top of page
Glossy 3D 48PEGASUS Logo.png

Job Displacement and Broader Societal Impacts

Education and entry pathways

Schoolwork adoption is not an achievement result

Student survey · United States ages 13 to 17 · Fielded in fall 2025

What teenagers reported

Pew Research Center surveyed 1,458 U.S. teens and their parents online from September 25 to October 9, 2025. The findings represent teens ages 13 to 17 living with parents, not all young adults or the full K–12 population. Fifty-four percent reported having used AI chatbots to help with schoolwork. About 10% said chatbots helped with all or most of their work, 21% some and 23% a little. Forty-five percent did not use chatbots in this way; rounding and small nonresponse prevent an exact total. [64]

Comparing surveys

Pew's earlier [*REDACTED*]-only question found 13% in 2023 and 26% in 2024. These two comparable question results show a rise in reported use of that tool. Do not put the later 54% value on the same trend line: it measures any AI chatbot. Differences in wording and denominators can mimic gains in adoption. [65] [64]

Separate opportunity, use and mastery

A student who asks for hints, a student who checks references, and a student who submits generated answers may all answer yes to schoolwork use. The adoption question does not estimate usage hours, independent achievement, plagiarism or the effect of a school policy. Likewise, 59% describing classmates' AI cheating as at least somewhat common is a perception measure, not an observed infraction rate. [64]

What a district can measure

Use course-specific instruments to separate use for feedback or accessibility from substitution for assessed work. Combine an anonymous student survey, a small teacher observation sample, and a delayed independent task. State which students were surveyed, whether users self-selected, and what happens to students not using the selected platform. Avoid collection of unnecessary personal prompts. Evaluation should tell educators what happened to thinking, not merely prove that licenses were activated.

Exhibit 10 · More than half of teens used AI for schoolwork help

Pew fieldwork September 25–October 9, 2025; published February 24, 2026. The first three categories sum to 54%. Responses are rounded and not all responses are shown. Do not splice with older [*REDACTED*]-only questions.

Pew Research Center, How Teens Use and View AI: primary source

Teachers want time back, but time is not a headcount estimate

Teacher-reported adoption · Independent workload study · Preserve instructional capacity

Recent U.S. evidence

The [*REDACTED*] and [*REDACTED*] surveyed 1,952 U.S. K–12 educators online July 28–August 14, 2026, including teachers, paraprofessionals and others working in classrooms. Fifty-nine percent said they used AI tools for work activities during the 2025–26 school year. Among users, 13% said they got a lot more time for valued work and 51% somewhat more; 29% reported no effect and 7% less time. These are perceived effects among users, not randomized changes in hours. The question included adaptive learning and interactive platforms as well as chatbots. [66]

Reported gains are not automatically redistributable

A few minutes saved in scattered tasks cannot simply be reassigned into budgeted teaching posts. Time may move to correcting outputs, mentoring students, adapting instruction, or meeting previously unaddressed needs. A cost model should account for setup, review, policy and training time as well as time returned.

A test of time savings

England's NFER evaluation for the Education Endowment Foundation used a randomized comparison for an AI lesson-planning assistant, Aila. Over ten weeks in autumn 2025, Year 3–6 teachers assigned access spent about 2.5 hours weekly on lesson and resource preparation versus 3.3 hours in comparison, saving about 49 minutes. Independent review did not find worse materials. Both groups could use other generative tools. The measured outcome was preparation time and assessed material quality, not pupil achievement, budget savings or positions made unnecessary. [67]

Put saved time back in the classroom

For a local pilot, track preparation, verification, revision and actual time reallocated, plus accessibility and error quality. Ask whether feedback to students improved, whether individualized support increased, and whether cuts would erase the time that review requires. Treat recaptured teacher time as a conditional dividend, not a forecast of staff reductions.

