Artificial Intelligence and Learning

Performance assistance durable learning truth privacy relationships and human capacity

Main finding

Artificial intelligence can improve the speed, fluency, completeness, and apparent quality of student work while the tool is available. Those gains are not sufficient evidence of learning. The decisive test is what the learner retains, explains, transfers, verifies, and judges after assistance is removed.

The strongest current evidence does not support either indiscriminate adoption or categorical rejection. General-purpose answer generation can become a cognitive crutch. Purpose-built tutoring, guided questioning, source-bounded research, and tools that strengthen a human educator can help under defined conditions. The educational question is therefore not whether AI appears in a classroom, but which mental act is being formed, what the tool does to that act, and how unaided capacity is verified. [1] [2] [3] [4]

Determinations

This table summarizes the major questions about artificial intelligence and learning and the conclusions supported by the current evidence.

Performance is not learning

A student can produce a correct answer without understanding it, an elegant paragraph without owning its reasoning, or executable code without being able to explain its structure. Generative systems intensify this familiar problem because they compress research, drafting, revision, calculation, and presentation into a single interaction. The visible product may improve while the invisible learner changes little.

Three outcomes must therefore be separated. Assisted performance is what a learner can do with the tool. Learning is a durable change in knowledge or skill. Formation is the development of habits by which a person attends, remembers, seeks truth, accepts correction, governs appetite, keeps commitments, and judges how knowledge ought to be used.

The guarded-math study by Bastani and colleagues is a warning about this distinction. Access to a standard GPT interface improved performance during practice but students performed worse when the tool was removed; a tutor design with safeguards largely mitigated the negative learning effect. The finding is causal and important, but it concerns one subject, age range, implementation, and period. It does not justify a universal percentage estimate for all AI use. [2]

The September 2026 release of PISA 2025 adds a broader but observational warning. Students reporting AI use for particular schoolwork tasks scored lower in science on average. That association cannot establish that AI caused lower performance; prior achievement, task selection, school practice, access, and self-selection may contribute. It is a signal for investigation, not a verdict. [5]

The mental acts education must protect

Every assignment contains a target mental act. If AI performs that act, the learner may complete the assignment while missing its purpose. A defensible policy therefore begins with the learning objective and identifies what must remain human before deciding which tools are allowed.

Protected work need not mean technologically isolated education. It means that schools deliberately preserve occasions for direct perception, memory, handwriting where useful, conversation, oral defense, calculation, making, laboratory observation, artistic practice, friendship, service, and embodied work. AI literacy becomes more credible when students know both how to use a system and how to proceed without it.

The removal test

• Can the student state the problem and relevant knowledge without consulting the system?
• Can the student reconstruct or defend the reasoning and identify uncertainty?
• Can the student detect a plausible error or unsupported citation?
• Can the student apply the idea to a materially different case?
• Can the student explain what assistance was used and remain accountable for the result?

Cognitive offloading memory and productive struggle

Cognitive offloading is the use of external action or storage to reduce internal demand. It is not inherently a failure. Notes, diagrams, calculators, libraries, collaborators, and search engines extend human capacity. The formative question is whether an external aid supports a mental process or prevents it from developing. [6]

Earlier research on the so-called Google effect found that people may remember where information can be found instead of the information itself. This is adaptive when access is dependable and the fact is peripheral. It becomes risky when foundational knowledge is needed to recognize error, connect ideas, reason fluently, or decide what to search for. [7]

Retrieval practice supplies a durable counterprinciple: trying to recall knowledge strengthens long-term retention more than repeated study alone. Productive struggle is not the celebration of confusion or needless difficulty. It is carefully calibrated effort before assistance arrives, followed by feedback, correction, and another opportunity to perform. [8]

Preliminary neurocognitive evidence

A 2025 MIT Media Lab preprint reported weaker neural connectivity, recall, and ownership among adults using an LLM for essay writing than among participants writing without tools. It is an interesting multimodal study, not proof that ChatGPT damages the brain. The sample was small, adult, geographically limited, task-specific, and the report had not completed peer review. It should generate replication and better study designs, not headlines that outrun the evidence. [9]

Tutoring is a spectrum not a product category

The word tutor conceals materially different arrangements. A human tutor can notice confusion, motivation, embarrassment, family context, disability, or a misconception that the system was not designed to see. An AI tutor can offer immediate practice and explanations at low marginal cost. Whether the combination helps depends on pedagogy, use, oversight, and what happens when the support is absent.

