HCC 3030 · Week 7
AI in Education What counts as learning?
The real promise is not easier schoolwork. It is better practice, better teaching, wider access, and new ways to build human capability.
LEARN
PRACTICE FEEDBACK TRANSFER AGENCY
This lecture asks a simple but demanding question: when AI helps a student produce a correct answer, did the student learn?
We will treat education as a broad system that includes schools, universities, job training, professional practice, accessibility, and lifelong learning. The opportunity is enormous, but only if the design protects the cognitive work that creates durable capability.
[Sources]
A useful opening problem
Two students submit equally excellent work.
STUDENT A Used AI to test ideas, request hints, and improve a draft. Can explain every choice and solve a new problem alone.
STUDENT B Used AI to generate the reasoning and polish the answer. Cannot reconstruct the argument without the tool.
Same product. Very different learning.
Do not ask students to moralize about the two learners yet. Ask what evidence would let us distinguish them. Their products may be indistinguishable, while their retained knowledge, transfer, confidence calibration, and independence differ.
This is why AI in education is not mainly a plagiarism story. It is a measurement and design story.
[Sources]
The central claim
AI can improve performance while improving, weakening, or leaving learning unchanged.
↑ BETTER LEARNING scaffolding, retrieval, feedback, reflection
→ SAME LEARNING faster production with no deeper change
↓ WEAKER LEARNING answer substitution, shallow practice, dependency
The technology does not determine the outcome. The learning design does.
OECD’s 2026 review makes this distinction directly: outsourcing tasks can improve performance without producing real learning gains. The Bastani field experiment shows that even the interface and guardrails can change the learning result.
The course should carry this distinction throughout the deck. We care about what a learner can remember, explain, transfer, and do when the support changes.
[Sources]
Education is a much bigger design space than a classroom chatbot
Learning happens across a lifetime, in places with very different goals.
elementary classroom university laboratory trade apprenticeship language program special education medical simulation workplace reskilling community college teacher development museum homework table field training
A reading tutor, a flight simulator, and a professional coach should not be judged by the same outcome.
Students often hear “AI in education” and picture a student chatting with an LLM. Widen the frame. Educational AI includes adaptive practice, speech recognition, computer vision, simulations, accessibility tools, scheduling, advising, analytics, and teacher support.
The learning goal may be conceptual understanding, physical skill, teamwork, judgment, language fluency, or safe performance under pressure.
[Sources]
Why the opportunity is so large
The world needs more teaching capacity, not fewer teachers.
44 million additional primary and secondary teachers needed worldwide by 2030
15 million needed in sub-Saharan Africa alone
$120B estimated annual investment needed to address the shortage by 2030
The economic case is strongest when AI expands expert attention, improves teaching quality, or reaches learners who currently go without support.
UNESCO estimates that 44 million additional primary and secondary teachers are needed by 2030, including about 15 million in sub-Saharan Africa. Its teacher campaign estimates roughly 120 billion dollars in annual investment is needed.
These numbers do not prove that AI solves the shortage. They explain why scalable tutoring, teacher coaching, translation, lesson adaptation, and administrative relief are economically important.
[Sources]
This is not an LLM-only field
Different educational problems call for different kinds of AI.
KNOWLEDGE TRACING estimate what a learner knows
RECOMMENDERS choose the next task or resource
SPEECH + NLP listen, read, translate, and give feedback
COMPUTER VISION observe handwriting, lab work, movement, and making
OPTIMIZATION schedule people, courses, rooms, and supports
GENERATIVE MODELS explain, simulate, create, critique, and role-play
Connect back to Week 2. Educational AI predates current chatbots by decades. Classical systems estimate mastery and sequence practice. Speech systems support reading and language. Vision can interpret diagrams, handwriting, physical demonstrations, or laboratory work.
Generative models add flexible dialogue and content creation, but they do not replace the need for student models, curriculum structure, and validated pedagogy.
