The Thinking AI Should Not Do for Us

AI reliably improves what students produce, but it does not reliably improve what students learn. That single distinction organizes everything in this post. It explains why some AI-rich classrooms are thriving and others are quietly hollowing out, and it tells academic leaders where to spend their limited money, time, and political capital.

The 2026 EDUCAUSE Horizon Report, Teaching and Learning Edition (Robert, Muscanell, McCormack, & Arnold, 2026) describes a sector under compounding pressure. Enrollments are falling, funding is unstable, and the public is asking whether a degree is worth what it costs (HR p. 8). Into that environment AI arrives promising efficiency, personalization, and scale. Three of the report's fifteen trends name AI directly in their titles, and AI runs through nearly every other one.

The timing creates a trap. Institutions most need credible evidence of learning at the very moment a technology makes student output look better than ever. A polished essay, a clean problem set, or a confident discussion post no longer certifies that a mind has changed.

Fan and colleagues (2025) captured this trap in a single randomized study. University students who revised essays with ChatGPT earned the largest score improvements of any group. Yet their knowledge gain and transfer were no different from students who had no extra support, and their self-regulated learning looked different in ways that suggested less monitoring and evaluation. The authors named the pattern metacognitive laziness: the tool absorbs the thinking the learner would otherwise have done.

Approach

This post begins with the Horizon Report and expands outward into the research literature. Part One defines engagement, reviews the strongest current evidence on AI and learning, examines the ICAP framework, extracts design principles, introduces the idea of technology as a substrate for human connection, and applies all of this to relational engagement and to clinical education in physical therapy. Part Two turns to leadership: the current state of institutional readiness, assessment and integrity, faculty development, governance, and the often neglected question of faculty wellbeing.

Horizon Report page references appear as (HR p. X). Studies are cited by author and year, with links in the references to sources that were opened and checked. Where a study is a preprint, has a small sample, or has drawn methodological criticism, that is noted, because a field moving this fast is especially vulnerable to overclaiming.

One caution frames the whole piece. The evidence on AI in education is young, mostly short-term, and dominated by single-site studies of tools that are already outdated. The conclusions here are best read as working hypotheses, grounded in learning science that long predates AI, rather than settled findings.

Engagement is not one thing, and most claims that AI "increases engagement" quietly measure the easiest kind. Before judging any AI tool, it helps to be precise about which engagement it is supposed to improve.

Four dimensions

Fredricks, Blumenfeld, and Paris (2004), in a review that still anchors the field, described engagement as a multidimensional construct with three parts:

• Behavioral engagement: participation, effort, attendance, time on task, completion.
• Emotional engagement: interest, enjoyment, belonging, and the absence of boredom or anxiety.
• Cognitive engagement: investment in understanding, willingness to exert effort on difficult ideas, and the use of deep rather than surface strategies.

Reeve (2013) added a fourth: agentic engagement, the student's proactive contribution to the flow of instruction. Agentically engaged students ask questions, express preferences, and shape the learning environment to suit their needs. In Reeve's longitudinal data, agentic engagement predicted achievement beyond the other three dimensions and also drew more autonomy support from teachers over the semester.

This fourth dimension matters for AI. A student who asks an AI to restructure a confusing reading, generate practice questions at the right difficulty, or challenge their argument is showing agentic engagement. A student who asks the same tool to produce the assignment is not. The technology is identical; the stance is opposite.

Why the distinction matters for AI

AI tools are exceptionally good at raising behavioral and emotional engagement. They are always available, endlessly patient, responsive in seconds, and pleasant to use. Those features produce more interaction and higher satisfaction ratings almost automatically.

Cognitive engagement is different. It requires effort, and effort is precisely what a helpful assistant is designed to reduce. A tool optimized to satisfy the user will, by default, remove the friction that cognitive engagement depends on. This is the core design problem of AI in education.

A useful illustration comes from research that predates generative AI. Deslauriers and colleagues (2019) found that students in actively taught physics classes learned more than students in polished lectures, yet felt they had learned less. The cognitive effort of active learning registered as confusion and difficulty, and students mistook fluency for learning. AI tools that make learning feel smooth risk exploiting exactly this illusion.

Motivation underneath engagement

Self-determination theory (Ryan & Deci, 2000) explains why some engagement lasts. It holds that people are most durably motivated when three needs are met: autonomy (a sense of choice and ownership), competence (a sense of growing mastery), and relatedness (a sense of connection to others).

AI can support or undermine each need:

• Autonomy grows when students choose how to use AI within clear boundaries. It shrinks under surveillance, detection, and blanket bans.
• Competence grows when AI scaffolds a student just beyond their current ability. It shrinks when AI performs the task, leaving the student with a product but no new capability.
• Relatedness grows when AI frees instructors to spend more time with students. It shrinks when AI becomes the student's main academic relationship.

That last point connects directly to the Horizon Report, which warns that students increasingly feel more comfortable seeking help from AI than from instructors, and that office-hour visits and informal interactions could decline (HR p. 9).

The research on AI and learning points in two directions at once, and the difference between the two directions is design. That conclusion is more useful, and more honest, than either "AI improves learning" or "AI harms learning."

The Horizon Report itself is thin here, which is worth noticing. Its examples mostly report adoption, usage, or satisfaction. Lone Star College's advising chatbot, for instance, is described by its conversation volume and accuracy rate (HR p. 13), not by any effect on persistence or learning. Most of the report's emerging practices are pilots with outcomes still to come (HR pp. 37 to 41). To understand what AI actually does to learning, we have to go to the research literature.

The long shadow of the two-sigma problem

The hope behind AI tutoring is old. Bloom (1984) reported that students tutored one-to-one performed about two standard deviations better than students taught conventionally, and challenged the field to find group methods that could approach that effect. Later work tempered the number. VanLehn's (2011) review found human tutoring effects closer to 0.8 standard deviations, with well-designed intelligent tutoring systems nearly matching them.

The lesson for AI is that tutoring works because of what tutors do: they diagnose misconceptions, prompt students to explain, withhold answers strategically, and give timely feedback. A system that simply answers questions is not a tutor in this sense. It is a very fast reference book.

When AI harms learning: Bastani and colleagues (2025)

Bastani and colleagues ran a field experiment with nearly 1,000 high school math students in Turkey, published in the Proceedings of the National Academy of Sciences. They compared a control group with two GPT-4 tools: "GPT Base," which mimicked a standard ChatGPT interface, and "GPT Tutor," which was prompted to give hints rather than answers and drew on teacher-provided solutions.

During practice, both tools raised performance substantially: 48% for GPT Base and 127% for GPT Tutor. But when access was removed for the exam, GPT Base students scored 17% worse than students who never had AI. GPT Tutor students largely avoided that harm. Chat logs suggested GPT Base students used the tool as a crutch, asking for answers rather than help.

The finding that matters most for leaders is subtle. Grades went up while students had AI. An institution watching only course grades, completion rates, or satisfaction would have seen success. The damage appeared only when the tool was taken away.

When AI helps learning: Kestin and colleagues (2025)

Kestin and colleagues ran a randomized crossover trial with 194 students in Harvard's largest introductory physics course, published in Scientific Reports. Each student experienced both conditions: one lesson in an already strong active-learning class, and one lesson at home with a custom AI tutor.

The AI tutor was not a general chatbot. It was built on the same pedagogical principles as the classroom: active learning, cognitive load management, growth mindset, sequential scaffolding, and accurate feedback. Crucially, it was fed the instructors' own step-by-step solutions to prevent errors. Median learning gains in the AI condition were more than double those of the active-learning class, with effect sizes estimated between 0.73 and 1.3 standard deviations. Median time on task was 49 minutes, compared with 60 minutes in class. Students also reported feeling more engaged and more motivated.

The authors are careful about scope. The study covered two topics, measured short-term outcomes, and involved content well suited to structured tutoring. They explicitly do not claim AI tutoring will outperform active learning for complex synthesis or higher-order thinking. And they recommend using AI to prepare students before class so that in-person time can be spent on problem solving, projects, and group work, which is a flipped model, not a replacement model.

When AI improves products but not learning: Fan and colleagues (2025)

Fan and colleagues' randomized lab study, published in the British Journal of Educational Technology, sits between these poles. Among 117 university students revising essays with support from ChatGPT, a human expert, writing analytics tools, or nothing extra, the ChatGPT group showed the largest essay score improvements. Yet knowledge gain and transfer did not differ across groups, and there were no differences in post-task intrinsic motivation. What did differ was the pattern of self-regulated learning processes, which the authors interpreted as a risk of dependence and metacognitive laziness.

Neural and self-report signals

Two further studies add texture, with important caveats.

Kosmyna and colleagues (2025) at the MIT Media Lab used EEG to study 54 participants writing essays over several sessions with an LLM, a search engine, or no tools. The preprint reports that brain connectivity scaled down with external tool use, with LLM users showing the weakest connectivity. LLM users also reported the lowest ownership of their essays and struggled to accurately quote their own work. In a fourth session, only 18 participants switched conditions. The sample is small, the study is not peer reviewed, and its interpretation has been debated. It is best treated as a provocative signal, not a verdict.

Lee and colleagues (2025), in a CHI paper from Microsoft Research and Carnegie Mellon, surveyed 319 knowledge workers about 936 real uses of generative AI. Higher confidence in AI was associated with less critical thinking, while higher self-confidence in one's own ability was associated with more. Critical thinking did not disappear so much as shift toward verifying AI output, integrating it, and stewarding the task. The implication for education is uncomfortable: novices, who have the least self-confidence and expertise, are exactly the people most likely to accept AI output uncritically.

A cautionary tale about the evidence itself

In May 2025, a meta-analysis of 51 studies by Wang and Fan reported large positive effects of ChatGPT on learning performance. It was widely shared, cited about 270 times, and recommended to educators as essential reading. On April 22, 2026, it was retracted over concerns about discrepancies in the analysis. Critics had noted early that most included studies had very small samples and that study quality had not been assessed.

The retraction is a useful warning for anyone building institutional strategy on this literature. Enthusiasm traveled faster than verification. Leaders should prefer randomized designs, delayed or unassisted outcome measures, and replication, and should be wary of any single headline effect size.