Stanford SCALE concludes that high-impact tutoring remains live, human-led instruction. Its 2026 review found only twenty high-quality causal student studies among hundreds of AI-and-education papers and no such student studies in United States K through 12 settings. The evidence becomes thinner as the direct human relationship decreases. [1] [10]

Access is not implementation. In two elementary trials, forty to forty-seven percent of students offered an AI platform never used it, and active users averaged only a few minutes per week. Human check-ins increased engagement but did not improve reading achievement. Procurement therefore cannot be credited as educational success without evidence of use, instructional integration, and durable outcomes. [10]

Default institutional comparison

When evaluating an AI tutoring initiative, compare AI alone, AI plus an educator or parent, human support without AI, and usual instruction. Measure both immediate task performance and delayed unaided learning. Without those conditions, a claim that the tool works may identify only novelty, access, or additional adult attention.

Writing authorship and voice

Writing is simultaneously a product, a method of thought, and an act of communication. A policy that treats only the final prose as the object of evaluation misses how a writer discovers a claim, orders evidence, hears rhythm, confronts contradiction, accepts correction, and takes responsibility for words addressed to another person.

The appropriate boundary changes with the objective. If a language learner is practicing sentence formation, generation of sentences displaces the lesson. If an advanced student is studying machine rhetoric, generated text may be the object of analysis. If a disabled student uses an accommodation, the institution should protect access without quietly lowering the intellectual standard or exposing sensitive data.

A practical authorship standard

• The student can identify every consequential AI contribution.
• The student has opened and evaluated the sources supporting material claims.
• The student can explain the structure, evidence, and contested choices.
• The educator has at least one comparable sample of protected human writing.
• Disclosure rules are proportionate, teachable, and consistent across similar assignments.

Research truth and source verification

A fluent system can state falsehood with confidence, merge incompatible sources, invent citations, omit minority interpretations, or summarize away the condition that changes a conclusion. AI literacy must therefore be joined to information literacy, disciplinary knowledge, and a positive duty of truthfulness.

Source-bounded systems reduce but do not eliminate error. They can misread a source, cite a passage that does not support the claim, omit contradictory material, or inherit an incomplete corpus. The student remains responsible for opening the evidence and the institution remains responsible for teaching what counts as authoritative in the discipline.

Truth before fluency

Schools should reward an accurate qualified answer over a confident complete one. Students should be allowed to say that evidence is insufficient. Assignments should include planted errors, conflicting sources, changed policies, ambiguous data, and claims requiring a refusal to infer. Verification is not a final proofreading step; it is part of the intellectual work.

Assessment and academic integrity

The arrival of generative AI does not make assessment impossible. It makes single-product assessment less credible. Institutions need an assessment architecture that combines authentic use, protected performance, process evidence, and human conversation rather than attempting to infer authorship from style alone.

Automated AI detectors are not a sound sole basis for discipline. Published research found bias against writers of English as an additional language, and generated text can be altered to evade detection. California’s model policy similarly recommends alternative assessments, documented authorship, and a meaningful opportunity for rebuttal. [11] [12]

The better response to uncertainty is procedural fairness: clear advance rules, evidence from more than one source, a conversation with the student, attention to disability and language status, proportionate consequences, and an educational path for correction. A school that cannot explain what evidence establishes misconduct should not outsource the judgment to a score.

Minimum protected-performance rule

Every substantial course or grade band should include recurring demonstrations of the capacities the institution claims to develop without generative assistance. The frequency and form should fit age, discipline, disability, and purpose. The result should inform teaching, not merely police students.

A developmental sequence for AI use

Age does not by itself determine readiness. Knowledge, self-command, literacy, disability, family context, and the design of the activity matter. Even so, a developmental sequence is necessary because younger learners have fewer internal standards with which to evaluate a fluent machine and may confuse responsiveness with relationship or authority.

California’s current guidance follows a related progression: recognition and adult mediation in the earliest grades, questioning and simple data concepts later, explicit study of bias in middle school, and increasing creation, citation, privacy, design, policy, and auditing in high school. The guidance is exemplary rather than mandatory, so local adoption and implementation remain matters for verification. [12] [13]

The right to remain humanly capable

Students should not be forced to use a generative system merely because an institution has purchased it. A meaningful alternative is especially important when an activity implicates conscience, privacy, biometric or emotional data, disability information, copyrighted work, or the student’s developing voice. Opting out should not mean losing access to the underlying educational opportunity.