[Sources]
The technical core
An adaptive learning system is a repeated cycle of inference and action.
1 OBSERVE answers, speech, time, steps, confidence
→ 2 INFER knowledge, misconception, strategy, state
→ 3 CHOOSE problem, hint, example, partner, pause
→ 4 UPDATE use the response as new evidence
Every arrow contains uncertainty. A wrong answer can reflect a misconception, a typo, anxiety, language, or a bad question.
This is the technical spine of the lecture. The system does not observe knowledge directly. It observes behavior and infers a changing latent state. Then it chooses an action intended to move that state.
Use the final sentence to resist overconfident profiling. The same behavior can have many explanations. A good system keeps uncertainty visible and gives the learner and teacher ways to correct the model.
[Sources]
Six pathways to value
Educational AI can do much more than answer questions.
1 PERSONALIZE pace, pathway, practice
2 FEEDBACK specific help at the moment of need
3 EXPAND ACCESS language, format, place, time
4 AMPLIFY TEACHERS planning, noticing, coaching
5 CREATE EXPERIENCES simulations, roles, worlds
6 BUILD CAPABILITY reskilling across a lifetime
This is the opportunity map. The rest of the lecture moves through personalization, teacher capacity, learning design, and institutional systems.
Ask students to name one application in each pathway that does not involve writing an essay with a chatbot.
[Sources]
01
Personalize practice The old dream of a tutor for every learner is becoming technically plausible.
This section explains the architecture and evidence behind personalized tutoring. The emphasis is on careful instructional design rather than chat alone.
[Sources]
The tutor architecture
A useful AI tutor needs more than a language model.
LEARNING GOALS What should become possible?
DOMAIN MODEL What is correct, connected, and important?
LEARNER MODEL What does this person likely understand now?
PEDAGOGICAL POLICY What should happen next?
INTERACTION How will the learner think, respond, and reflect?
The LLM may power the conversation. The surrounding system gives the conversation educational purpose.
Walk from top to bottom. A tutor needs explicit learning goals, dependable domain knowledge, a representation of learner state, a policy for choosing the next move, and an interface that elicits thinking.
A generic chatbot has broad language competence but may lack course alignment, stable sequencing, validated solutions, and a reliable account of what the learner has practiced.
[Sources]
The learner model
Personalization begins with a useful, revisable hypothesis about the learner.
KNOWLEDGE concepts and procedures
MISCONCEPTIONS systematic wrong models
STRATEGY how the learner approaches problems
STATE effort, confusion, confidence, fatigue
CONTEXT language, tools, goals, accessibility
AGENCY what the learner chooses and contests
A learner model should be a hypothesis, not a permanent label.
Personalization is often reduced to difficulty level. A richer model includes what the learner knows, the misconceptions they hold, the strategies they use, and the context in which they are learning.
The learner must be able to challenge or correct the model. A transient pattern should not become a hidden track that limits opportunity.
[Sources]
Knowledge tracing
The system estimates mastery from a sequence of imperfect clues.
Q1 ✓ quick
→ Q2 ✕ confident
→ Q3 ✓ after hint
→ MODEL BELIEF 0.68 probability of mastery
Bayesian approaches update explicit probabilities from prior knowledge, learning, guessing, and slipping.
Deep approaches learn patterns across longer, richer interaction histories.
Knowledge tracing treats mastery as a hidden state. Classic Bayesian Knowledge Tracing updates a probability based on whether the learner knew the skill, learned it, guessed, or slipped. Deep Knowledge Tracing uses recurrent neural networks to learn patterns from interaction sequences.
Do not overteach the math. The conceptual point is that the model estimates, updates, and can be wrong.
[Sources]
Adaptive sequencing
The next activity should sit between boredom and overload.
TOO EASY fluency without growth
PRODUCTIVE CHALLENGE effort with a reachable next step
TOO HARD noise, guessing, disengagement
space a concept mix problem types change representation fade a hint ask for explanation
The tutor’s action space includes far more than choosing an easier or harder question. It can vary representations, space practice, interleave concepts, ask for self-explanation, or gradually remove support.
