isn’t cheating, it is judgment. The ease of AI offloading only amplifies the need to manage or nudge students’ choices about cognitive load.

Learning goals can look formulaic, but students often don’t understand the real purpose of cognitive difficulty: there is a reason fitness coaches constantly clarify the value of discomfort. We have always been cognitive coaches (and not just professors), but AI makes it even more important to start with motivation and transparency.

If the default student use of AI is harmful offloading (findings 4 & 6), AI also offers desirable offloading of extraneous load (finding 8). “Load Reduction Instruction” (LRI; Martin et al., 2025) suggests AI could improve feedback (humans tend to give too much at once), structure practice, and build better scaffolding. AI’s ability to customize offers a way to increase engagement and effort (the choice to use more cognitive load). AI can turn your “when does train B catch up with train A” problems into ones that look relevant and worth the effort for each student. Video games offer a customized experience designed to keep every user pleasantly frustrated: too easy or too hard, and you quit. AI tools can keep students at the right (customized) level of interest and difficulty.

Determining which parts of cognitive effort are extraneous is nuanced, and students rarely understand it. Is correcting grammar or formatting citations important for learning, or can I offload it? Hong et al. (2025) asked students to work through reflective cycles in which AI brainstormed an outline, followed by individual critique, peer revision, and journaling about the offloading decision (N = 240 first-year English majors across four colleges; half in a control that ran the same cycle with traditional brainstorming). The students who offloaded the initial brainstorming, but kept the critique, revision, and reflection, showed significantly greater gains in critical thinking (analysis, evaluation, and reflection). In other words, offloading the divergent first draft could free capacity for higher-order work if that is the objective—see also Anders & Dux Speltz (2026). 

More importantly, the part of the cognitive load which is intrinsic is relative to the learner’s prior knowledge. Kalyuga et al (2003) named this the “expertise-reversal effect” which makes it hard (impossible?) for human teachers to design assignments with the right balance for every student. You might need an explanation of term A to understand my assignment, but it is extraneous and wasted cognitive effort for your neighbor. This explains why AI’s ability to reduce mental effort seems contradictory; if AI helps you reduce extraneous load, you can focus on higher order thinking, but if AI reduces intrinsic effort for the desirable difficulty, it also reduces learning. Li et al., 2025 connects AIs ability to lower extraneous cognitive load was key to improved learning critical. We will need more specific studies to learn where this line is and how to adjust it, but since AI can adjust easily, designing tools that monitor and lower extraneous load could become a key design insight. 

Research also suggests AI can boost metacognition by forcing students to pause and reflect (Singh et al., 2025; Xu et al., 2025). In another high-quality study, active, mediated use of ChatGPT promoted complex critical thinking (Suriano et al., 2025). Frameworks such as the “cognitive mirror” propose using AI to reflect students’ thinking back to them to prompt self-regulation (Tomisu et al., 2025). These are early results, but they suggest AI might also be used to make desirable difficulty unavoidable.

To use or not use AI is less important than how we design friction or desirable difficulty (with or without AI) to be unavoidable. The new problem for teachers is that AI makes harmful offloading easier, and our assignments need to foresee this. That might mean allowing AI use (increasingly hard to forbid) while being more intentional about guiding students toward the right offloading choice. Lodge and Loble (2026) call this the “performance paradox”: as AI makes output better, it also makes it too easy for students to undermine learning. At the same time, AI offers a new way to reduce extraneous load and push students into metacognition.

How students perceive their interaction with AI also matters; the same interaction can create empowerment or dependency (Nasr et al., 2025; Edwards & Edwards, 2025). Users with higher self-confidence do more critical thinking while using AI, whereas higher confidence in the AI is associated with less critical thinking and less effort (Lee et al., 2025). This also seems to be task-specific.

While AI literate students, on average, may be less anxious and less dependent on AI, among anxious students, higher AI literacy intensifies the felt gap between your own abilities and the machine’s (Wu et al., 2026). This means we also have to design for increasing self-confidence, and not just more AI fluency.

First, we will need to pair exposure to AI capability with explicit reassurance about students’ own developing competence. Self-efficacy grows most from mastery experiences (Bandura, 1997), so we will need to scaffold assessments so that students gradually see they can do the work without the tool. Second, human teachers will need to talk openly with students about how AI makes them feel: a confident “AI power user” may be more fragile under stress than they look.