Reading the evidence together

Read together, these studies support one clear claim. The underlying model was not the variable that mattered most; the pedagogy built around it was. The same class of technology produced harm in one design and large gains in another. The question "Does AI improve learning?" is therefore badly formed. The better question is: under what design conditions does AI improve durable learning, for whom, and measured how?

The evidence also carries limits that should temper every conclusion. Nearly all studies are short-term. Contexts are narrow. "Learning" is measured differently in each. And the tools studied in 2023 and 2024 are already outdated, which cuts both ways: newer models may be better tutors, but also more persuasive crutches.

The ICAP framework is the most practical lens available for judging what any learning activity, including an AI-supported one, actually asks of learners. It was developed by Michelene Chi, grew out of her earlier work on "differentiated overt learning activities" (Chi, 2009), and was most fully stated by Chi and Wylie (2014). Its core move is to define engagement by what learners observably do, then link those behaviors to underlying cognitive processes and expected learning outcomes. That lets an instructor judge an activity's likely depth from the outside, without measuring thinking directly.

The four modes

ICAP sorts learning behaviors into four modes, listed here in descending order of predicted learning.

The ICAP hypothesis predicts the ordering Interactive > Constructive > Active > Passive. The hypothesis is ordinal: it predicts which mode should produce more learning, not how much more.

Distinctions that matter in practice

• Active is not the same as active learning. Many activities labeled active learning are merely active in ICAP terms. A clicker question that asks students to recall a fact, or a worksheet filled in from the slides, involves doing something but generating nothing new.
• Generation separates active from constructive. Copying a definition is active; restating it in one's own words, giving a new example, or explaining why it matters is constructive. The same task, such as note-taking, can fall in either mode depending on what the learner actually produces.
• Interactive requires both partners to be constructive. Two students talking is not automatically interactive. If one explains and the other listens, the listener is passive; if they take turns reading answers aloud, they are active. Genuine interaction needs sustained turn-taking in which each contribution builds on the last. A partner can be a peer, a tutor, or a well-designed system, provided the exchange is truly co-constructive.
• Modes are nested. Constructive behavior includes active behavior, and interactive includes constructive.

The evidence base

Chi and Wylie (2014) reviewed a large body of laboratory and classroom research and argued that the predicted ordering held across many studies, including classic findings that self-explanation outperforms rereading. Menekse, Stump, Krause, and Chi (2013) found the predicted ordering on deeper test questions in engineering classrooms. Wiggins and colleagues (2017) found that interactive activities produced larger gains than constructive or active ones in an intensely active undergraduate biology course. The framework is also consistent with the broader finding that active instruction outperforms lecture in STEM (Freeman et al., 2014).

Translating ICAP into practice is harder than it looks. When Chi and colleagues (2018) trained teachers in ICAP and asked them to design activities, the teachers improved, but many activities they intended as constructive or interactive turned out to be merely active on close inspection. Designing genuinely generative tasks is a skill, not a reflex.

The framework also faces recent challenges. At least one lab study with 161 undergraduates learning from a difficult text found the opposite ordering, with passive learners performing best, most likely because constructive and interactive tasks overloaded learners with limited prior knowledge. That study was identified only through a search summary and was not reviewed in full here, so it should be treated as a signal rather than a verified finding.

Strengths

• Practical and observable. Instructors can classify an activity by what students produce, without specialized measurement.
• Diagnostic. It explains why some active learning disappoints: the activity was active but not constructive.
• Design-oriented. It shows how to upgrade a task, first by adding a generative demand, then by adding genuine dialogue.
• Explanatory. Apparently conflicting findings on group work, note-taking, or discussion often reflect differences in which mode students actually entered.
• Discipline-agnostic. It applies across the sciences, the humanities, and professional education.

Limitations and critiques

• Overt behavior is an imperfect proxy for thinking. A student listening to a lecture may be silently self-explaining, and a student drawing a concept map may be reproducing one from memory. ICAP classifies an activity's design intent more reliably than any individual's actual cognition.
• Prior knowledge and cognitive load are boundary conditions. Constructive and interactive tasks assume learners know enough to generate something meaningful. For true novices facing complex material, generative tasks can overload working memory, and well-designed direct instruction, including worked examples, may outperform them. This is consistent with cognitive load theory and the expertise reversal effect (Kalyuga, 2007).
• The interactive advantage is the least consistent. The gap between constructive and interactive is smaller and less reliable than the gap between active and constructive. Interaction carries costs: coordination effort, unequal participation, and the risk that one partner's errors spread.
• Quality within a mode varies. Two constructive activities can differ greatly in value depending on the prompt, the feedback, and alignment with goals. ICAP describes the type of engagement, not its quality.
• Emotional and motivational dimensions sit outside the framework. ICAP is a theory of cognitive engagement. Belonging, interest, and agency are addressed by the other frameworks in What We Mean by Engagement.
• Ordinal, not quantitative. It tells you which mode should win, not by how much or whether the extra time is worth it in a given context.

ICAP, backward design, and the question of what counts as learning

ICAP and Understanding by Design (discussed in Design Principles) approach the same problem from opposite ends, and putting them together exposes a weakness running through much of the current research on AI and learning.

ICAP is a theory of activity; backward design is a theory of evidence. ICAP predicts that the mode of an activity determines the depth of the resulting understanding: passive engagement supports recall, active supports application in familiar contexts, constructive supports transfer, and interactive supports the co-creation of new ideas. Backward design insists that before any activity is chosen, the designer must decide what evidence will count as understanding (Wiggins & McTighe, 2005). Read together, the implication is direct: an ICAP mode is only the right choice relative to the evidence it is meant to produce. A course whose assessments ask only for recall gives no reason to prefer constructive over passive activities, and an instructor cannot judge whether an activity is "deep enough" until the target performance has been defined.

The ICAP evidence itself shows why the order matters. The predicted advantage of constructive and interactive modes appears most clearly on measures of deep understanding and transfer. Menekse and colleagues (2013) found the ICAP ordering on deeper test questions; the recent counter-finding in which passive learners did best involved a difficult text and comprehension-style tests with learners who lacked prior knowledge. Whether ICAP "works" depends heavily on what the outcome measure asks students to do. In other words, the assessment determines whether the activity design matters at all. That is the backward-design hypothesis confirmed from inside the ICAP literature.

Much of the AI research reverses this order. Many studies begin with a tool, deploy it, and then measure whatever outcome is convenient. When the evidence of learning is not defined first, and defined at the level of understanding the course actually values, studies can report "learning gains" that are not learning in the sense backward design or ICAP would recognize. Several recurring patterns are visible in the evidence reviewed in this post:

The underlying problem has a name in learning science. Soderstrom and Bjork (2015) distinguished performance, what can be observed during or immediately after instruction, from learning, the relatively durable change in knowledge or skill that supports later retention and transfer. Performance is an unreliable index of learning; conditions that boost performance in the moment can impair long-term learning, and vice versa. AI is a uniquely powerful performance enhancer. Research that measures performance and calls it learning will systematically overstate what AI does for students.

The better studies show the alternative. The most informative work reviewed here comes closest to backward design. Bastani and colleagues defined learning as unassisted exam performance after the tool was withdrawn, which is exactly how they detected harm hidden behind rising practice scores. Kestin and colleagues reported the cognitive level of their test items, most of which targeted applying and analyzing rather than remembering, which makes the meaning of their gains interpretable, even though their measures were still short-term. Fan and colleagues separated product quality from knowledge gain and transfer, which is what allowed them to see that the two diverged.

A standard worth adopting. For both course design and research, the combined lesson of ICAP and backward design suggests a simple discipline:

1. Define the evidence of understanding first, at the depth the course or study actually values, including transfer to new situations.
2. Measure that evidence without the AI tool present, unless using the tool well is itself the learning goal.
3. Include a delayed measure where possible, to distinguish learning from momentary performance.
4. Only then choose the ICAP mode, and the AI role, most likely to produce that evidence.
5. Treat satisfaction, usage, and product quality as secondary indicators, never as proof of learning.

For health professions programs, this is not abstract. The evidence that ultimately matters is performance with a real patient, often without any tool and always under time pressure. An AI-supported activity should be judged by whether it improves that performance, measured that way, and not by how polished the AI-assisted work looked along the way.

ICAP and AI

ICAP is especially useful for AI because the same underlying technology can place a learner in any of the four modes, depending on how the tool is designed and how the task is framed.

This mapping echoes the evidence in What the Evidence Actually Says. Bastani's unguided tutor let students remain passive or active by requesting answers, while the guardrailed version pushed them toward constructive work; Kestin's tutor was explicitly designed to demand generation. One open question is whether dialogue with an AI counts as truly interactive in Chi's sense. An AI can build on a student's input, but whether that co-construction produces the same benefits as human dialogue, including social accountability, perspective-taking, and shared meaning, is not yet established.

ICAP in health professions education

For DPT and other clinical programs, ICAP offers a quick audit of any teaching sequence:

• Passive: a lecture on differential diagnosis of low back pain.
• Active: students annotate the slides or complete a red-flag checklist from their notes.
• Constructive: students work through a novel case, generate a ranked differential, and justify each hypothesis.
• Interactive: pairs reason through a case together, challenging and revising each other's hypotheses; or one student examines a standardized patient while a peer observes and then debriefs with them.

Two cautions apply. Early in the curriculum, novices may need worked examples and modeled reasoning before constructive tasks become productive. And psychomotor skills do not map cleanly onto ICAP's largely verbal and conceptual orientation: practicing a mobilization technique with expert feedback is valuable even when it is "active" by ICAP's definitions.

Summary

ICAP's most durable contribution is its distinction between activity that keeps learners busy and activity that makes them generate. Its main limits are that observable behavior imperfectly reflects thinking, that generative tasks assume adequate prior knowledge, and that the interactive advantage is its least consistent claim. Used as a design and diagnostic heuristic rather than a strict law, it offers a sharp question for every AI use discussed in this post: which mode does this put the learner in?

The central design principle is simple to state: AI promotes real engagement when it creates situations that demand human thinking, and erodes engagement when it supplies the thinking. Everything else in this section elaborates that idea.