Productive challenge is personal and dynamic. Friction helps only when it serves a learning mechanism.
[Sources]
Feedback
Good feedback changes the learner’s next move without stealing the move.
ANSWER “The result is 42.” fast, low diagnosis
ERROR LOCATION “Check your second assumption.” preserves search
HINT “What relationship stays constant?” opens a path
QUESTION “Why does that step follow?” elicits reasoning
REFLECTION “What would you try differently next time?” builds strategy
The fastest response is not always the best educational response. A tutor should choose how much information to reveal based on the goal, the learner’s attempt, and the cost of continued struggle.
AI’s advantage is not infinite explanation. It is the ability to tailor the next prompt and wait for the learner to do the thinking.
[Sources]
Evidence · Intelligent tutoring systems
Decades before current chatbots, well-designed tutors were already producing meaningful learning gains.
50 controlled evaluations in a 2016 meta-analysis
0.66 SD median gain over conventional instruction
50th → 75th approximate percentile shift
The boundary: gains were larger on tests aligned to the tutor’s goals, and weak implementations produced small effects.
Kulik and Fletcher synthesized 50 controlled evaluations. The median effect was 0.66 standard deviations, roughly a move from the 50th to the 75th percentile.
The result is impressive but not universal. Effects depended on implementation and on how closely the test matched the tutor’s instructional objectives. A 2025 K-12 review found generally positive results but emphasized small samples, short durations, and weaker differences against strong non-intelligent tutoring.
[Sources]
Evidence · Generative AI tutor
A carefully built AI tutor beat an active-learning lesson in a short university physics trial.
PARTICIPANTS 194 eligible students in a crossover randomized trial
LEARNING 0.73–1.3 SD estimated effect range after accounting for ceiling effects
TIME 49 min median AI-tutor time versus a 60-minute class lesson
EXPERIENCE 4.1 / 5 engagement rating versus 3.6 in class
This was not open-ended ChatGPT. Experts supplied the sequence, solutions, prompts, videos, and pedagogical rules.
The study took place in two lessons in an introductory Harvard physics course. Students crossed between the AI tutor and active-learning classroom conditions. The AI group learned more, spent less median time, and reported higher engagement.
Do not generalize this to all courses. The system was heavily engineered, used expert-written solutions, controlled sequencing, and targeted understanding, application, and analysis in a bounded context.
[Sources]
Access and inclusion
Personalization can also mean changing the doorway into learning.
LANGUAGE translation, conversation practice, local examples
MODALITY speech, text, image, caption, tactile or simplified view
PACE repeat, pause, preview, revisit without social cost
EXPRESSION answer by speaking, drawing, demonstrating, or building
SUPPORT reading guidance, executive-function prompts, structured steps
PLACE reach learners outside specialist or well-resourced settings
Inclusion requires choice. “Personalized” should not mean a permanent easier track or constant surveillance.
AI can transform content across language and modality, give private repetition, recognize multiple forms of response, and support planning. This can be especially meaningful for multilingual learners and learners with disabilities.
Keep expectations high and options visible. A model should not infer that a learner needs permanently simplified work based on a thin behavioral signal.
[Sources]
02
Amplify teachers The best educational AI may help a teacher notice and respond to more learners.
This section shifts from student-facing tools to teacher-facing systems. The goal is to expand teacher capacity and judgment, not to automate away the relationship.
[Sources]
A teacher’s real workflow
Teaching is a continuous cycle of design, attention, response, and revision.
PLAN goals, examples, sequence
→ TEACH explain, model, question
→ NOTICE confusion, progress, participation
→ RESPOND feedback, grouping, support
↺ draft alternatives summarize evidence surface patterns generate practice prepare follow-up
AI products often target lesson-plan generation because it is easy to demonstrate. The deeper opportunity is connecting planning to evidence from learning, classroom activity, feedback, and revision.