AI-assisted practice and independent learning can diverge

Randomized math trial · Türkiye · Relative scores versus control

Why the outcome changes the story

In a classroom-randomized high-school mathematics study in Türkiye, [*REDACTED*] and colleagues compared a control group, a [*REDACTED*] based assistant configured like a general chatbot, and a tutor configured with teacher-supplied solutions and progressive hints. During assisted practice, the base and tutor groups scored 48% and 127% higher, respectively, relative to the control group. These are relative changes from the control mean, not percentage-point changes or national grade gains. [68]

The tool was then removed

On an exam without the AI, the [*REDACTED*] Base group scored 17% lower relative to the control group. The negative effect was largely mitigated in the guarded-tutor arm, whose exam difference was not statistically distinguishable from control. That does not prove the exact effect is zero. It also does not show the guarded tutor produced a positive independent learning gain in this setting. [68]

Why attempts and explanations matter

Students can finish more questions while practicing fewer of the decisions the exam requires. This mechanism is plausible and supported by observed interaction patterns in this trial; it should not be inferred for every learner from product telemetry alone. Teacher-designed scaffolding can move effort back onto the student: insist on an initial attempt, request a reason, offer the smallest useful hint and ask the student to solve a changed problem. [68] [69]

How to audit a pilot

Define the knowledge or method that should transfer before picking the tool. Collect an assisted practice measure, an immediate unaided measure and a delayed unaided transfer measure. Specify whether accessibility supports remain available; removing the AI is different from removing established accommodations. If assisted scores rise and independent scores fall, stop scaling until the activity is redesigned. A good-looking exercise is not enough to justify a rollout.

Exhibit 11 · Assisted practice was not the same as learning

One randomized experiment tested two designs in a Turkish high-school mathematics setting. Reported effects are relative to control scores, not percentage-point changes. [*REDACTED*] Tutor’s unassisted exam difference was not statistically distinguishable from control; it is not plotted as precisely zero.

[*REDACTED*] et al., Proceedings of the National Academy of Sciences: primary source

Tutoring evidence depends on who or what receives assistance

Three experiments · Different populations · Do not pool outcomes

Human tutors can remain the pilot's center

Tutor CoPilot supplied suggestions to human tutors in K–12 online math tutoring rather than placing an unrestricted chatbot directly in front of students. Research by [*REDACTED*] and colleagues reports a four-percentage-point increase in exit-ticket mastery, from 62% to 66%; the lower-rated tutor group improved nine points, from 56% to 65%. These are site-and-program-specific short-run pass rates and the subgroup should not be taken as a promise for every inexperienced tutor. [70]

A teacher-facilitated program

A World Bank working paper on a six-week after-school program in Edo State, Nigeria, reported a 0.31-standard-deviation treatment gain on a combined assessment of English, digital skills and AI knowledge. The intervention bundled generative AI with structured sessions and supervision. The randomized comparison supports the package's local effect, not a clean estimate of the independent contribution of the software. [71]

A college design is a different test

[*REDACTED*] and colleagues studied a course-specific tutor against active-learning instruction in an undergraduate physics course. Their randomized study reported higher learning gains in less time for the tutor condition on the measured lessons. The finding informs course-specific tutoring design, not the broader effect of replacing all classroom time or the durability of gains across semesters. [72]

Replication before scaling

Show each case with its age group, subject, teacher role, comparator, test timing, outcome and product configuration. Do not rank an exit-ticket percentage-point effect against standard deviations on a composite or a physics intervention test. Where possible, test the least expensive adequate alternative: teacher feedback, established tutoring, peer support or non-generative formative software. Only scale if effectiveness, accessibility, supervision and full operating costs are credible locally.

Higher education should teach with tools and test beyond them

Proctored experiment · July 2026 working paper · Not a universal college effect

What a recent experiment found

[*REDACTED*] and [*REDACTED*] randomly assigned undergraduate students to access off-the-shelf generative AI or not while studying an unfamiliar topic and writing in proctored sessions. On the immediately following unaided knowledge test, access increased scores by 0.27 standard deviations; gains persisted one week later. The authors also report that delayed essay gains occurred especially for students who used AI for explanation rather than generating text. The experiment identifies the effect of access under its design; students' selection into usage modes is not itself a randomized intervention. [73]

What it adds to the evidence

The result is not inconsistent with a negative unassisted exam effect elsewhere. Subjects, baseline skills, the intervention, and the outcome can differ. A course can benefit from directing time to reading and asking conceptual questions while losing practice if students outsource the assessed skill. This July 2026 study is a working paper, not conclusive evidence about all higher education effects. [73] [68]

Rebuild syllabi around intelligible permissions

Mark each assessment's AI permission and rationale. Give students a clear explanation of when sources must be independently checked, whether tool-assisted editing is allowed, when disclosure is required, and what does or does not count as an accommodation. Include a non-tool path when equivalent work is required and universal access is missing.