Privacy consent and the student data life cycle

Educational privacy cannot be reduced to whether a vendor signs a contract. Student prompts may disclose identity, location, disability, family conflict, immigration status, health, beliefs, unfinished writing, or misconceptions. Outputs and interaction logs may then influence instruction, grading, discipline, support, or product development. Governance must follow the data from collection through deletion.

The Children’s Online Privacy Protection Act requires covered services to provide notice and obtain verifiable parental consent before collecting personal information from children under thirteen, subject to defined exceptions. Compliance with COPPA is a floor, not a complete educational judgment about developmental appropriateness, school records, state law, or whether a tool should be used. [14]

California’s model policy recommends authorized tools, privacy by design, annual informed guardian consent for autonomous student-interaction tools, consent for AI grading or feedback, contractual limits, deletion, audits, and a parent right to human review. These provisions are a strong governance baseline, but publication of a model does not establish local adoption or faithful practice. [12]

Relationships wellness and artificial companions

Education is relational before it is computational. A teacher’s responsibility, a parent’s knowledge, a peer’s disagreement, and a community’s memory cannot be reproduced by a system that generates responsive language. An AI companion may feel patient and available while lacking mutual obligation, embodied presence, moral agency, confidentiality in the ordinary human sense, and responsibility for the learner’s good.

California guidance emphasizes authentic human connection and cautions that chatbots or emotional-support uses require professional review. The public evidence does not yet establish the long-term effects of sustained companion use on childhood attachment, friendship, loneliness, self-command, or help-seeking. Institutions should therefore apply a precautionary standard and preserve human care as the governing relationship. [13]

Wellness evidence must be designed, not assumed

A wellness claim should distinguish immediate satisfaction from durable well-being. Useful measures include sleep, attention, belonging, loneliness, help-seeking, conflict, physical activity, attendance, persistence, and access to trusted adults. These measures require context: prior condition, family stability, disability, friendships, COVID-era disruption, workload, and other technology use. Short self-reports alone cannot establish benefit or harm.

Equity accessibility and family capacity

AI can lower barriers through translation, speech and text support, flexible explanation, captioning, and rapid feedback. It can also widen inequality when some learners receive disciplined adult mediation while others receive only a chatbot, when paid tools outperform free access, or when low-resourced communities bear greater surveillance and experimentation.

The appropriate comparison is not simply users versus non-users. It is AI plus strong human support, AI with weak support, comparable human support without AI, and ordinary instruction. The effect of the technology may be inseparable from the relationships and implementation surrounding it.

Family authority and capacity

Families need intelligible notice of the purpose, tool, data flows, alternatives, learning boundary, and responsible human decision-maker. Consent is not meaningful when refusal removes the educational opportunity or when technical descriptions conceal foreseeable uses. Schools must also avoid converting parents into unpaid technology administrators as a condition of access.

For home-based and distance-learning settings, the same principle applies with roles reversed: a family may hold greater operational authority, but the program remains responsible for clear curriculum, developmental guidance, privacy, instructor preparation, protected student work, and support when family capacity is limited.

Teacher agency and institutional responsibility

Teachers should not be expected to absorb institution-wide technological risk through individual syllabus statements. They need authority to protect disciplinary methods, time to redesign assessment, access to approved alternatives, and clear responsibility above the classroom for procurement, privacy, accessibility, records, safety, and evaluation.

The OECD’s 2026 Digital Education Outlook reaches a similar distinction: unguided outsourcing can improve performance without producing learning, while purpose-built pedagogical tools and deliberate sequencing are more promising. It recommends developing foundational capacity without AI, learning with educational AI, and then engaging general-purpose systems with critical competence. [15]

No procurement credit without learning evidence

Licenses, workshops, login counts, pilot announcements, and testimonials are implementation indicators. They are not evidence of retained knowledge or strengthened judgment. Institutions should publish the objective, comparison condition, participation, attrition, protected assessment, outcome interval, subgroup results, limitations, and decision taken after review.

A learning-centered institutional standard

The following standard translates the investigation into a review that a school, college, university, program, or family can use without presuming that one policy fits every discipline or developmental stage.

Use categories

This table defines the categories used to distinguish assistance that supports learning from uses that perform or conceal the learner’s work.