The teacher brings goals, knowledge of students, group dynamics, values, and responsibility. AI can expand the number of alternatives and patterns the teacher can consider.
[Sources]
Classroom orchestration
AI can help a teacher see the room at more than one scale.
INDIVIDUAL Who is stuck on what? attempts, hints, confidence, pace
GROUP Which ideas need discussion? shared errors, contrasting strategies
CLASS Where should time move? reteach, regroup, extend, pause
COURSE What should change next time? sequence, examples, assessment
The dashboard should help the teacher ask better questions, not pretend to read students’ minds.
A teacher cannot watch every learner’s process at once. Educational software can aggregate attempts and patterns, then surface questions worth investigating.
Keep the interface epistemically humble. “Six students may be confusing rate and amount” is more useful than a definitive label about ability.
[Sources]
Evidence · Human-AI tutoring
Tutor CoPilot improved student mastery by giving human tutors real-time access to expert teaching moves.
700+ tutors
1,000+ K-12 students in underserved communities
+4 pp topic mastery overall
up to +9 pp for lower-rated tutors
350,000+ messages: access to the tool increased probing questions and reduced generic praise. Tutors kept control over whether and how to use suggestions.
The randomized study involved more than 700 tutors and 1,000 students in a virtual math tutoring program serving federally supported schools. Students whose tutors had access to Tutor CoPilot were four percentage points more likely to pass session exit tickets.
Gains reached nine percentage points for lower-rated tutors. The authors estimate about 20 dollars per tutor per year based on observed use. The important design pattern is AI coaching a person in context, with the person retaining judgment.
[Sources]
Assessment support
AI can help teachers give more feedback, but only the teacher can decide what the work means.
FIRST PASS organize responses and locate common patterns
EVIDENCE LINK connect a comment to a rubric criterion or passage
ALTERNATIVES draft feedback at different levels of specificity
FOLLOW-UP generate practice for the next misconception
Keep human judgment for: ambiguous work · high-stakes grades · context · accommodations · appeals · care
AI can reduce the cost of a first pass through many student responses, suggest rubric-aligned comments, and help create targeted follow-up. This can increase feedback frequency.
High-stakes evaluation needs teacher judgment, especially when work is unconventional, context matters, or the student contests the result. The Department of Education recommends human control, inspectability, explainability, and override.
[Sources]
Create experiences, not just content
Generative AI can turn a lesson into a world that responds.
HISTORY negotiate a treaty with competing interests
HEALTH interview a virtual patient whose symptoms evolve
ENGINEERING diagnose a failing system under time pressure
LANGUAGE practice a conversation with adaptive difficulty
BUSINESS run a market, supply chain, or crisis scenario
SCIENCE design a study and face plausible experimental results
The opportunity is a low-cost practice environment for judgment, not an endless worksheet generator.
Content generation is useful, but responsive simulation is more transformative. Students can practice interviewing, diagnosis, negotiation, design, troubleshooting, and communication with consequences that change.
The simulation must be grounded, bounded, and debriefed. A plausible narrative is not automatically an accurate model of the world.
[Sources]
Advising and student support
Many students do not need another portal. They need help navigating a complicated institution.
QUESTION “Can I graduate next spring?”
→ RETRIEVE program rules, completed courses, deadlines
→ EXPLAIN options, tradeoffs, missing information
→ CONNECT human adviser when judgment or exception is needed
financial-aid navigation course planning career exploration campus resources early outreach
AI can lower the cost of asking basic questions, explain complex policies, compare pathways, and prepare a student for a human advising conversation.
These systems should retrieve from authoritative institutional sources, distinguish rules from suggestions, and escalate exceptions, risk, or distress to qualified people.
[Sources]
The economics of teacher time
The most valuable automation may be the work around teaching, not teaching itself.