Give both types of credit

An unaided explanation shows subject mastery. An assisted work product can show selection, supervision and revision. An assignment can test both without equating a product demonstration to a skill test. In a budgeting module, for example, students first compute key estimates from a source table, then ask an assistant for an alternative scenario, verify the output and defend which assumptions changed the recommendation.

Assessment integrity needs evidence, not a detector verdict

Student rights and accountability · Multiple sources of evidence · Maintain review channels

The reliability problem

A 2023 study evaluating particular [*REDACTED*] detectors and writing samples reported elevated false-positive rates for non-native English writing. That finding does not establish that all detectors, versions or datasets have the same error rates. It does show why institutions must verify the detector's applicability before treating its scores as meaningful. The true case mix matters: a low false-positive rate on a vendor benchmark can still yield many false accusations when prohibited use is rare. [74]

A worked arithmetic example

Suppose 1,000 assignments include 20 actual policy violations; an illustrative detector catches 90% of those and falsely flags 2% of the 980 others. Approximately 18 true flags and 20 false flags result. Fewer than half the flags are correct even though the headline sensitivity sounds high. This arithmetic is illustrative only; none of the assumed rates is a claim about a real detector.

Move evidentiary weight onto learning

Before a problem occurs, design checkpoints: notes, source discussion, explanation of decisions, revision evidence and a transfer task. At review, distinguish authorship uncertainty from policy violation. Evaluate any school-approved accessibility tool in context. Do not use detector output alone to discipline students. Provide notice of the issue, human review, an opportunity to explain and an appeal process appropriate to the institution.

Data discipline applies to assessment too

Do not upload identifiable drafts to an unapproved detector to investigate suspected misuse. Limit access to records and use published retention rules. Review disparities by assignment type and grade or language group where sample size and privacy rules permit. Report overturned flags as well as upheld decisions; tracking only cases that survive a filter conceals the cost of false positives. [75] [74]

Teach AI literacy as an extension of foundational skills

Transferable knowledge · Verify against sources · Learn when not to delegate

Frameworks identify more than prompting

UNESCO's student framework organizes 12 competencies around a human-centered mindset, ethics, techniques and applications, and system design, across understanding, application and creation. Its teacher framework adds professional competence in pedagogy and learning. OECD and the European Commission published a primary and secondary AI literacy framework in 2026. These are curricular resources, not mandated universal learning standards. [76] [77] [78]

Make the skill visible

Ask students to compare answers against independently checked sources, identify unsupported certainty, explain how input quality affects retrieval or generation, and explain when to involve a person. Require factual, mathematical and domain reasoning before tool use. For younger learners, discuss classifiers and uncertainty with examples or paper exercises instead of requiring any individual AI account.

Fundamentals are not optional

The OECD adult skills survey distinguishes literacy, numeracy and adaptive problem solving, each with its own scale. Those domains are not measurements of AI literacy, but they clarify the foundational skills learners need before asking a model to explain, judge or adapt. A fluent prompt cannot substitute for recognizing whether an answer's units or cited source make sense. [79]

Curriculum example

In a school-to-career materials-planning exercise, provide a short supplier table and a client constraint. Have students calculate the order unaided. Next, ask them to compare their results with an approved tool's proposal, verify calculations, identify a misleading assumption and produce a short recommendation. The final assessment should score both baseline competence and the ability to use assistance responsibly. Tools change quickly; verification, numeracy and a well-framed problem remain portable.