Desirable difficulties

The learning science behind this principle long predates AI. Robert and Elizabeth Bjork's work on desirable difficulties shows that conditions which slow apparent progress, such as spacing practice over time, interleaving topics, testing rather than restudying, and generating answers before seeing them, produce more durable learning (Bjork & Bjork, 2011). Roediger and Karpicke (2006) showed the same for retrieval: students who practiced recalling material retained far more a week later than students who reread it, even though rereading felt more productive.

Consumer AI is engineered to remove friction. That is its commercial value. Educational AI often needs to do the opposite: introduce the right friction at the right moment, then step back. A tutor that answers instantly removes the generation effect. A tutor that asks "What do you think happens next, and why?" restores it.

Two foundations: Understanding by Design and Scientific Teaching

Two books written long before generative AI anticipate most of what the AI research is now rediscovering. Both are worth returning to, because they give faculty a design process that keeps AI in its proper place.

Understanding by Design: start with the understanding, not the tool

In Understanding by Design, Grant Wiggins and Jay McTighe (2005) proposed backward design, a three-stage planning sequence:

1. Identify desired results. What should students understand and be able to do, and which big ideas and essential questions frame that understanding?
2. Determine acceptable evidence. What performances would convince us that students genuinely understand, as opposed to having memorized or reproduced?
3. Plan learning experiences and instruction. Only then decide which activities, resources, and tools will help students produce that evidence.

Backward design puts AI exactly where it belongs: in Stage 3, as one possible means, chosen after the goals and the evidence are clear. Much of the current confusion about AI comes from reversing the order, starting with a tool and asking what to do with it. Backward design asks instead what understanding a course is for, and only then whether a given AI use helps students build it or lets them bypass it.

Stage 2 matters even more in an AI era. Wiggins and McTighe argued that understanding is revealed through transfer: the ability to use knowledge thoughtfully in new, authentic situations, not to reproduce it in familiar ones. That is precisely the distinction between Bastani's practice scores and exam scores, and between Fan's essay quality and knowledge transfer. A course designed backward from transfer tasks is far harder to hollow out with AI, because a polished product is not the evidence it asks for.

Wiggins and McTighe also described six facets of understanding: a person who truly understands can explain, interpret, apply, take perspective, empathize, and show self-knowledge, including awareness of the limits of their own understanding. AI can readily generate products that look like the first two. It is much less able to stand in for the last three, which depend on a person's own vantage point, relational attunement, and metacognition. For clinical education, the empathy and self-knowledge facets connect directly to the compassion and discernment arguments elsewhere in this post: they are forms of understanding that must be demonstrated by the learner, in person, because they cannot be outsourced.

Scientific Teaching: treat teaching with the rigor of research

In Scientific Teaching, Jo Handelsman, Sarah Miller, and Christine Pfund (2007), building on an earlier Science article by Handelsman and colleagues (2004), argued that teaching should be approached with the same rigor, experimentation, and evidence that faculty bring to their research. Their framework rests on three pillars:

• Active learning, grounded in evidence about how people learn, rather than transmission of information.
• Assessment, especially frequent formative assessment that tells both students and instructors whether learning is actually happening.
• Diversity, meaning teaching designed to include and support all learners, recognizing that the same instruction affects different students differently.

Each pillar speaks to AI. Active learning is the design principle of this section in another form: AI should create the activity, not replace it. Formative assessment is the antidote to the trap Bastani exposed, where grades rose while learning fell; an instructor who regularly checks unassisted understanding would notice. And the diversity pillar anticipates the equity concerns raised throughout this post, from biased AI detectors to the risk that some students receive human mentoring while others receive only automated help.

Scientific teaching also models the stance this post recommends toward AI evidence. It treats each course change as a hypothesis to be tested, not a conviction to be defended. In a field where a widely cited meta-analysis on AI and learning was retracted, and where most studies are short-term and small, that disposition is not academic fussiness. It is the only responsible way to adopt a technology whose effects on learning are still being discovered.

Using them together

The two books complement each other. Understanding by Design tells faculty how to plan: begin with enduring understanding, define the evidence of transfer, then choose activities and tools. Scientific Teaching tells faculty how to improve: build in active learning, gather formative evidence, attend to every learner, and revise based on what the evidence shows. Applied to AI, the combination produces a simple discipline: decide what students must understand, decide how they will show it without the tool, choose AI uses that serve that goal, and check whether they actually did.

Five design moves

The research and the Horizon Report's examples suggest five practical moves.

1. Make AI ask before it tells. Bastani's GPT Tutor and Kestin's physics tutor both withheld answers and prompted reasoning. Any course-level AI tool can be configured, or simply instructed, to respond with a question, a hint, or a request for the student's attempt before offering explanation. This moves the student from ICAP's passive mode toward constructive.

2. Put AI in roles that require a human response. Mollick and Mollick (2023) described several pedagogical roles for AI: tutor, coach, mentor, teammate, student, simulator, and tool. The most cognitively demanding roles are the ones where AI cannot finish the task without the student. When the AI plays a confused student, the learner must teach. When the AI plays a skeptical colleague, the learner must defend. When the AI plays a patient, the learner must assess.

3. Ground AI in expert material. Kestin's tutor worked partly because it was fed instructors' solutions. Bastani's guardrailed tutor drew on teacher-written solutions too. Grounding reduces errors and keeps the tool aligned with what the course actually values. It also keeps faculty expertise at the center rather than outsourcing it to a general model.

4. Make the process visible. Purdue's Charlie gives rubric-aligned feedback on drafts but asks students to evaluate and contextualize the suggestions rather than accept them (HR p. 12). Process portfolios, AI interaction logs submitted with work, and brief reflections on what the student accepted, rejected, and why all turn AI use into an object of metacognition. This directly counters the metacognitive laziness Fan described.

5. Close the loop with humans. Kestin and colleagues explicitly recommend using AI before class so that in-person time can focus on higher-order work. The Bodyswaps crisis simulation for chiropractic students works partly because a structured human-led debrief follows it (HR p. 13). AI rehearsal plus human consolidation appears to be a stronger pattern than AI alone.

The Horizon Report's examples, sorted by what they demand

The report's examples become more useful when sorted by the kind of engagement they target and what they ask of the learner.

Two patterns stand out. First, the examples with the strongest claim to cognitive engagement all contain a built-in demand: a patient who must be calmed, a suggestion that must be judged, a question that must be generated. Second, the Kenyan WhatsApp pilot is a reminder that design principles mean nothing without access. In mobile-first, low-bandwidth settings, engagement starts with whether the tool reaches the learner at all.

Curiosity as the scarce resource

Pikes Peak State College's "Outwondering the Algorithm" framework deserves special attention because it reframes the goal (HR p. 40). It uses a breathing metaphor: divergent thinking (wondering, exploring, gathering) is the inhale, and convergent thinking (synthesizing, deciding, making) is the exhale. Years of answer-focused schooling, the framework argues, have trained students to hold their breath and skip straight to conclusions. AI makes that habit both easier and more visible.

If AI can supply competent answers to most well-formed questions, the scarce human capacity becomes forming good questions in the first place. An engaging AI environment, on this view, is one that rewards question generation, comfort with ambiguity, and the ability to notice what is missing. Assignments can make this concrete: students submit the ten questions they asked an AI about a case, ranked by usefulness, with a note on which ones changed their thinking.

A caution on personalization

The Horizon Report notes interest in AI-personalized textbooks, such as Google's Learn Your Way experiment, while observing that adoption at scale is unlikely in the near term because of shared-objective, copyright, accessibility, and tenure-credit barriers (HR p. 25). Personalization is attractive, but it carries a hidden cost. If every student receives a different text, the shared reference point that makes class discussion possible weakens. The most engaging environments may personalize the path while preserving a common destination and a shared community conversation about it.

A practitioner parallel: IDEO

These principles are not confined to academic research. IDEO, the design firm most closely associated with human-centered design and design thinking, has arrived at a similar position through practice. Through its online learning arm, IDEO U, the firm teaches where AI strengthens creative work and where it does not, and frames the goal as developing judgment rather than tool skills: knowing when AI helps and when it doesn't (IDEO U, n.d.-a). Its AI course opens with responsible use and risk, asks learners to use AI to organize observations but then apply their own judgment to decide what matters, and closes by having each learner write a personal statement of how they intend to use AI responsibly.

IDEO U's general learning design follows a cycle of see, try, share, and reflect, which it grounds in learning science: skills are built through repeated practice rather than observation, knowledge is constructed through interaction with others, and reflection on one's own thinking supports transfer to new contexts (IDEO U, n.d.-b). In the vocabulary of this post, that cycle moves learners from ICAP's passive mode into constructive and interactive work and adds the metacognition that Fan and colleagues found AI can erode. In 2026, IDEO U was recognized for integrating generative AI into its curriculum by augmenting its human-centered frameworks rather than letting the technology overshadow them (Fiske, 2026).

Two cautions apply. IDEO U serves adult professionals rather than degree-seeking students, and its public materials describe design philosophy and reach rather than measured learning outcomes; the recognition cited comes from its course platform vendor. IDEO's contribution is best read as a practitioner's design stance that converges with the research, not as evidence of effect. That convergence is itself useful: a leading design firm and the experimental literature point to the same conclusion, that AI belongs inside a human-centered process in which the learner keeps the thinking and the judgment.

A simple design test

For any proposed AI activity, three questions reveal most of what matters:

1. Who does the thinking? If the AI could complete the activity without the student, redesign it.
2. What does the student produce that the AI did not? If the answer is nothing, the activity is at best active, not constructive.
3. Where do humans re-enter? If no instructor or peer ever sees the student's reasoning, the activity is missing its consolidation step.

For online and hybrid learning, the Community of Inquiry framework (Garrison, Anderson, & Archer, 2000) adds a further layer to the picture of engagement. It describes a meaningful educational experience as the overlap of three presences: cognitive presence (constructing meaning through sustained reflection and discourse), social presence (participants projecting themselves as real people), and teaching presence (the design, facilitation, and direction of learning).

AI can contribute meaningfully to teaching presence through feedback, scaffolding, and timely prompts. It cannot supply social presence among humans, and it cannot replace the instructor's judgment about what this group of learners needs. The danger is not that AI adds to teaching presence. The danger is that an institution mistakes AI-supplied teaching presence for the whole community.