GIVE BACK formatting · transcription · routine communication · resource search
REINVEST IN conversation · observation · feedback · mentoring · curriculum judgment
The value equation is not “minutes saved.” It is what those minutes become.
$20 / tutor / year Tutor CoPilot’s usage-based estimate shows why narrow, embedded support may scale economically.
Time-saving claims are common and often weak. Ask what happens after time is saved. If the institution increases workload, reduces staffing, or adds review burden, the educational benefit may disappear.
The Tutor CoPilot estimate illustrates how a narrow system can be inexpensive at the margin. Cost still includes integration, training, data governance, support, and evaluation.
[Sources]
The teacher role expands
Teachers need AI competence, but they should not be turned into full-time technology auditors.
HUMAN-CENTERED MINDSET protect agency and purpose
ETHICS fairness, privacy, access, accountability
FOUNDATIONS understand capabilities and limits
AI PEDAGOGY design learning with the tool
PROFESSIONAL LEARNING evaluate and improve practice
UNESCO organizes 15 teacher competencies across these five dimensions and three levels: Acquire, Deepen, Create.
UNESCO’s framework treats teacher competence as more than prompt-writing. It includes a human-centered mindset, ethics, foundations and applications, pedagogy, and professional learning.
Institutions must provide approved tools, evidence, professional development, support, and clear lines of responsibility. Individual teachers cannot be the only safety layer.
[Sources]
03
Protect the learning When AI can perform the task, education must become clearer about why the learner is doing it.
This is the conceptual center of the lecture. Shift from what AI can do to which cognitive work students still need to do themselves.
[Sources]
Performance is not the same as learning
A smooth performance today can hide fragile capability tomorrow.
Ease can be useful. It can also remove the very effort that changes memory, strategy, and understanding.
Soderstrom and Bjork review the distinction between temporary performance during acquisition and relatively permanent learning. Conditions that make practice look easy can produce weaker retention.
This does not mean difficulty is always good. The question is whether the effort engages a mechanism that builds knowledge or skill.
[Sources]
What counts as learning?
Look for capability that survives a change in time, task, context, or support.
RETENTION Can I retrieve it later?
TRANSFER Can I use it in a new situation?
EXPLANATION Can I show why it works?
ADAPTATION Can I recover when the problem changes?
METACOGNITION Do I know what I know and what I need?
AGENCY Can I choose, question, and continue learning?
This slide offers a practical operational definition. No single test measures every dimension. Different courses will weight them differently.
The AI-rich question is whether the learner can still explain, adapt, and judge when the tool gives an incomplete answer or disappears.
[Sources]
Learning science gives us design tools
The strongest study activities make the learner retrieve, connect, explain, and revisit.
AI can generate and personalize these conditions. It can also bypass every one of them.
Dunlosky and colleagues rated practice testing and distributed practice as high-utility techniques. Butler’s experiments showed repeated testing improved transfer to new inferential questions relative to repeated studying.
An AI tutor can schedule retrieval, vary examples, and ask for self-explanation. An answer engine can let the learner skip them.
[Sources]
Design the amount of help
Assistance should rise when needed and fade as capability grows.
1 ASK What have you tried?
2 FOCUS Where does your reasoning feel uncertain?
3 CUE Recall the relevant relationship.
4 HINT Try representing the problem another way.
5 MODEL Show one step, then return the work.
MORE LEARNER WORK MORE SYSTEM SUPPORT
The ladder is a practical interaction policy. Start by eliciting the learner’s current thinking. Add the smallest support that can move them forward. Return responsibility as soon as possible.
Sometimes direct instruction is appropriate, especially for safety, missing prerequisites, or severe confusion. The design question is deliberate dosage, not a rule that hints are always better.
[Sources]
Evidence · Guardrails change learning
In a field experiment, an unstructured AI tool improved practice scores and then hurt unaided exam performance.