An evidence ladder for claims about AI and learning

The quality of an educational claim depends on design, not confidence. Evidence should be rated according to what it can establish and the distance between the study and the setting in which a decision will be made.

Current evidence judgment

• Strong foundational evidence supports retrieval practice and shows that external access changes what people remember. [6] [7] [8]
• Moderate causal evidence supports carefully designed tutoring and guardrails in several bounded settings. [2] [3] [4]
• Current observational evidence supplies population-level warning signals but cannot isolate causation. [5]
• Preliminary neurocognitive evidence is hypothesis-generating and should not be generalized beyond its design. [9]
• Evidence remains insufficient for long-term claims about early childhood, friendship, emotional dependence, equity, moral judgment, and institutional transformation. [1]

No single study settles the educational use of AI. The responsible conclusion is conditional: protect the central mental act, use the least substitutive effective support, retain human authority, and require delayed unaided evidence before claiming learning.

What remains open and how to participate

Public sources can establish published objectives, policies, guidance, approved tools, program descriptions, contracts when released, and reported results. They generally cannot establish what happens in classrooms and homes, how consistently rules are applied, how students experience the tools, which accommodations work, whether parents receive meaningful notice, or what learners can do after assistance is removed.

Recommended contributions

This table identifies what each participant can contribute to close the investigation’s remaining evidence gaps without treating missing information as proof of failure.

Contributions should distinguish direct experience from inference, identify the relevant institution, grade, course, period, and tool, and protect student identity. A testimony can reveal an implementation question but cannot by itself establish prevalence or causation. Conflicting accounts should be retained and investigated rather than averaged into false certainty.

Sources and evidence notes

1. Stanford SCALE The Evidence Base on AI in K 12 2026 review of the quality scope and gaps in student outcome research.

2. Bastani and colleagues Generative AI without guardrails can harm learning Peer reviewed randomized field experiments distinguishing practice performance from later independent performance.

3. Kestin and colleagues AI tutoring outperforms in class active learning Randomized college physics study of one purpose built tutor and lesson context.

4. Wang and colleagues Tutor CoPilot Field evidence on AI assistance delivered through human tutors, including differential effects by tutor rating.

5. OECD Education GPS PISA 2025 September 8 2026 release covering science and broader competencies including reported AI use; associations are not causal estimates.

6. Risko and Gilbert Cognitive offloading Foundational review defining external actions used to reduce cognitive demand.

7. Sparrow Liu and Wegner Google effects on memory Experiments on memory for information and where it can be found.

8. Karpicke and Roediger The critical importance of retrieval Experimental evidence supporting retrieval practice for long term retention.

9. MIT Media Lab Your Brain on ChatGPT 2025 preprint using behavioral linguistic and EEG measures; preliminary and limited in generalizability.

10. Stanford SCALE Access is Not Enough Two elementary randomized trials of AI platform access and human check-ins, published in 2026.

Sources and evidence notes continued

11. Liang and colleagues GPT detectors are biased against non native English writers Peer reviewed evidence relevant to fairness and the limits of detector-based discipline.

12. California Department of Education Model artificial intelligence policy Nonmandatory model covering integrity human review privacy consent vendors development and monitoring; reviewed June 26 2026.

13. California Department of Education Learning with AI Learning about AI Updated state guidance on human-centered developmental and privacy-conscious use; exemplary rather than mandatory.

14. Federal Trade Commission Children Online Privacy Protection Rule Federal requirements for covered online services collecting personal information from children under thirteen; amended in 2025.

15. OECD Digital Education Outlook 2026 International synthesis distinguishing unguided performance gains from learning and emphasizing pedagogical purpose.

16. UNESCO Guidance for generative AI in education and research Human-agent age-appropriate governance framework for educational use.

17. United States Department of Education Student Privacy Policy Office resources Official resources concerning student records privacy and educational technology.

18. Institute of Education Sciences What Works Clearinghouse handbooks Standards for judging causal educational evidence and study quality.

Interpretive limits

Publication dates, policy language, study samples, comparison conditions, outcomes, and review status should travel with every claim. Product and institutional materials establish what is offered or promised. Participant testimony establishes experience. Only suitable comparative designs can estimate causal effects, and even strong causal studies may not transfer across ages, subjects, populations, tools, or implementations.