That danger is easier to see, and to govern, through the lens of technology as a substrate (L'HommeDieu, 2026a), a framework developed for healthcare and extended here to the learning environment.

Technology is a substrate, not an actor

The framework begins by rejecting the question "Are you for or against this technology?" as the wrong shape. Technology is not a moral actor in the life of an organization. It is a substrate: a medium on which human decisions grow. Whether a given tool ends up enabling compassion and connection or extinguishing them is not a property of the silicon. It is a property of the human decisions made in designing and deploying it: what the tool is built to do and to refuse, what it is asked to do, whose time it is asked to save, what it is allowed to replace, and what conditions of human encounter it is required to protect.

Applied to the Community of Inquiry, this reframes the AI question. An AI tutor is not inherently a threat to social presence or a guarantee of teaching presence. The same underlying technology can widen the space for human community or quietly fill it. Whether it becomes an enabler or an extinguisher is determined by two sets of decisions, made at two points: design and deployment. Design decisions shape what the tool is built to do and what it refuses to do, such as whether it asks before it tells, whether it is grounded in expert material, and whether it hands off to a human. Deployment decisions shape the conditions of its use, such as whose time it saves, what it is allowed to replace, who oversees it, and which learners it reaches. A tool designed without careful consideration of its function is a poorly designed tool, and good deployment cannot fully rescue it. A well-designed tool, deployed carelessly, can still extinguish what it was built to support. Both sets of decisions belong to people, and both must be made with the human encounter in view.

This is also why the evidence in What the Evidence Actually Says diverges so sharply, and the divergence illustrates both levels. Bastani's guardrailed and unguarded tutors ran on the same underlying model; what differed was their design, specifically whether the tool was built to give answers or to give hints grounded in teachers' solutions. The Madrid physiotherapy studies in A Health Professions Lens show the deployment level: a similar tool produced little engagement when handed to students without structure, and measurable gains when embedded in a coherent clinical curriculum. The underlying technology was the same; the outcomes were opposite, because the decisions were different.

Attention is the currency of connection

The framework's second principle is that compassion requires attention. Tools that add attention to the human encounter enable it; tools that siphon attention away extinguish it. In clinical care, the patient's nervous system does not measure the technology directly. It measures the attention the technology leaves behind.

The same is true in a learning community. Students do not experience an instructor's AI tools; they experience whether the instructor has more or less attention for them. An AI that absorbs routine grading and frees an instructor for a weekly conversation with each struggling student has widened the attention available for social and teaching presence. An AI that absorbs the feedback conversation itself, so that students receive polished comments from no one in particular, has narrowed it, however good the comments read.

Emulation is not connection

The third principle draws a line that matters more each year. Large language models can emulate the linguistic surface of empathy with remarkable fidelity. They cannot be empathic, because empathy requires a nervous system capable of sharing another's affective state (Singer & Klimecki, 2014). They cannot be compassionate, because compassion, a virtuous response that seeks to address another's suffering through relational understanding and action (Sinclair et al., 2016), requires sentience and an authentic desire to relieve suffering. AI has neither.

This distinction gives precision to the claim that AI cannot supply social presence. Social presence is the experience of being with real people who perceive, care about, and respond to one another. An AI can produce sentences that read as warm, encouraging, and attuned, sometimes more reliably than a depleted instructor between classes. But a student encouraged by a chatbot has not been seen by anyone. The words are present; the person is not.

The practical risk is not that students are fooled once. It is that, over time, emulated warmth comes to feel like enough, and the harder, riskier work of human connection on which belonging depends begins to seem unnecessary.

Four hazards in the learning environment

The framework identifies four hazards that appear consistently when AI is deployed without governance. Each has a direct counterpart in education.

Sycophancy deserves particular attention in education because it is invisible to the student. A learner who receives agreeable, articulate feedback has no reason to suspect that a skilled human mentor would have pushed back. Cognitive presence, in the Community of Inquiry sense, requires productive disagreement and challenge. An AI optimized for user satisfaction tends to remove exactly that friction.

Three questions for every deployment

The framework proposes three structural questions to be asked before approval, again after deployment, and whenever conditions change. Each translates directly to teaching and learning.

1. The displacement question. Does the tool displace administrative or cognitive overhead so the instructor has more time and attention for students, or does it displace parts of the educational encounter itself? An AI that drafts accessibility alt text frees time; an AI that replaces the advising conversation removes the encounter.
2. The oversight question. Is a competent human in the loop, with the authority and the time to override the tool when it is wrong, or has oversight become nominal? AI-drafted feedback that an instructor reads, revises, and signs is on the enabler side. AI feedback released at scale without review is on the extinguisher side, and the rubber-stamping is itself a structural failure no AI can fix.
3. The population question. For a confident, well-supported student embedded in a strong peer network, an AI study partner is a useful adjunct. For an isolated, anxious, or struggling student, the same tool may become a substitute for the human help they most need. Deploying one tool identically to both groups accepts harm to the second in order to capture benefit for the first.

The population question connects directly to equity. The Horizon Report warns that some students may mainly receive automated support while others receive human attention (HR p. 12). The substrate frame explains why that split is not neutral: it tends to fall along exactly the lines of vulnerability where human connection matters most.

What this means for educators

The framework's application to healthcare education (L'HommeDieu, 2026b) adds two points that bear directly on engagement.

First, discernment is becoming a competency. The ability to distinguish AI emulation of empathic language from the actual relational work of teaching or care has to be explicitly taught, modeled, and assessed. Students who cannot tell the difference will drift toward letting AI handle relational tasks and accept the outputs as if they were the work itself. In health professions programs, this belongs alongside differential diagnosis and procedural skill, not in an elective.

Second, faculty modeling is part of the hidden curriculum. Students learn as much from how instructors use technology as from what instructors say about it. Faculty who use AI thoughtfully and name their choices teach discernment. Faculty who silently delegate every relational task to AI teach that delegation is the norm. Faculty development on technology use is therefore a teaching topic, not an IT topic.

The generational context adds urgency. Konrath and colleagues' meta-analysis documented a substantial decline in dispositional empathy among American college students between 1979 and 2009, with partial recovery since (Konrath et al., 2011; Konrath et al., 2023), and Turkle (2015) describes how mediated communication has displaced practice in face-to-face conversation. Students now entering health professions programs may arrive less practiced in embodied presence than earlier cohorts. For them, deliberate attention to social presence and compassion is not a soft addition. It is a structural counterweight.

The decision belongs to those who choose the conditions

The framework closes with accountability. Vendors optimize for the metrics a contract specifies, and instructors absorb whatever conditions the system hands them. The moral weight of technology decisions therefore sits with the people who choose the conditions: in education, the deans, program directors, and committees who decide which tools are adopted, what they are allowed to replace, and whether efficiency gains are reinvested in human relationship or captured as savings.

This returns the argument to the Community of Inquiry. Teaching presence, social presence, and cognitive presence are not features a platform can supply. They are conditions an institution chooses to protect or allows to erode. AI can strengthen all three when it is designed and deployed to liberate human attention. It can hollow out all three when it is designed or deployed to replace human encounter. The underlying technology may be the same either way. The line runs through the decisions, at both the design and the deployment stage, and Part Two is about who makes them.

A working definition

Pulling these threads together, this post uses the following working definition. An engaging AI-supported learning environment is one in which AI raises the quantity and quality of students' constructive and interactive thinking, supports their autonomy, competence, and relatedness, and strengthens rather than replaces the human community of inquiry. Satisfaction and time on task are welcome, but they are not evidence on their own.

An engaging learning environment is not only one where students think hard; it is one where they feel known. AI can protect that, or it can quietly erode it, and the difference again comes down to design and intent.

What the Horizon Report sees

The report's third social trend is titled plainly: AI is reshaping relationships between students and instructors (HR p. 9). Students increasingly rely on AI for explanations and drafting, and some feel more comfortable asking AI than asking faculty. Office-hour visits and informal interactions may decline, reducing opportunities for mentoring. Meanwhile, unclear rules and AI detection tools are breeding mutual suspicion: instructors watch more closely, and students focus on proving they did the work rather than taking intellectual risks.

The report also warns of an uneven future in which some students mostly interact with automated support while others receive human attention (HR p. 12). Because relationships play a significant role in shaping student growth, that split would be an equity problem, not just a service-quality problem.

What students report

The HEPI Student Generative AI Survey 2026, based on 1,054 UK undergraduates surveyed in December 2025, captures the emotional landscape. About 15% of students reported using AI for companionship, advice, or to address loneliness. Roughly equal shares said AI made them feel lonelier (20%) and less lonely (21%). The report describes the overall picture as polarized, and its qualitative comments range from students who say AI frees time for deeper analysis to one who wrote, in effect, that they were no longer using their brain at all.

The Horizon Report's own panel offered a pointed observation on emotional AI. Students may be seeking emotional engagement with AI because they lack meaningful opportunities to engage with humans, and chatbots are filling a void (HR p. 26). That framing shifts the question from "should students use AI for support?" to "what human connection is missing that makes AI the default?"

What the research on chatbot use and wellbeing shows

A four-week randomized controlled study by Fang and colleagues at the MIT Media Lab and OpenAI (n = 981, more than 300,000 messages) tested different chatbot modes and conversation types. In its revised version, the assigned conditions themselves showed no significant effects on the primary outcomes. The striking pattern was elsewhere: participants who voluntarily spent more time with the chatbot each day showed higher loneliness, less socialization with real people, more emotional dependence, and more problematic use. Those who saw the AI as a friend or trusted it highly showed more dependence.

The authors stress that these associations are correlational and that the sample was general adults, not students. Still, the pattern lines up with the Lee findings on trust and critical thinking. In both cognitive and emotional domains, how much a person trusts and leans on the AI seems to matter more than which features the AI has.

Designing for belonging, not substitution

Several Horizon Report examples point toward AI that strengthens human connection rather than replacing it.