GPT BASE · PRACTICE +48% relative performance while the tool was available
GPT BASE · EXAM −17% relative performance after the tool was removed
GPT TUTOR · PRACTICE +127% with problem-specific hints and solution guardrails
GPT TUTOR · EXAM no significant loss but no significant gain over control either
Nearly 1,000 high-school math students. Students using the open tool often asked for and copied answers.
Students used either textbooks, an open GPT-4 interface, or a guarded GPT tutor during practice. Both AI conditions improved assisted practice performance. On the unassisted exam, the open-interface group scored 17 percent below control. The guarded tutor eliminated the harm but did not create a significant independent-learning gain.
The most important result is the contrast between interfaces. Students’ behavior changed because one system readily provided full answers while the other promoted attempts and hints.
[Sources]
Offloading, scaffolding, and metacognition
The same AI action can support learning at one moment and replace it at another.
AI DOES WHEN IT HELPS WHEN IT HURTS
SUMMARIZES after the learner compares and critiques before the learner encounters the source
EXPLAINS after an attempt reveals a gap instead of asking the learner to explain
PLANS makes strategy options visible chooses every step and priority
CHECKS tests a learner’s judgment becomes the only judge
Metacognition grows when learners predict, monitor, compare, and revise their own understanding.
Do not label a feature as inherently good or bad. Summarization can support comparison after reading, or replace reading before it begins. Checking can calibrate judgment, or make the learner dependent on external validation.
Ask AI to elicit a prediction, request confidence, compare the learner’s reasoning with another approach, and prompt revision. These moves make metacognition part of the interaction.
[Sources]
A better role for conversational AI
The tutor should often ask the question that makes the learner do the next piece of thinking.
ELICIT “What do you think is happening?”
→ PROBE “What evidence supports that?”
→ CHALLENGE “When would that rule fail?”
→ REFLECT “What changed in your model?”
Future direction: Give the learner an AI “student” to teach. The system reveals misunderstandings only when the human explanation is incomplete.
Conversational AI can take multiple pedagogical roles. It can be a Socratic tutor, a debate opponent, a skeptical reviewer, a role-play partner, or a novice that the student must teach.
The teachable-agent idea is powerful because explanation exposes gaps. The system’s goal is not to sound smart. It is to create a reason for the learner to make their model explicit.
[Sources]
Assessment has to change
Assess the process, the explanation, and the transfer, not only the polished artifact.
PROCESS EVIDENCE notes, drafts, choices, prompts, revisions
ORAL DEFENSE explain decisions and answer follow-up questions
TRANSFER TASK apply the idea to an unfamiliar case
LIVE PERFORMANCE solve, design, discuss, or demonstrate in context
REFLECTION name what AI changed and what remains uncertain
AI-INCLUDED TASK evaluate, correct, and improve an AI contribution
AI makes product-only assessment increasingly weak. Process evidence and oral defense reveal ownership of ideas. Transfer tasks test whether capability survives a change in context. AI-included tasks test verification and judgment.
No assessment is perfectly secure or complete. Use several kinds of evidence, aligned to the actual capability the course values.
[Sources]
Integrity in an AI-rich course
A useful policy tells students what kind of help preserves the purpose of the work.
USE brainstorm alternatives · request feedback · test an explanation · translate your own work
DISCLOSE generated text or code · substantial restructuring · AI-created examples or data
DO YOURSELF the target reasoning · unassisted practice · personal reflection · assessment conditions
A simple disclosure: tool · purpose · material contribution · what I verified · what remains mine
A single campus-wide ban or permission statement is rarely enough. The acceptable role of AI depends on the learning goal. Students need a short explanation of why some assistance is allowed and some work must remain unassisted.
Disclosure should be easy enough to do. The goal is not an exhaustive machine log. It is an honest account of material contribution, verification, and responsibility.
[Sources]
AI literacy becomes part of the curriculum
Students need to become capable users, critics, and co-creators of AI systems.