Artificial Intelligence Tool Suitability

Practical guidance for schools universities educators parents and students

Main finding

The strongest educational fit does not belong to one brand. It belongs to a mode of use. AI is most compatible with a well-formed mind when it asks the learner to reason, stays bounded to inspectable sources, makes assistance visible to a teacher or parent, and is followed by independent performance. AI is least compatible when it supplies finished answers, substitutes for relationships, or removes the productive effort through which knowledge and judgment develop.

Recommended order of preference

• Human teaching and independent effort remain the foundation.
• Source-bounded tools help students examine defined materials and trace claims.
• Guided tutors ask questions, reveal hints gradually, and keep the learner responsible for the solution.
• General assistants support critique, comparison, and revision only after the learner has attempted the work.
• Answer generators, AI companions, and automated surveillance do not qualify as ordinary educational substitutes.

What the matrix evaluates

This table identifies the questions used to judge whether an artificial-intelligence tool or mode protects the development of a well-formed mind.

This is a suitability judgment, not a permanent product ranking. Features, prices, policies, and models change. Schools should reassess products before procurement and should test learning outcomes after deployment.

1 Evidence that governs the recommendations

Current evidence supports a conditional conclusion. Generative AI can improve practice performance and access to feedback, but the same tool can weaken unaided performance when students use it as a shortcut. Design and supervision determine whether AI functions as a tutor or a crutch.

Guardrails change the result

Strong evidence: A large field experiment in high-school mathematics found that unrestricted GPT-4 access improved practice performance but harmed later unaided performance. A guarded tutor version designed to withhold direct answers avoided the same damage. The published title states the central result: generative AI without guardrails can harm learning.

Implication: A school should not treat access to a capable model as an educational intervention. It must specify how the model behaves, what the learner must do, and how independent mastery will be tested.

Human support remains decisive

Experimental evidence: Tutor CoPilot supplied real-time AI guidance to human tutors in a randomized trial involving 900 tutors and 1,800 K-12 students. Students working with supported tutors were 4 percentage points more likely to master topics; the gain reached 9 percentage points for students assigned to lower-rated tutors.

Recent evidence: Stanford SCALE reports that access alone is insufficient and that human support improves engagement with AI tutoring. Its 2026 review warns that adoption is moving faster than causal evidence across contexts.

Implication: The best current case is AI strengthening the teacher, tutor, or parent. Evidence is weaker for replacing that person.

How students use the tool matters

A 2026 randomized experiment found gains when students used AI to explain concepts, while short-run advantages from automating the writing task did not persist in the same way. Another study of mathematical proof found better homework performance but no significant exam gain. These results support delayed, unaided assessment and process-level evidence rather than judging the visible product alone.

Evidence limits

• Many vendor studies measure adoption, satisfaction, time saved, or conversation features instead of delayed learning.
• A product may have strong privacy controls and still be educationally unsuitable for a particular task.
• A feature described as Socratic should be tested against actual conversations and unaided outcomes.

2 Suitability standard

This table defines the dimensions of educational suitability and contrasts a strong formative fit with warning signs that a tool may displace learning.

Suitability labels

This table defines the labels used in the tool matrix so each rating can be interpreted consistently and within its stated limits.

3 Tool matrix at a glance

This table compares tools and modes by their best educational role, overall fit and the conditions required for responsible use.

What rises to the top

Khanmigo and source-bounded notebook systems are the clearest current matches for routine learning because their visible design constrains the model’s role. ChatGPT Study Mode, Claude Learning Mode, and Gemini’s guided features can support serious learning, but they remain embedded in broad assistants that can also generate finished work. Teacher platforms are strongest when they return time and insight to a human educator rather than automate the educational relationship.

4 Guided tutors

Khanmigo

Why it fits: Khan Academy describes Khanmigo as guiding learners without giving direct answers. It supports personalized tutoring, writing and debate prompts, coding feedback, teacher summaries, and parent access to interaction history and moderation alerts. Access for minors requires a parent or participating school or district.

Best use: Mathematics and coding hints, guided writing conferences, debate practice, and formative feedback after the student has attempted the work.

Required safeguards: Limit sessions; preserve notebook or scratch work; ask the learner to restate the reasoning; require periodic closed-tool performance; review flagged conversations; compare gains against ordinary Khan Academy practice or human tutoring.

Judgment: Good fit with safeguards from later elementary through college. For younger students, adult presence and short sessions matter more than product capability.

ChatGPT Study Mode

Why it fits: Study-oriented use can ask questions, break concepts into steps, adapt explanations, and encourage a learner to work through a problem. Institutional ChatGPT Edu accounts add administrative and privacy controls that consumer access may not provide.