• Social presence as a teachable skill. Grand Canyon University's Social Presence Literacy Lab treats connection as a design literacy, giving faculty tools to build belonging and relational trust into technology-rich courses (HR p. 9).
• Students as partners. The University of Utah's student-led AI symposium flipped the usual hierarchy, with students facilitating and faculty listening. Attendees reported seeing AI knowledge, and each other, differently (HR p. 9).
• Peer structures. The University of Denver pairs every incoming student with a trained peer mentor (HR p. 10). In an AI-saturated environment, deliberately designed human relationships become more valuable, not less.
• AI that hands off. Lone Star's advising chatbot is described as freeing advisors for complex cases (HR p. 13). The report calls for clear handoffs from AI to people when needs are complex or high-stakes (HR p. 12).

Where the line sits: emotion tracking

The Horizon Report's most instructive cautionary example is EDmotionsAI, a K to 12 "emotional management system" that tracks students' emotions in real time (HR p. 26). The report judges it unlikely to become a higher education trend because it risks undermining the educator-learner relationship itself, raises privacy concerns, and touches on deeper questions about personhood.

This example is useful because it sits close to legitimate tools. The ELSEI Co-Regulator also infers learner states, but from learning-system behavior rather than biometrics, and it emphasizes transparency and learner agency (HR p. 37). The ethical difference is not whether AI notices student states. It is whether students know, consent, and remain in control, and whether the information flows toward supporting the student or toward monitoring them.

Transparency as a relational practice

California's companion chatbot legislation, which the Horizon Report cites, would require operators to clearly disclose that a chatbot is not human (HR p. 28). Institutions do not need to wait for legislation. Telling students plainly when they are talking to AI, what it can and cannot do, how their data are used, and where to find a human is a basic act of respect.

Transparency runs both ways. Students increasingly ask whether faculty used AI to create materials or grade work. Faculty who model honest disclosure of their own AI use make it easier to ask the same of students, and they begin repairing the trust that detection-first policies have damaged.

The mentor mindset and wise feedback

David Yeager's research on motivation offers one of the clearest accounts of what relational engagement requires, and why AI alone cannot supply it. His work, synthesized in 10 to 25: The Science of Motivating Young People (Yeager, 2024), was developed with adolescents but extends naturally into the late teens and twenties, which includes most undergraduate and many graduate students.

Status, respect, and three mindsets

Yeager's starting point is that young people are acutely sensitive to status and respect. When they interact with socially powerful people such as teachers, coaches, managers, or clinical instructors, they read between the lines of every comment to judge whether they are being honored or dismissed. Feedback is never just information; it is also a message about how the person giving it regards the learner.

From this, Yeager describes three ways adults tend to lead young people, organized along two dimensions: standards and support.

The mentor mindset is not a midpoint between the other two. It holds both dimensions at full strength. Yeager's profiles of effective teachers describe concrete practices: normalizing struggle as part of learning rather than a sign of low ability, asking questions instead of supplying answers, being transparent about the reasons behind expectations, allowing work to be redone, and making sure students know the teacher will be there for them. One physics teacher he profiles coaxes frustrated students to explain their thinking before solving problems with them rather than for them, a description that reads almost exactly like the guardrailed tutor in Bastani's study (Evans, 2025).

Wise feedback: the evidence

The mentor mindset has a well-tested core practice: wise feedback, a term originating in work by Cohen, Steele, and Ross (1999). In a set of randomized field experiments, Yeager and colleagues (2014) had seventh-grade teachers give critical feedback on student essays. Half the students also received a brief handwritten note conveying, in essence, that the teacher was giving these comments because they held very high standards and believed the student could reach them. Students who received the note were roughly twice as likely to revise and resubmit their essays. The effects were strongest for African American students, especially those who had begun to lose trust in school, a group for whom critical feedback can easily be read as a sign of bias rather than belief.

Later analyses of the same cohort found that this brief trust-restoring treatment, delivered in spring of seventh grade, improved African American students' eighth-grade discipline outcomes and later four-year college enrollment (Yeager, Purdie-Vaughns, Hooper, & Cohen, 2017). A few sentences that resolved a question of trust changed the trajectory of the relationship, and with it the trajectory of the student.

Yeager emphasizes that wise feedback is not a magic phrase. The note worked because it was credible: it came from a real teacher who actually held those standards and actually believed in that student. Delivered by someone who does not mean it, the same words would likely fail or backfire.

What this means for AI

Yeager's framework sharpens several arguments already made in this post.

AI defaults tend to fall into the enforcer and protector quadrants. A sycophantic chatbot that validates weak work and smooths away difficulty is a protector: high support, low standards. An answer engine that does the work for a struggling student is also a protector, because it removes the challenge rather than helping the student meet it. On the institutional side, AI detection regimes are enforcers: high standards, enforced through suspicion, with little support. Neither produces the growth the mentor quadrant does. The guardrailed tutors that worked in Bastani's and Kestin's studies look, by contrast, like attempts to engineer mentor-like behavior: hold the standard, refuse to do the work, and scaffold the student toward it.

Wise feedback depends on a credible relationship, which AI cannot supply. The power of the wise feedback note lay in a social signal: a person with authority, who knows this student, believes they can meet a high bar. This is exactly the line drawn in the substrate framework. An AI can generate the sentence "I have high standards and I know you can reach them" with perfect fluency. But no one is holding those standards, and no one believes anything about the student. The words are present; the mentor is not. Early work from Yeager's own group suggests that untrained AI-generated motivational messages were perceived by adolescents as conveying less of a growth mindset than teacher-written messages, with prompt-trained AI coming closer (Clapper et al., 2023); this was a conference poster, and its published abstract is ambiguously worded, so it should be read as preliminary.

The workable design is a division of labor. AI can help with the parts of feedback that are informational: identifying specific gaps against a rubric, suggesting where an argument needs evidence, flagging errors in reasoning. An exploratory meta-analysis of 41 studies found no significant difference in learning performance between AI-generated and human feedback, and its authors recommended a hybrid model that keeps the empathetic and contextual features of human feedback (Kaliisa et al., 2025). Wise feedback suggests what those human features are. The instructor reviews the AI-supported comments, decides what matters most, and adds the part only a person can credibly give: a statement of high standards, a statement of belief in the student, and an invitation to revise with help available. That combination preserves both halves of the mentor mindset. It also answers the oversight question from the substrate framework: the human is not rubber-stamping, but supplying the relational core.

Mentor-mindset AI policy for students. The same logic applies to how institutions frame AI rules. Telling students "AI use is cheating and we will catch you" is an enforcer message. Telling them "use whatever helps" is a protector message. A mentor message sounds different: here is what this assignment is meant to build in you, here is why AI would short-circuit it, we expect you to do this thinking yourself, and here is the support available if you get stuck. Students who understand the reason behind a standard, and believe the people setting it are on their side, are far more likely to meet it.

For clinical educators, the mentor mindset is already familiar. The best clinical instructors hold students to the standard patient safety requires while making it clear they will help them reach it. Wise feedback names what those instructors do intuitively, and it explains why that work cannot be delegated to a tool, however fluent.

A practical test for relational engagement

After any AI-supported activity, ask: do students now have more to talk about with their instructor and peers, or less? Good designs generate human conversation. A simulation produces a debrief. A feedback tool produces a revision conference. A curiosity exercise produces a class discussion of the best questions. Poor designs end the conversation at the screen.

Everything in Part One applies with more force in programs that prepare clinicians, because the skill must transfer to unassisted performance with a real patient. A physical therapist reasons at the bedside, often without a tool and always without time to verify one. Bastani's finding, that learners who leaned on unguided AI performed worse once it was removed, reads like a warning written for licensure preparation and clinical readiness.

Why the stakes are higher

Three features of clinical education sharpen every concern raised so far.

• Transfer to unassisted performance is the whole point. A DPT student who can produce an excellent AI-assisted plan of care but cannot reason through a patient's presentation independently has not met the standard.
• Errors carry patient consequences. Lee and colleagues found that critical thinking shifts toward verifying AI output and that verification is weakest where expertise is lowest. Novice clinicians are that group.
• Privacy is regulated. Entering identifiable patient information into an unvetted tool is a compliance failure, not a judgment call. Students need explicit rules and practice de-identifying cases before they ever reach a clinical site.

Two studies from the same group tell the design story

Physical therapy now has its own small but revealing evidence, and two studies from the same Madrid research group illustrate the design principle almost perfectly.

In a pilot randomized study published in JMIR Formative Research (Ferrer-Peña et al., 2025), 46 third-year physiotherapy students were assigned either to use ChatGPT as a virtual patient for four weekly clinical cases or to follow the usual curriculum. Students received predefined prompts but, deliberately, no training in how to use the tool. Engagement collapsed. Only four or five of the 23 students assigned to the tool actively used it (the paper reports both figures). Students informally cited unfamiliarity with AI, preference for human interaction, perceived extra workload, and technical difficulties. Satisfaction did not differ between groups.

A year later, the same team reported a different approach at the EDULEARN26 conference (Pérez-González et al., 2026). This time the LLM simulator was embedded in a structured course built on the International Classification of Functioning, Disability and Health (ICF), configured to portray three patients with complex chronic pain, and integrated with ICF-based assessment assignments. Twenty-eight final-year students completed six simulator interactions before their clinical examination; nine controls followed the standard curriculum. Tutors rated the intervention group substantially higher on the Lasater Clinical Judgment Rubric (Cohen's d = 0.99). Simulator performance correlated moderately to strongly with real-patient clinical scores (rho = 0.618). And, tellingly, student satisfaction was unrelated to clinical performance.

The second study is a quasi-experimental pilot with a very small control group, so its effect size should be read cautiously. But the contrast between the two is instructive. Unstructured access produced little engagement. A simulator embedded in a coherent clinical framework, with a clear purpose and assessment link, produced measurable gains. The satisfaction finding echoes Deslauriers: how students feel about a learning experience is not a reliable guide to what they learned.

What medical education adds

Medical education is further along. A 2026 systematic review in JMIR Medical Informatics found that LLM-based virtual patient systems for history taking can simulate many conditions but rarely represent multimorbidity, which is precisely the complexity physical therapists face daily. Outpatient PT caseloads are full of patients with diabetes, obesity, depression, and multiple joint complaints at once. Simulators built on single-condition cases risk teaching an overly tidy version of clinical reasoning.