UNDERSTAND how AI represents, predicts, and generates
USE choose tools and collaborate effectively
VERIFY check evidence, uncertainty, bias, and fit
DESIGN shape a system around people and goals
GOVERN question power, data, access, and accountability
UNESCO specifies 12 student competencies across human-centered mindset, ethics, AI techniques and applications, and AI system design.
AI literacy should not collapse into prompt tips. Students need conceptual knowledge, practical collaboration skill, verification habits, design capability, and civic judgment.
UNESCO’s framework positions students as responsible users and co-creators, with progression from understanding to applying to creating.
[Sources]
04
Build the system around learning An educational model sits inside a data, labor, market, and governance system.
This final substantive section moves from classroom interaction to institutional infrastructure, economics, evaluation, and future direction.
[Sources]
Learning evidence infrastructure
The long-term opportunity is a shared memory of growth across tools, courses, and time.
ACTIVITY attempts · discussion · making · practice
→ EVIDENCE what the learner can show
→ MODEL skills · concepts · strategies · confidence
→ ACTION feedback · pathway · support · credential
A lifelong learner model could connect school, work, and self-directed learning. It could also become an extraordinarily sensitive dossier.
Imagine a learner model that persists across a course or career, carrying evidence of concepts, skills, strategies, and goals. It could support continuity, competency-based education, and personalized reskilling.
The same infrastructure could expose intimate patterns about ability, disability, behavior, and aspiration. Treat persistence as a design choice, not a default.
[Sources]
Privacy and equity
Personalization asks for more data precisely where power is already unequal.
COLLECTION What evidence is truly necessary to support learning?
INFERENCE Which sensitive traits or states can the system infer?
ACCESS Who gets the strongest models, devices, support, and connectivity?
ERROR Who is mislabeled, underchallenged, or overdisciplined?
CONTEST Can a student or teacher inspect and correct the record?
EXIT Can the institution leave the platform without losing its learning history?
Education combines children, mandatory institutions, long records, and consequential decisions. That makes data governance unusually important.
Equity is not only equal model accuracy. It includes access to high-quality tools, the risk of lower expectations, language and disability support, recourse after error, and institutional dependence on a vendor.
[Sources]
The political economy of educational AI
Whoever controls the platform can shape curriculum, labor, data, and the definition of success.
The economy of AI is not outside the classroom. It is embedded in the architecture.
A platform’s defaults can influence what teachers teach, what students practice, and what administrators count. Data and integration can create switching costs that give vendors durable power.
Institutions should evaluate interoperability, data portability, teacher control, evidence, total cost, and the ability to exit. A cheaper model can be expensive if it narrows the curriculum or creates permanent dependency.
[Sources]
A practical evidence ladder
Educational AI should earn claims through progressively harder tests.
1 USABILITY Can learners and teachers use it?
2 ENGAGEMENT Do they choose to persist?
3 ASSISTED PERFORMANCE Does work improve while AI is present?
4 RETENTION Does knowledge persist later?
5 TRANSFER Can learners handle a new case?
6 REAL OUTCOMES Do completion, capability, access, or equity improve?
7 LIFECYCLE Do benefits last across cohorts and system changes?
This ladder is a teaching synthesis. Many studies stop at usability, engagement, or assisted performance. Those measures matter, but they cannot establish durable learning.
Retention and transfer require delayed and changed assessments. System-level claims about equity, completion, or cost need broader and longer evaluation.
[Sources]
Human-AI collaboration
The strongest learning system combines machine scale, teacher judgment, and learner agency.
AI continuous practice rapid adaptation many representations pattern detection
TEACHER goals and standards context and relationships classroom orchestration responsibility and care
LEARNER effort and curiosity goals and identity reflection and choice ownership of growth
LEARNING
This is the joint-system claim. AI can provide scale and adaptation. Teachers can interpret context, set standards, orchestrate social learning, and care for people. Learners contribute effort, goals, reflection, and ownership.