Best use: High-school and college explanation, practice questions, oral rehearsal, counterarguments, error diagnosis, and critique of the student’s own attempt.

Required safeguards: Use a course instruction that withholds finished answers; keep sources visible; disclose assistance; follow with an oral, handwritten, live, or supervised task. Do not allow switching to unrestricted generation during protected practice.

Judgment: Conditional to good for mature learners. A study interface does not by itself prove learning; the assignment and verification system remain decisive.

Claude Learning Mode

Why it fits: Anthropic states that Learning Mode guides reasoning, uses Socratic questions, emphasizes core concepts, and offers templates within course projects. The product is positioned for higher education.

Best use: Conceptual explanation, thesis testing, evidence questions, literature-review planning, and feedback before final submission.

Required safeguards: Students must read cited work, maintain their own notes, produce the substantive argument, and defend it without the model.

Judgment: Conditional to good for college use. Public outcome evidence remains less developed than the feature description.

5 Source bounded research and study tools

Gemini Notebook and NotebookLM

Why it fits: Google describes its education notebook as a walled garden grounded in sources supplied by the user. It can produce explanations, questions, study materials, and audio overviews while linking responses to the source collection. School accounts receive administrative and data protections under Google Workspace for Education terms.

Best use: Close reading across assigned documents, comparison of primary sources, retrieval from a course packet, question generation, and preparation for a human seminar or oral examination.

Required safeguards: The teacher curates the source set; students distinguish quotation, paraphrase, and model inference; every important claim is opened in context; audio summaries supplement rather than replace reading; the final assessment includes unaided explanation.

Judgment: Good fit with safeguards from middle school onward. It is among the strongest current tools for truth and provenance because the evidence boundary is visible, but a weak or one-sided source set still produces weak learning.

Elicit and academic research assistants

Why they fit: Elicit searches academic literature, extracts information, and attaches sentence-level citations. Similar academic search systems can help students discover vocabulary, competing findings, methods, and gaps.

Best use: College and graduate literature discovery, evidence tables, study screening, and comparison of methods or findings.

Required safeguards: Treat the tool as an index and extraction assistant. Open each paper, inspect methods and limitations, record inclusion decisions, search outside the platform, and cite the original source.

Judgment: Conditional for advanced learners. These tools can strengthen evidence discipline, but automated synthesis can hide search omissions, poor study quality, or disagreement.

General citation and answer engines

Citation links improve inspectability, but a citation is not proof that the source supports the sentence. These systems are appropriate for discovery only when the learner opens the cited material and can explain why it is credible, relevant, and sufficient.

6 Teacher controlled platforms

MagicSchool and MagicStudent

Why it may fit: MagicSchool provides teacher planning, differentiation, rubrics, feedback, and teacher-led student tools. Its public materials describe administrative controls, content moderation, role-based access, zero data retention with model providers, and third-party privacy certifications.

Best use: Reducing clerical preparation, generating options that a teacher edits, creating bounded practice, and giving students AI-literacy exercises under teacher supervision.

Risk: Large menus of generators can encourage quantity over curricular coherence. AI-generated individualized education program drafts, grading, or comments require special caution because errors can affect consequential decisions and relationships.

Judgment: Conditional. Strong privacy claims and school controls improve institutional suitability, but educators must review content and retain authorship of consequential judgments.

SchoolAI Spaces

Why it may fit: Teachers design Spaces, control how the AI responds, and view student activity through Mission Control. Public materials describe configuration, crisis flags, non-human disclosure, privacy commitments, and emerging evidence about higher-order student interactions.

Best use: Teacher-designed simulations, guided practice, targeted questioning, and real-time identification of students who need human help.

Risk: Dashboard inferences can appear more precise than they are. A conversation pattern is not automatically evidence of comprehension, mental state, or mastery. Monitoring can also change student behavior and trust.

Judgment: Conditional. Promising for keeping the teacher in control, provided that human follow-up and independent assessment validate what the dashboard suggests.

Brisk Teaching

Why it may fit: Brisk works inside materials educators already use and can create lessons, differentiate texts, generate feedback, inspect revisions, and build student activities from selected resources.

Best use: Teacher preparation, translation, accessibility, question generation, and first-pass feedback that the teacher reviews.