The AAMC's Principles for the Responsible Use of AI in and for Medical Education (version 2.0, July 2025) offer a useful benchmark that translates readily to physical therapy. The seven principles are: maintain a human-centered focus; ensure ethical and transparent use, including preparing learners to explain technology use to patients; provide equal access; foster education and continuing professional development for educators; develop curricula through interdisciplinary collaboration; protect data privacy across admissions, teaching, simulation, assessment, and evaluation; and monitor and evaluate tools where they are actually used.

Two of these deserve emphasis for PT programs. Preparing students to explain AI use to patients is a communication competency, not a technical one, and it belongs in professionalism curricula. And monitoring tools in their place of use means evaluating learning outcomes, not just usage.

Practice from a DPT classroom

Touro University's DPT program offers a concrete classroom example. Through a faculty innovation grant, third-year students used a generative AI assistant to design age-specific weekly physical activity plans for children, following structured prompting guidelines and evidence-based activity standards. Students then peer-reviewed each other's plans to compare how different prompts shaped outputs and completed an individual guided reflection. Faculty reported that students quickly noticed the absence of evidence and references in AI-generated content and understood their responsibility to scrutinize it before use.

The design carries several of the moves described earlier: structured prompts, a defined clinical standard, peer comparison that made AI variability visible, and reflection that turned AI use into an object of judgment.

A model sequence for clinical skills

Combining the evidence, a defensible sequence for teaching a clinical skill such as a subjective examination and differential reasoning might look like this:

1. Prepare. Students review content and complete retrieval practice without AI.
2. Rehearse with AI. Students practice with an AI standardized patient, configured for realistic complexity and multimorbidity, that withholds diagnoses and responds only to what is asked.
3. Defend to a human. Students explain and justify their reasoning to a faculty member or peer group, with the AI transcript as evidence of their process.
4. Perform with hands. Students examine a real person in lab or clinic, where palpation, observation, handling, and the therapeutic relationship cannot be simulated.
5. Reflect. Students compare their AI rehearsal with their human performance and identify what transferred and what did not.

In this sequence AI multiplies rehearsal; humans verify judgment; the body teaches what only bodies can. The AI never does the reasoning. It creates more occasions for the student to do it.

Protected competencies

Health professions programs may also need a short, explicit list of competencies that must always be demonstrated without AI, the clinical equivalent of Sydney's secured "Lane 1." For DPT education, strong candidates include examination and differential diagnosis in real time, recognition of red flags requiring referral, safe patient handling and guarding, communication of risk to patients, and documentation of one's own clinical reasoning. Naming these openly helps students understand that limits on AI are about patient safety, not suspicion.

The central fact for leaders is a gap: AI use by students and staff is nearly universal, while institutional strategy, policy, tools, and training lag well behind. Support has to start from that reality, not from an imagined starting line where everyone is waiting for permission.

Students are already there

The HEPI 2026 survey of 1,054 UK undergraduates found that 95% use AI in at least one way and 94% use it for assessed work. The share directly including AI-generated text in assessed work rose to 12%, from 8% in 2025 and 3% in 2024. Nearly two-thirds (65%) said assessment had changed significantly in response to AI.

Institutional support lags. Only 36% felt encouraged by their institution to use AI, only 38% said they were provided with AI tools, and fewer than half (48%) felt teaching staff were helping them build AI skills for their careers, even though 68% considered those skills essential. Arts and humanities students were particularly likely to feel under-supported.

Institutions are working in pockets

In the United States, the 2025 EDUCAUSE AI Landscape Study surveyed nearly 800 higher education professionals in late 2024. As reported by GovTech, 57% said their institution considered AI a strategic priority, up from 49% the year before. Yet 55% said AI strategy was happening in pockets rather than through an institution-wide approach. Only 2% reported new funding sources for AI costs, 39% had AI-related acceptable use policies, and just 9% felt their cybersecurity and privacy policies were adequate for AI risks. Larger institutions were more likely to have policies and to treat AI as an investment, a gap EDUCAUSE described as a digital AI divide.

EDUCAUSE's January 2026 report, The Impact of AI on Work in Higher Education (Robert, 2026), adds a sharper picture of faculty and staff. Among 1,960 respondents:

• 94% had used AI tools for work in the past six months, but only 54% were aware of any policies or guidelines for that use.
• 56% had used AI tools that their institution did not provide.
• 23% of faculty said their institution provided none of the AI tools they wanted for work, compared with 10% of all respondents.
• 63% of faculty had used AI to create learning activities or assessments.
• Only 13% said their institution was measuring the return on investment of AI tools.
• The top challenges were AI's pace of change (60%), lack of expertise (55%), lack of best practices (48%), and lack of time to learn new skills (46%).
• Among the most-cited risks was the loss of fundamental skills requiring independent thought (51%).

What the gap means

Three implications follow for leaders.

First, policy vacuums do not stop use; they only make it unsafe. EDUCAUSE found that uncertainty about policy almost never deterred people from using AI. Faculty and students are filling the vacuum with whatever tools they can find, often unvetted ones. Clear guidance and vetted tools are therefore risk management, not just service.

Second, faculty are being asked to lead a transformation they were not resourced for. They are redesigning assessments, learning new tools on their own time, and absorbing student anxiety, frequently without institution-provided tools or training. EDUCAUSE explicitly recommends codifying AI-related responsibilities in job descriptions partly to mitigate burnout among staff and faculty asked to take on more work. That recommendation deserves attention, and Protecting the People Who Do the Work returns to it.

Third, the lack of measurement is itself a strategic risk. With only 13% measuring return on investment, most institutions cannot tell whether their AI investments improve learning. Given Bastani's warning that harm can hide behind rising grades, institutions need outcome measures that include delayed and unassisted performance, not just adoption and satisfaction.

The Horizon Report's framing

The Horizon Report places all of this inside a larger squeeze: declining enrollments, strained budgets, and pressure to prove value (HR pp. 8 and 14). It warns that if decision-makers treat AI as a way to do more with fewer people, the result could be heavier teaching loads, reduced support roles, and lower-quality feedback and mentoring (HR p. 11). That warning frames everything in the sections that follow. The question for leaders is not only how to support AI adoption, but how to do so in a way that protects the human work AI cannot replace.

The single most useful thing leaders can give faculty and students is clarity about what AI use is expected, permitted, and prohibited, assignment by assignment. The second most useful is permission and support to stop relying on AI detection and start verifying learning directly.

Why clarity comes first

Ambiguity is the root of much of the anxiety the Horizon Report describes. When rules are unclear, instructors watch more closely and students focus on proving innocence rather than taking intellectual risks (HR p. 9). HEPI's respondents described anxiety about false accusations of misconduct. The EDUCAUSE work study found only about half of faculty and staff felt confident that existing policies gave them clarity.

Clarity is also a learning condition, not just a compliance condition. Self-determination theory predicts that students are more autonomously motivated when the boundaries of a task are clear and the reasons for them are explained. A student who knows why an assignment prohibits AI is more likely to accept the limit than one who simply suspects a trap.

Frameworks that work at the assignment level

Two approaches stand out because they operate where learning actually happens: in individual assignments.

National University's GAiIT framework (Generative AI Inclusion Threshold) defines five icon-labeled levels of permitted AI use and embeds the specified level in every course assignment (HR p. 21). The icons create a shared visual language that students learn once and recognize everywhere.

The University of Sydney's two-lane approach goes further by linking AI permission to assessment security. As described by Sydney's educational innovation team (Liu & Bridgeman, 2023), "Lane 1" assessments are supervised and assure that students have mastered required knowledge and skills; these are assessments of learning. "Lane 2" assessments are unsupervised, assume AI may be used, and focus on motivating learning and teaching responsible AI engagement; these are assessments for learning. Sydney's guidance makes one point with unusual honesty: an unsupervised assessment that prohibits AI is treated as an invalid combination, because there is no reliable way to prevent or detect AI use when nobody is watching. Lane 1 assessments should be used sparingly and strategically, ideally at the program level, while most assessments sit in Lane 2.

The Sydney model aligns with guidance from Australia's higher education regulator. TEQSA's Assessment reform for the age of artificial intelligence (Lodge, Howard, Bearman, Dawson, & Associates, 2023), developed through expert collaboration, argues that assessment should engage students appropriately and authentically with AI, and that trustworthy judgments about learning require multiple, inclusive, and contextualized forms of evidence rather than any single high-stakes task.

Why detection is the wrong foundation

AI detection fails on both accuracy and fairness. Liang and colleagues (2023), in a study published in Patterns, tested several widely used GPT detectors and found that they consistently misclassified non-native English writing as AI-generated while accurately identifying native writing. Simple prompting strategies, meanwhile, allowed AI-generated text to bypass detection entirely. A tool that flags honest multilingual students and misses determined cheaters is worse than no tool.

The Horizon Report notes growing distrust of detectors, including lawsuits over false positives, and cites a policy article arguing that detection tools should not be used in higher education because of methodological imperfections, procedural unfairness, and unverifiable outputs (HR p. 27). Its panel captured the human cost: students fear false accusation, and that fear is eroding classroom trust.

Grand Canyon University's response is instructive. It eliminated reliance on AI detection and adopted a verification-centered framework in which faculty verify student mastery through professional judgment and structured learning checks (HR p. 20). Reported benefits included reduced investigative burden, restored instructional authority, and integrity efforts refocused on learning outcomes.

Yeager's framework from Relational and Emotional Engagement clarifies why this shift matters. Detection-first integrity is an enforcer posture: high standards communicated through suspicion, with little support. Students read it, accurately, as a statement that they are not trusted. Verification can be designed as a mentor posture instead: the standard stays high, the reason for it is explained, and the conversation in which a student demonstrates understanding becomes an occasion for feedback and support rather than prosecution.

What verification looks like in practice

Verification is not a single technique. It is a portfolio of evidence that a student can do what the credential claims.

• Oral explanation. Brief viva voce conversations or interactive oral assessments in which students explain and extend their submitted work.
• Observed performance. Practical examinations, OSCE-style stations, live problem solving, or in-class writing.
• Process evidence. Drafts, version histories, AI interaction logs, and reflective notes showing how the work developed.
• Transfer tasks. Short unassisted tasks that apply the same concept in a new context, the direct counterpart of Bastani's "access removed" exam.
• Program-level checkpoints. A small number of secured assessments placed strategically across a program rather than in every course.