Build career pathways around supervised practice

Entry opportunities · Paid learning · Evidence employers can recognize

Protect what junior work used to teach

Routine first-year tasks can be inefficient and still teach useful judgment when paired with feedback. If a tool performs drafts, searches or simple calculations, employers should preserve deliberate opportunities to build basic competence and observe senior decisions. Training plans can use realistic cases whose correct answers are initially withheld, short observation rotations, and scheduled review of errors. Do not infer from classroom experiments that a particular occupation will disappear. [69]

A tested institutional structure

U.S. Registered Apprenticeship combines paid work, structured learning and mentorship, related instruction, and a recognized credential. Similar design principles can be adapted to employer-led training, though a program should not call itself registered without meeting the formal requirements. Put paid release time, instruction cost, and supervision responsibilities into the operating budget. [80]

Make the credentials usable

A short course completion measures participation, not competence or placement. Give employers a work sample with clear evaluation criteria and separable evidence of unaided competence, assisted execution, and review. Include task context and version where relevant. Assess independently rather than relying on activity logs or a tool-certified skill claim. Apprentices should be able to take the evidence to another employer.

Measure who gets through the path

Track applicants, interviews, paid placements, progression and retention for eligible groups with adequate privacy. Distinguish participants who change jobs by choice from involuntary displacement. Compare against earlier cohorts or matched alternatives and report uncertainty. Avoid using course enrollment as a proxy for workforce resilience. Locate mentorship where students actually need it, including adult career transitions and community colleges.

Digital equity is about the conditions of productive access

Device and broadband · Access to support · No compulsory paid account

Access has layers

A household with mobile internet is not necessarily able to run a laptop workflow, read dense tables, review long drafts or join a live session. NCES reports home connectivity for U.S. children ages 3 to 18 separately for computer access and smartphone-only access by selected characteristics. Those data describe internet-device access, not ownership of an AI subscription or measured exposure by classroom. [81]

Different bottlenecks dampen adoption

Device reliability, fast connections, language support, accessible features, protected time and knowledgeable help are different requirements. A tool may be broadly available but not function equally under weak service or on devices students share. Paid higher tiers or available parent coaching can change what using the same service means. A classroom survey that asks only whether students have ever used AI cannot identify these bottlenecks. [69] [64]

Design equivalent paths from the start

Provide an approved baseline tool or a nondependent alternative for assignments, open supervised access hours, put critical reading and evidence in offline-friendly formats, and offer offline or lower-bandwidth routes where relevant. Assign technical help rather than making lack of comfort a measure of achievement. Do not require consumer sign-ups or personal accounts to complete mandatory work.

Monitor equity as implementation quality

Report uptake, device or access failure, evaluation completion and learning gains with appropriate comparison groups. Compare how access and usefulness differ jointly; for instance, ask whether students with reliable laptops but limited study time complete the same number of practice attempts. Suppress or combine small cells under suitable privacy rules. Treat significant unmet access as an implementation hazard to correct before comparing learners' outcomes.

Accessibility calls for a tested accommodation route

Disability inclusion · Language and modality · Technology as support, not substitute

Potential utility in specific tasks

Speech-to-text, contrast enhancement or controlled summaries can help some learners access instructions; translated explanations may help some families or multilingual students participate. Generated accessibility output can also be inaccurate or incomplete. Check whether a description captures a chart's material information, whether captions match key content and whether translation preserves qualifiers. Offer an accessible non-AI equivalent when the tool fails. [82] [69]

Distinguish platform from accommodation

An institution should not use a policy against student-facing generative tools to unintentionally remove required assistive technology. New York City's 2026–27 guidance explicitly excludes assistive technology provided through an Individualized Education Plan or Section 504 plan from its AI-and-screen-time restriction and recognizes additional necessary tools. The exact treatment of a student's situation must follow applicable local processes. [83]

Conformance is not the full user experience

WCAG 2.2 offers testable web-content criteria, including text alternatives, keyboard access and navigable controls. Pair conformance review with actual tasks: start a session without a mouse, get meaningful screen-reader feedback, correct transcription, navigate error messages and access mathematical notation. Evaluate the tool and any generated artifact it expects students to read. A vendor claim or one successful login is not a usability study. [82]

Maintain an appeal and fallback

Include affected students and accessibility specialists in low-risk design testing without asking students to expose sensitive information unnecessarily. Record what the approved route is for a failure, how quickly help arrives, and whether a human can review an incorrect automated result. Evaluate outcomes separately rather than assuming one assistance feature resolves every need.

Student data protection is a procurement issue

U.S. baseline examples · FERPA and COPPA are distinct · Seek qualified advice

Start with the data flows

Before using a product, record the parties with access to student information, what they receive, the purpose, retention period and which subprocessors can use it. Include prompts, files, transcripts, screenshots, account details and embedded platform integrations. A tool reviewed last year can create a new data flow when it adds an assistant or changes a setting.