Risk: Automatic feedback and writing inspection may flatten voice, misread intention, or create unwarranted suspicion. Teachers must not outsource evaluation of originality or character.

Judgment: Conditional as a teacher assistant. Student learning depends on the quality of the source, prompt, review, and follow-up task.

7 Writing coding and creative tools

Writing assistants and Grammarly

Appropriate role: Spelling, punctuation, accessibility, formatting, and targeted revision after a student has written a complete first draft. A student may compare one suggested sentence with the original and explain the choice.

Boundary: Generating the thesis, structure, paragraphs, examples, or voice replaces the very capability most writing assignments are meant to reveal. Disclosure does not restore lost practice.

Judgment: Limited conditional use. The more a feature changes meaning or organization, the later it should enter the process and the stronger the authorship record should be.

GitHub Copilot and coding assistants

Appropriate role: Explaining an error, generating test cases, comparing implementations, refactoring code a student understands, and working with professional tools after foundations are established.

Boundary: Beginners can produce functioning code they cannot trace, test, secure, or modify. Schools should protect syntax, algorithms, debugging, and design practice before allowing routine completion or agentic implementation.

Verification: Require students to predict output, explain each dependency, write tests, repair a seeded defect, modify the program live, and complete periodic unaided exercises.

Judgment: Conditional for advanced high school and college. Poor fit when the assignment measures basic construction or debugging fluency.

Image audio and video generators

Appropriate role: Exploring visual alternatives, storyboarding, testing prompts as a form of media literacy, accessibility, or producing components when manual craft is not the objective.

Boundary: Do not substitute generation for observation, drawing, performance, composition, or craft practice when those are the learning goals. Protect consent, likeness, copyright, provenance, and cultural context.

Judgment: Conditional. Require a concept created by the student, a process record, source disclosure, and a critique of what the model introduced or distorted.

8 Uses that conflict with formation

Finished answer generation

When the assignment exists to develop reading, memory, calculation, writing, or reasoning, a finished AI answer replaces the evidence of learning. It should be prohibited during foundational practice and protected assessment. Allowing disclosure does not solve the educational loss.

AI companions and substitute relationships

A conversational system may feel patient and attentive, but it does not assume the obligations of a parent, teacher, friend, counselor, pastor, or clinician. Schools should not market ordinary chatbots as relationships or use them to replace belonging, spiritual direction, counseling, or mentorship. A support tool must identify itself as a machine and route risk or distress to a qualified human.

Emotion recognition and behavioral inference

Systems that infer attention, emotion, honesty, risk, or character from faces, voice, writing, or interaction logs can create false precision and intrusive records. Routine use is a poor fit. A narrow exception would require a demonstrated educational or safety need, validated performance for the population, minimal collection, short retention, human review, an appeal process, and a non-punitive alternative.

Automated grading of consequential work

AI can organize evidence or suggest feedback, but it should not make final decisions about grades, discipline, disability services, admission, or student character. The responsible adult must inspect the student’s actual work, understand the model’s basis, and remain accountable for the judgment.

Unrestricted use by younger children

Elementary education should prioritize language with adults, reading, memory, handwriting, number sense, play, nature, making, and direct experience. Teacher-facing AI may reduce preparation time, but unsupervised generative chat should not become a child’s default path to explanation or imagination.

Mental fitness test

• Does the tool increase sustained attention or reward rapid switching?
• Does it lead back to books, people, physical activity, and real work?
• Can the student stop using it without anxiety or loss of confidence?
• Does it strengthen help-seeking from people or allow avoidance of them?
• Is the learner becoming more capable without the tool over time?

9 Recommendations by age

This table adjusts artificial-intelligence use by developmental stage, identifying suitable assistance and the human capabilities that must remain protected.

Homeschool and family application

Families can apply the same progression used in a classical curriculum. Students first acquire material, then analyze it, then express and defend a judgment. AI enters only after the relevant human faculty has received real practice. Parents should see the interaction history, state the permitted role, and ask the child to explain the work away from the screen.

Post COVID consideration

Students who experienced disrupted instruction, family instability, isolation, weak friendships, or declining wellness may need more human structure before more digital independence. AI can offer low-friction help, but it can also make avoidance easier. Schools should distinguish remediation from dependence and pair tool access with relationships, routines, physical presence, and measurable recovery.

10 Assignment rules and verification

Every AI-permitted assignment should contain a short tool contract. The contract makes the educational purpose visible and gives students a stable standard across products.