For health professions, much of this already exists. Practical examinations, clinical competency checks, and supervised clinical education are forms of verification that AI cannot easily fake. Leaders in these programs can build on strengths rather than inventing new systems.

Teaching AI use as a learning outcome

The Horizon Report notes a shift underway: rather than asking students to prove the absence of AI, some institutions are asking them to demonstrate thoughtful and safe AI use as a feature of assessment (HR p. 27). HEPI's recommendations point the same way, urging institutions to teach AI knowledge and skills explicitly at both general and subject-specific levels and to publish clear, assessment-specific guidance.

In practice, this can mean Lane 2 assignments with rubric criteria for AI use itself: Did the student verify claims against sources? Did they recognize and correct an AI error? Did they explain what they accepted, rejected, and why? These criteria turn AI literacy into a graded, teachable competence.

What leaders can do

1. Adopt a shared, assignment-level labeling system across the institution or program, so students learn one language.
2. Retire AI detection as evidence in misconduct cases, and say so publicly.
3. Fund and protect time for faculty to convert high-risk unsupervised assessments into either secured or AI-integrated designs.
4. Map Lane 1 assessments at the program level, so verification is concentrated where it matters most rather than duplicated in every course.
5. Treat a clear AI policy as a teaching resource: explain the reasons, not just the rules.

Faculty development for AI works best when it is treated as a social, disciplinary, and pedagogical process, not a technical training event. Most institutions are currently relying on the least effective version: EDUCAUSE found that 80% of respondents said their institution was encouraging faculty and staff to develop AI skills on their own.

Why tool training is not enough

The TPACK framework (Mishra & Koehler, 2006) explains the problem. Effective technology integration requires the intersection of three kinds of knowledge: technological, pedagogical, and content knowledge. A workshop that teaches faculty how to prompt a chatbot builds only technological knowledge. It does not help a physical therapy instructor decide whether an AI patient simulator supports or short-circuits clinical reasoning in her particular course. That judgment lives at the intersection of all three.

The research in Part One makes the same point from a different angle. The difference between Bastani's harmful tool and Kestin's effective tutor was pedagogy, and the difference between the Madrid group's failed pilot and its successful simulator was curricular integration. Faculty need support at exactly that intersection.

The two foundations introduced in Design Principles offer a ready-made curriculum for this kind of faculty development. Backward design (Wiggins & McTighe, 2005) gives faculty a shared planning language that places AI decisions after goals and evidence, and scientific teaching (Handelsman et al., 2007) gives them an evidence-gathering habit for testing whether those decisions worked. Both have long track records in faculty development, which means institutions can build AI support on frameworks faculty may already know rather than inventing new ones.

Communities of practice

Wenger's (1998) concept of communities of practice describes groups of people who share a concern, deepen their expertise through regular interaction, and develop shared repertoires of practice. AI adoption fits this model well, because the knowledge needed is new, local, and evolving faster than any formal curriculum.

The Horizon Report offers several working examples:

• CUNY's Don't AI Alone is a grassroots community that maps, connects, and showcases AI work across 26 campuses, with a living database of more than 120 projects (HR p. 10).
• The Situated Social Mentorship Framework addresses uneven AI readiness and anxiety through structured self-assessment, peer mentorship, and instructional design support, moving faculty from isolated experimentation to collaborative practice (HR p. 9).
• Instructional Design Fellows trains faculty as peer mentors who conduct collaborative course reviews, a low-cost and replicable model (HR p. 10).
• AI and Accessibility Snackables offers short, problem-focused modules, each supporting one actionable instructional improvement, reinforced through lab sessions (HR p. 13).
• Society and AI treats open, scholar-led publishing as shared infrastructure for developing professional judgment about AI in teaching (HR p. 9).

The common thread is that faculty learn from colleagues facing the same problems in similar contexts. A nursing instructor learns more from another clinical educator's experiment with AI simulation than from a vendor demonstration.

Students as partners

Students use AI more than faculty, a gap the EDUCAUSE 2025 study noted and that EDUCAUSE respondents in 2026 named among the most urgent risks (47% cited student AI use outpacing faculty and staff skills). That gap is usually framed as a threat. It can also be a resource.

The students-as-partners movement in higher education (Healey, Flint, & Harrington, 2014) argues that students can contribute meaningfully to curriculum, pedagogy, and research as co-creators rather than consumers. Two Horizon Report examples apply this directly to AI:

• Northwestern's TRAIL Guides employs undergraduates as AI literacy guides who consult with individual faculty on assignment redesign, co-facilitate faculty workshops, and write about how students actually use AI (HR p. 39). The program explicitly reframes students from risks to be managed into partners in shaping responsible use.
• The University of Utah's Student-Led AI Symposium inverted the usual roles, with students setting the agenda and leading every session while faculty listened. More than 130 people attended, and a weekly AI Tinker Lab followed (HR p. 9).

These models also address the trust problem. When students help design AI policy and practice, they are less likely to experience it as surveillance.

Tools that build engagement into course design

Some existing platforms already embody the principle of AI that supports rather than replaces human engagement, and faculty development can introduce them as concrete starting points. Perusall is a useful example for reading-heavy courses, including DPT and other health professions programs. Students annotate assigned readings together in a shared, social space before class, asking questions and answering each other. The platform uses automated scoring of annotation engagement and generates confusion reports and discussion summaries so instructors can see where students are struggling and plan class time accordingly. The company reports that the large majority of students complete readings on the platform, a vendor figure worth testing locally.

Perusall illustrates several design moves from Design Principles: the work of making sense of the reading stays with students, peers interact around it, and AI supports the instructor's teaching presence rather than substituting for student thinking. It also fits the flipped model Kestin and colleagues recommend, with preparation before class and higher-order work during it.

Time, recognition, and expectations

The EDUCAUSE work study found lack of time to learn new skills among the top challenges, and recommended codifying new AI-related responsibilities in job descriptions. That recommendation is more important than it sounds. Assessment redesign, simulation development, and AI literacy teaching are real work. If they are invisible in workload models, promotion criteria, and annual reviews, they will be done by a committed few until those few burn out.

Leaders can make this work visible in concrete ways:

• Count AI-related course redesign as scholarship of teaching and learning when it is evaluated and shared.
• Offer course releases or stipends for faculty who redesign high-enrollment or high-stakes courses.
• Recognize community-of-practice leadership in service expectations.
• Fund small grants, like Touro's faculty innovation grant, that let faculty pilot and evaluate ideas in their own courses.

A practical development pathway

The pathway matters because faculty start in different places. EDUCAUSE found 23% of faculty lacked institutional access to any AI tools they wanted, while others were already building custom assistants. One-size training fails both groups.

Good governance for AI is proportional: light and fast for low-risk experimentation, careful and deliberate for uses that touch student data, patient information, high-stakes decisions, or accessibility obligations. Most institutions currently have the opposite problem, with slow committees for everything and no guidance for most of what people actually do.

Governance that can keep pace

The Horizon Report highlights two complementary approaches.

Frontier Nursing University is treating governance itself as "policy infrastructure" for teaching innovation (HR p. 37). Recognizing that traditional committee structures can unintentionally slow change, it is piloting a domain-based model meant to clarify authority pathways, reduce redundancy, and improve coordination. Notably, the changes are still in consultation and iterative review, with faculty and staff engaged in dialogue rather than handed a finished structure.

The S.E.C.U.R.E. framework (Security credentials, Ethical use, Confidential information, Use of personal data, Rights protection, Evaluation of outputs) is a Creative Commons licensed screening tool that lets staff check their own generative AI uses against six risk domains (HR p. 20). It enables low-risk use without formal approval while flagging uses that need more scrutiny. That is proportional governance in a single page.

The principle for leaders is to reserve committee time for decisions that genuinely need it, and to give everyone else a fast, transparent self-check for routine use.

Privacy and security

The Horizon Report's technological trends warn that student and faculty data are increasingly at risk, and that AI amplifies the risk both when people enter sensitive data into unvetted tools and when AI makes phishing easier to produce at scale (HR p. 11). EDUCAUSE's data make the exposure concrete: 56% of faculty and staff respondents had used AI tools their institution did not provide, and in 2025 only 9% felt their institution's cybersecurity and privacy policies were adequate for AI risks.

Two strategies reduce this exposure.

• Provide vetted tools. When institutions supply capable, approved tools, people have less reason to use unvetted ones. Boise State's boisestate.ai runs inside the university's own cloud environment, protects data, avoids vendor lock-in, and gives all 27,000 students and faculty access at lower cost than commercial subscriptions (HR p. 12). The Netherlands has gone further: all 113 public tertiary institutions are co-creating a shared, privacy-first platform on Dutch sovereign infrastructure rather than buying a commercial product (HR p. 38).
• Teach data judgment, not just rules. In health professions programs especially, students need explicit practice recognizing protected health information and de-identifying cases before using any AI tool. The AAMC's privacy principle extends this across admissions, simulation, assessment, and advising.

Accessibility

Accessibility is both an obligation and an opportunity. Under the federal ADA Title II rule, public colleges and universities must make websites, mobile apps, and digital course materials conform to WCAG 2.1 Level AA. In April 2026 the Department of Justice extended the compliance date for large public entities to April 26, 2027, and a parallel HHS rule under Section 504 now requires colleges receiving HHS funding, public or private, to comply by May 11, 2027, as summarized in EDUCAUSE Review. The HHS rule matters for health professions schools, which often receive HHS funding.

The Horizon Report notes a real risk here: without hands-on support, some instructors may avoid video and media-rich materials simply because they cannot make them accessible in time, narrowing learning options (HR p. 19). AI can help, and several examples use it for exactly this purpose, such as the Title II readiness course that pairs accessibility guidance with custom AI prompts and decision matrices (HR p. 21). But the report is clear that human review remains necessary and that the work should not fall entirely on individual faculty.

The convergence is worth naming: AI-supported accessibility remediation is one of the clearest cases where AI can reduce faculty burden without reducing learning, provided leaders fund the review and support that make it reliable.