The speed and fluency of AI could make it harder for students to become critical readers and lead to more fast “System 1” thinking (Kahneman, 2011). This could make AI a massive “Dunning–Kruger” accelerator (Kruger, J., & Dunning, D., 1999). One solution is to ask for fewer artifacts (where the temptation to offload is greatest) and more thinking and judgment: explain your reasoning, reframe this question, verify the output, contest these findings, surface hidden assumptions, defend this conclusion.  We need to redesign against the false fluency AI creates.

If you still want students to produce artifacts, you also need to raise standards and require better work as well as deeper thinking. What can students do with AI that they could not do on their own? Catching cheaters (offloaders who misunderstand the real goals of the homework) is increasingly futile. Now is a good moment to design assignments that increase cognitive load in the most important places. Well-designed AI tools can increase retrieval practice and critical thinking — which means teachers designing prompts and secure custom bots rather than simply allowing open-tool use. You can find templates, examples, and instructions for custom bots at https://weteachwithai.com/creating-ai-simulations-and-custom-bots/ or in Teaching with AI (2nd edn).

If you are going to use AI in an assignment, both the design of the tool and student agency are critical. But the most important design choice for teachers has always been how to direct students’ choices about offloading. AI just increases the urgency. You do not need to use AI in your teaching — but you cannot ignore that its existence makes the wrong offloading ramp easier to take.

Design Principles for Teaching with AI

1  REFINE & CLARIFY learning goals.

Students try harder and learn more if they understand the purpose of your tasks. The new AI environment only increases the need to clarify why your assignment matters and what is the most important desirable difficulty. Cheating is less likely when instructors explain why the content matters for future professional practice and when tasks are authentic (Rüth et al., 2024).

2. DECIDE if AI supports learning.

Will AI support or impede your desired difficulty? Can you use AI to make friction unavoidable, or the wrong kind of offloading less useful? Could AI reduce extraneous cognitive load? Align your assignment design to the objective and use AI only if you can design it to support that.

IF NO, remember that forbidding AI without increased motivation and transparency can leave unguarded AI products as a tempting efficiency. An AI policy or syllabus statement is not enough. Good assignments (with or without AI) inspire students because they care (the purpose is clear and relevant) they can (the task and its process are understood) and they matter (the connection between the work and their goals is visible).  

3  PRESERVE student thinking.

Build specialized AI tools (custom bots) that coach rather than answer. AI improves durable learning only when it is designed to question, withhold solutions, scaffold, provide one hint at a time, and fade its support. 

4  CUSTOMIZE the friction.

Use AI to calibrate challenge to the edge of each student’s ability and tie the task to what they care about: you can now personalize difficulty and relevance at once. AI can make it easier to keep every student productively frustrated: stretched, while neither bored nor overwhelmed

5  AMPLIFY to raise the bar.

Use AI for what humans cannot do at scale, like fast, private, formative feedback, or support for complex work students could not do alone. As AI tempts students into less thinking and judgment, you will want to require more of both. AI may change where you want desirable difficulty; design work with AI to make it unavoidable. LLM polishing reduces writing-complexity variance by 21–50% across more than 880,000 texts while leaving core content intact (Sourati et al., 2026).

6  MEASURE process and higher-order thinking.

Grade reasoning, not artifacts, and build AI-off reflection checkpoints where you measure learning and not dependence. Ask students to probe, dissect, critique, and verify AI output. Design prompts that surface student thinking so it is both built and visible. At the same time, you will need to consider how AI will change work in your discipline: calculators didn’t eliminate the need to learn math, but they did change the skills needed to be an accountant. 

7  EMPOWER access.

Anticipate access gaps and design for equity: structured access narrows gaps while unstructured access widens them, so ensure every student can participate with the same level of model. Note finding 12 that students will need instruction and support to engage with AI tutors.

EQUIP the humans.

This might be more of an institutional design necessity, but supporting the humans around students with AI rather than students themselves (finding 13) shows promise for learning. It is also essential for equity. Teachers need to understand AI before they can teach it. This might improve teaching directly (AI improvement of assessments or assignments) or indirectly (AI doing extraneous tasks that allow more human teacher focus on students).

Susan Ray has built a “friction bot” that helps faculty build friction and process into assignments.

ChatGPT version https://chatgpt.com/g/g-6a4d6047afb081919eeed3e503e6bfaf-assignment-rebuilder-add-friction-to-assessments

BoodleBox version: https://box.boodle.ai/a/@AssignmentRebuilderAddFrictiontoAssessments

All of our custom tools and simulations were stress-tested on the best models but set to run and interact with students using only free models. You can find examples and templates in Teaching with AI (2nd edn) or at weteachwithai.com, including custom bots at weteachwithai.com/creating-ai-simulations-and-custom-bots.

References