Verification methods

• Oral defense that probes the student’s reasoning and evidence.
• Supervised writing or problem-solving on a related but new task.
• Process portfolio showing notes, drafts, source checks, and revisions.
• Live modification of code, model, design, experiment, or performance.
• Delayed quiz or explanation after AI access ends.
• Comparison between AI output, primary evidence, and the student’s final judgment.

Disclosure standard

Disclosure should identify the tool, the purpose, the material supplied, and the substantive effect on the work. Prompt transcripts may be useful, but volume alone is not evidence of honesty or learning. The student must be able to explain what changed and why.

11 Procurement and pilot requirements

This table identifies the evidence an institution should require before adopting, expanding or renewing an artificial-intelligence system.

Pilot design

• Begin with one capability, one age band, and a short period.
• Compare AI plus human support, AI alone, and a credible non-AI practice where feasible.
• Measure work produced with AI and performance after access is removed.
• Review conversation samples for direct answers, fabricated sources, bias, unsafe advice, and unproductive dependence.
• Include students, teachers, parents, counselors, privacy staff, and learning-support staff in the review.
• Stop or narrow the pilot if independent performance, trust, attention, or relationships deteriorate.

Decision rule

Adopt only when the tool improves a defined educational result without weakening independent competence, privacy, relationships, or equitable access. Time saved for adults counts only when the recovered time reaches students through better teaching, feedback, or care.

12 Recommended policy position

Schools and universities should reject both blanket prohibition and blanket adoption. The institution should authorize specific tool roles within a developmental sequence and preserve a meaningful non-AI path when a product is not essential.

Institution wide minimum

• Define protected human capabilities by age and discipline.
• Require school-managed accounts for routine student use whenever available.
• Give instructors assignment-level control within a common institutional floor.
• Require an independent attempt before AI on formative work unless accessibility needs justify another sequence.
• Use source-bounded or guided modes before unrestricted generation.
• Prohibit confidential and protected data in unapproved systems.
• Retain human responsibility for grading, discipline, counseling, admission, and disability decisions.
• Require recurring unaided performance and delayed learning measures.
• Give families plain-language notice, meaningful access to policies, and an appeal or alternative process.
• Review vendors and educational evidence at least annually and after major feature or policy changes.

Final judgment

AI can help form a capable mind when it is designed and governed as a temporary scaffold, an evidence tool, or an assistant to a responsible human teacher. It becomes incompatible with formation when the learner’s own memory, reasoning, voice, judgment, or relationships become optional. The test is simple to state and demanding to enforce: after using the tool, is the student more able to understand, explain, verify, choose, and act without it?

13 Selected sources

Product pages document current features and policies. They do not prove educational outcomes. Research sources are listed separately so that product claims and measured evidence remain distinct.

Learning evidence

Bastani and colleagues Generative AI without guardrails can harm learning - large field experiment in high-school mathematics.

Wang and colleagues Tutor CoPilot - randomized human-AI tutoring trial involving 900 tutors and 1,800 students.

Stanford SCALE Evidence base on AI in K-12 - 2026 research review and human-support findings.

Contractor and Reyes Experimental evidence on the learning impact of generative AI - randomized study distinguishing augmentation from automation.

Chen and colleagues AI support for learning mathematical proof - homework gains, no significant exam gain, and overreliance concerns.

Guided learning and source bounded tools

Khanmigo - guided tutoring, teacher tools, parent controls, and access rules.

Google Gemini for Education - under-18 experience, data protections, guided learning, and verification features.

Google Gemini Notebook - source-bounded education notebook and Workspace protections.

Anthropic Claude for Education - Learning Mode and higher-education deployment.

OpenAI ChatGPT Edu - institutional deployment and workspace controls.

Teacher platforms and advanced tools

MagicSchool trust and privacy - school controls, privacy claims, and certifications.

SchoolAI - teacher-designed Spaces, Mission Control, safety, and emerging evidence claims.

Brisk Teaching - teacher workflow, feedback, student activities, and privacy information.

Elicit - academic search, extraction, literature review, and sentence-level citations.

GitHub Education - Copilot access and educational use.

Scope limit

This case file evaluates publicly documented tool modes and selected research available in September 2026. Institutions must verify current contracts, product settings, age eligibility, accessibility, state requirements, and data practices before adoption. Product behavior should be tested directly because a vendor may change its models, defaults, integrations, or terms.