Equity of access

The Horizon Report warns repeatedly about a digital AI divide, between institutions and between students (HR pp. 12 and 24). HEPI found only 38% of students said their institution provided AI tools. When institutions provide nothing, students who can pay for better tools gain an advantage. The AAMC names equal access as one of its seven principles for exactly this reason.

Equity is not only about licenses. The Kenyan Three Dots pilot shows that in mobile-first, low-bandwidth settings, the delivery channel itself determines who can participate (HR p. 39). And the uneven-support concern applies here too: if some students mainly get automated help while others get human mentoring, AI could widen gaps it was meant to close (HR p. 12).

Practical equity measures include institution-provided access to capable tools for every student, explicit AI orientation for all incoming students (a HEPI recommendation), monitoring who uses AI support and who receives human support, and attention to students with limited devices or connectivity.

Environmental cost

The Horizon Report identifies growing awareness of AI's environmental footprint as an environmental trend (HR p. 17). Estimates vary, but greater reliance on AI means more data center electricity and cooling. The report suggests AI policies and procurement should expand to include sustainability expectations, that AI literacy should encourage intentional rather than default use, and that some students may resist AI for environmental reasons, which pushes faculty to be clear about when AI use is required versus optional.

This consideration reinforces the design principle from Part One. If AI is used where it demands thinking and adds real learning value, and avoided where it merely substitutes for effort, the environmental and pedagogical cases point the same way.

A proportional governance map

An AI transition handled carelessly is a textbook setup for faculty occupational distress: rising demands, shrinking control, unclear standards, eroding relationships, and work that goes unrecognized. Leaders who want AI to improve teaching must protect the people who teach.

Faculty are worried, and stretched

A national survey by Elon University and the American Association of Colleges and Universities, released in January 2026, found deep faculty concern. Ninety-five percent expected generative AI to increase student overreliance on these tools, 90% expected it to diminish students' critical thinking, and 83% expected shorter attention spans. Eighty-six percent believed AI would likely affect the work and role of those who teach.

The workload picture is mixed. Tyton Partners' Time for Class 2025 survey, summarized by D2L, reported that among instructors using generative AI daily, 36% experienced a decrease in workload, 26% an increase, and 39% no change. The same survey found that trying to monitor student AI use had created more work for instructors and administrators. AI can save time on some tasks while generating new labor in others, especially surveillance and redesign.

Two other findings from that survey are worth holding onto. First, 84% of students said they turn to people, an instructor or peers, when struggling in a course, while instructors overestimated how often students turn to AI instead. Second, only 28% of institutions had an active generative AI policy. Students still want human help; institutions are still not giving faculty clear frameworks.

A burnout lens

Two established frameworks explain why the current moment is risky.

The Job Demands-Resources model (Bakker & Demerouti, 2007) holds that burnout results when job demands persistently exceed job resources, while engagement grows when resources such as autonomy, support, feedback, and development keep pace. AI adoption adds demands: learning tools, redesigning assessments, investigating suspected misconduct, managing student anxiety, meeting accessibility deadlines. Unless resources grow alongside, the balance tips.

Maslach and Leiter's work on areas of worklife identifies six domains where mismatches between people and their jobs drive burnout: workload, control, reward, community, fairness, and values. An AI transition can strain every one of them.

The values row deserves particular attention. Most faculty, and especially clinical faculty, entered education because of what happens between people: the mentoring conversation, the moment a student's reasoning clicks, the professional identity that forms through relationship. If AI is introduced in a way that threatens those experiences, the resulting distress is not simply about workload. It is a conflict with the reason people do the work at all.

Yeager's mentor mindset applies to leaders as much as to teachers. Faculty experience their institution's AI strategy the way students experience feedback: as a signal of how they are regarded. A leadership team that issues new expectations for assessment redesign, AI literacy, and accessibility without time, tools, or training is acting as an enforcer. One that lowers expectations to avoid friction, leaving every instructor to improvise, is acting as a protector. The mentor posture holds faculty to a genuinely high standard of AI-informed teaching and pairs it with the support needed to meet it, and says so explicitly.

The "do more with fewer" trap

The Horizon Report names the trap directly. If decision-makers treat AI as a way to do more with fewer people, the result could be increased teaching loads, reduced support roles, and lower-quality feedback and mentoring (HR p. 11). Under budget pressure (HR p. 14), the temptation is real.

The evidence in Part One suggests why it would be self-defeating. The AI successes, Kestin's tutor, the Madrid simulator, Bodyswaps with debriefing, all depended on expert human design and human consolidation. AI raised the value of expert faculty time rather than replacing it. An institution that uses AI to cut faculty would be removing the very ingredient that made the AI effective.

Using AI to give time back

There is a better path. AI can absorb some genuinely low-value work, and the savings can be redirected deliberately toward the human core.

• Administrative offloading. EDUCAUSE respondents most often named automating repetitive processes (70%) and offloading administrative burdens (65%) as AI's promising opportunities.
• Routine student questions. Lone Star College's advising chatbot is described as freeing advisors for complex cases (HR p. 13).
• Accessibility remediation. AI-assisted workflows can speed captioning, alt text, and document repair, with human review (HR p. 21).
• Preparation insight. Tools that summarize where students are confused before class can help faculty spend class time where it matters.

The leadership move is to make the reinvestment explicit. If an AI tool saves advisors ten hours a week, say publicly where those hours go: more mentoring appointments, smaller advising caseloads, or time for outreach to struggling students. If the savings silently become budget cuts, faculty will learn that efficiency is a threat, and adoption will stall.

Supporting students' wellbeing too

Students carry their own version of this strain. HEPI respondents described anxiety about false misconduct accusations, and about 15% reported using AI for companionship or to address loneliness. The Fang study's association between heavy chatbot use and loneliness is a reminder that AI can quietly displace the human contact students need. The same leadership principle applies: use AI to expand access to human support, not to substitute for it, and make human help easy to find.

What leaders can do

1. Treat AI adoption as a workload change and budget for it, including release time and recognized service.
2. Commit publicly that efficiency gains will be reinvested in teaching and mentoring, and report on it.
3. Involve faculty and students in tool selection, policy, and evaluation, which restores control.
4. Retire detection-first integrity regimes that add monitoring labor and damage relationships.
5. Monitor faculty wellbeing alongside AI adoption metrics, using established burnout and engagement measures.

Everything in this post comes together in a single idea: AI should raise the demand for human thinking and human connection, and leaders should build the conditions that make that possible. The tools below translate that idea into decisions.

For any AI use in teaching: four questions, in order

This framework is a synthesis of the studies reviewed here, not an established model. It should be tested against local experience and revised.

For academic leaders: a ten-point checklist

• [ ] Publish an assignment-level AI labeling system that students and faculty share across the program.
• [ ] Map secured ("Lane 1") assessments at the program level, including protected clinical competencies in health professions.
• [ ] Retire AI detection as evidence of misconduct; fund verification through oral, observed, and process-based assessment.
• [ ] Provide vetted AI tools to every student and faculty member, with clear data rules.
• [ ] Offer a proportional governance path: self-check for low risk, review for high risk.
• [ ] Build faculty development around communities of practice and student partners, not one-time tool training.
• [ ] Count AI-related course redesign and mentoring in workload, review, and promotion.
• [ ] Commit publicly to reinvesting AI efficiency gains in teaching and mentoring.
• [ ] Measure learning outcomes, including delayed and unassisted performance, not just adoption and satisfaction.
• [ ] Track faculty and student wellbeing alongside AI metrics.

The two failure modes to avoid

The evidence points to two symmetrical mistakes.

The first is uncritical adoption: deploying general-purpose AI at scale, celebrating usage and satisfaction, and missing the quiet erosion of learning that Bastani and Fan documented. Its signature is rising grades with no evidence of durable learning.

The second is defensive refusal: banning AI, relying on detectors, and retreating to surveillance. Its signature is rising suspicion, lost trust, unfair accusations, and graduates unprepared for workplaces where AI is ordinary.

The path between them is design: intentional, evidence-informed, human-centered use, supported by leaders who supply clarity, tools, time, and protection.

Yeager's three mindsets map neatly onto these choices. Uncritical adoption is the protector's error at institutional scale: high support, in the form of always-available AI, with the standard for learning quietly lowered. Defensive refusal is the enforcer's error: high standards defended through surveillance, with support withdrawn. The design path is the mentor's: keep the standard for genuine learning high, and invest in the tools, relationships, and time that help students and faculty meet it.

The evidence so far does not say AI is good or bad for learning. It says design decides, and that leadership decides whether good design is possible.

On engagement, the research converges on a principle older than AI: people learn by doing the thinking. AI tools that ask before they tell, simulate situations that demand a response, ground themselves in expert material, make the learner's process visible, and hand off to humans for consolidation can multiply opportunities for that thinking. Tools that answer, draft, and summarize on demand can raise output while quietly lowering learning. The difference is often invisible in satisfaction scores and grades, which is why it matters so much.

On leadership, the evidence is just as consistent. Students and faculty are already using AI, mostly without clear guidance, vetted tools, or recognized time. The most valuable things leaders can supply are not technologies but conditions: clarity at the assignment level, verification instead of detection, communities of practice, proportional governance, equitable access, and protection for the human relationships that make education worth having.

In health professions education the stakes are concrete. A physical therapist must reason, communicate, and touch with judgment that no tool can lend at the bedside. AI can give students far more rehearsal than any program could provide before. It cannot give them the judgment itself, and it should never be allowed to replace the mentors who help that judgment form.

The Horizon Report ends by noting that it does not predict a single future but surfaces patterns that can inform strategy. The pattern here is clear enough to act on.

Questions to carry forward

• Which competencies in your own field must always be demonstrated without AI, and who should decide?
• If an institution's grades rose after adopting AI, as Bastani's practice scores did, how would anyone notice if learning fell?
• What would it look like to measure an AI initiative by the hours of mentoring it created rather than the hours of labor it removed?
• The Horizon Report's panel is drawn largely from instructional technology and teaching centers. How might clinicians, students, and patients frame these questions differently?
• If AI makes faculty work easier in some ways and harder in others, who in the institution is responsible for noticing the difference, and for acting on it?