Do not make the learner a passive recipient in a teacher-AI partnership. The learner’s agency is one of the outcomes education should protect.
[Sources]
State of the field · August 2026
The news is not one breakthrough. It is a clearer picture of when educational AI helps and when it does not.
2025 Guardrails matter Open assistance lifted practice scores but weakened unaided performance.
2025 Pedagogy can scale A tightly designed tutor produced large short-term gains.
2025 AI can coach people Real-time suggestions improved tutor moves and student mastery.
2026 Policy follows evidence OECD argues for a learning partner, not a shortcut.
2026 · PREPRINT Population signals emerge 3.2 million interactions link faster completion to weaker retention.
NOW The next test is duration Whole courses, transfer, equity, workload, and dependence.
Use this as a news briefing. The 2025 studies show that instructional design, guardrails, and human-AI configuration change outcomes. OECD’s 2026 report synthesizes the emerging pattern.
The 2026 Faster Completion, Less Learning paper is a preprint, not peer-reviewed evidence. It analyzes 3.2 million learning interactions and reports post-ChatGPT declines in time on AI-susceptible problems and weaker proctored retention. Present it as a serious signal that needs replication, not a settled causal verdict.
[Sources]
Near-term horizon
The next wave will make high-quality practice more personal, multimodal, and connected to real work.
PERSONAL TUTOR remembers goals, misconceptions, and progress across years
TEACHER COMMAND CENTER turns class evidence into grouping, questions, and follow-up
MULTIMODAL FEEDBACK watches a proof, lab, performance, drawing, or physical skill unfold
SIMULATION STUDIO creates responsive patients, clients, crises, and worlds
UNIVERSAL ACCESS LAYER translates language, modality, reading level, and interface in real time
APPRENTICESHIP COPILOT guides practice at a bench, bedside, field site, or factory
MASTERY MAP connects evidence from school, work, and independent projects
CAREER TRANSITION COACH builds a personalized path from current skill to a changing role
Keep this slide energetic. Each idea is near enough to prototype now, though not necessarily ready for broad deployment.
The common thread is continuity: models that see more forms of work, remember progress, and connect learning to authentic contexts. The corresponding risks are surveillance, dependency, weak evidence, and unequal access.
[Sources]
Farther horizon
Some of the most exciting futures would change the shape of education itself.
LEARNING WORLDS ON DEMAND learn by acting inside coherent simulated worlds
COLLECTIVE INTELLIGENCE CLASSROOMS groups combine ideas and examine disagreement
TEACHABLE DIGITAL MINDS agents learn from students and reveal gaps
LIFELONG CAPABILITY PASSPORTS portable evidence beyond transcripts and seat time
ROBOTIC LEARNING LABS autonomous instruments run student-designed experiments
NEUROADAPTIVE ENVIRONMENTS consensual adaptation to load and attention
COMMUNITY CURRICULUM MODELS locally built tutors in local languages
PUBLIC AI FOR EDUCATION tutoring and translation as shared infrastructure
Speculative, not inevitable. Each future is also a choice about power, privacy, and what we want people to become.
This is the high-horizon opportunity slide the user requested. Let students enjoy the imaginative range before returning to design criteria.
These possibilities are not predictions. They are prompts for technical and institutional design. Ask which future expands agency, which creates dependency, and what evidence would make each worth pursuing.
[Sources]
A closing design test
When you encounter an educational AI idea, ask: 01 What human capability should become stronger?
02 What cognitive work must the learner still do?
03 What should AI do, and what should the teacher do?
04 How will we test retention, transfer, and agency?
05 Who gains access, power, data, and the right to contest?
BETTERLEARNING not only
Return to the two students from the opening. The purpose of educational AI is not to make every artifact smoother. It is to expand the capability, confidence, access, and agency of learners while strengthening the people and institutions that support them.
The final design test puts learning first, assigns clear roles, demands appropriate evidence, and treats power as part of the system.
[Sources]