Two different federal regimes

FERPA applies to covered educational institutions and regulates education-record disclosures, subject to specified exceptions. The U.S. Education Department's privacy guidance discusses conditions for studies, audits and other uses. COPPA separately regulates covered operators' collection and use of personal information from children under 13. FTC guidance limits school authorization to the educational context for the school's benefit, not unrelated commercial purposes. These are separate checks and do not exhaust state or local obligations. [75] [84]

Set approval gates

Use synthetic teaching material for early testing. For a live pilot, specify purpose limitation, settings that govern secondary use and training, restricted staff access, audit and breach provisions, data export, verified deletion and contract-change notices. Do not assume a student or teacher accepting a consumer terms-of-service screen answers these questions. Require the vendor to explain defaults and where choices actually take effect.

Keep procurement and academic authorities separate

A curriculum leader should confirm educational value; a privacy or security lead should validate protections; an accessibility lead should check inclusion; and a program owner should maintain the approval. A product may meet one objective and fail another. Maintain a public-facing description in plain language, a contact for questions and a practical route to stop processing or change scope if the product changes.

U.S. education policy operates at several levels

Federal grant priorities · State implementation · Local classroom rules

A grant priority is not a classroom mandate

The U.S. Department of Education's July 2025 grant guidance described allowable AI-supported educational uses subject to existing program and regulatory requirements. Its final priority on advancing AI in education, published April 13, 2026 and effective May 13, can be used in discretionary competitions. It is not a requirement to buy AI, a conclusion that any named product is safe, or a substitute for checking the relevant grant program. [85] [86]

Tennessee uses local policy as one lever

Tennessee Public Chapter 550 requires local education boards and public charter governing bodies to adopt policies on instructional and assignment uses by students, teachers and staff. It also addressed certain Tennessee public higher-education governing boards. The statewide duty to maintain policies is distinct from the local choice about if, where or for whom a specific tool is permitted. [87]

Compare with a restricted local rollout

New York City Public Schools adopted a 2026–27 moratorium on student-facing generative AI for grades 2-K through 8, subject to its guidance and accommodation exceptions, and limited high-school use to vetted purposes while introducing AI literacy modules. The mayor's announcement describes a one-year policy with a review, not the prohibition of all school technology or a national rule. [83] [88]

What can coexist

A federal grant may support AI curriculum development, a state may require local rules, and a locality may prohibit a younger cohort from using a particular class of product. A policy inventory should specify issuer, jurisdiction, population, scope, exceptions, term and review owner. Distinguish final, effective rules from proposals and convenience summaries; do not turn elective teaching recommendations into purported mandates.

International education frameworks make different choices

England product guidance · Australia school framework · UNESCO competencies

England focuses on product safety and procurement

The Department for Education's product safety standards, updated January 2026, distinguish intended purposes and cover matters such as age-appropriate filtering, social and cognitive development, safeguarding, and monitoring. They chiefly guide developers and suppliers and can inform school purchasing. On November 10, 2026 the department opened a consultation on proposed edtech and AI product safety certification. A consultation is not an operative certification scheme or a guarantee about the status of any vendor. [89] [90]

Australia applies a school-wide framework

The Australian school framework organizes responsible generative AI use around teaching and learning, human and social wellbeing, transparency, fairness, accountability, and privacy, security and safety. Education ministers endorsed the review in June 2025. The framework helps institutions develop practices rather than setting a single universal product approval. [91]

Cross-border portability has limits

UNESCO's competency resources help define what students and educators should know; England's standards help ask whether a product meets the needs of a specific learner population; Australia's framework offers school governance principles. These are not interchangeable. Keep the educational objective, practical safeguards, legally binding local requirements and voluntary design principles separately in any procurement checklist. [76] [77]

What a common checklist can ask

Across jurisdictions, an institution can request intended use by grade, evidence in comparable classrooms, age-appropriate interfaces, limits for secondary data uses, human redress, accessible alternatives, and a documented process when product design changes. Validate jurisdiction-specific rules at implementation. Translate framework concepts into tests rather than assessing compliance by the mere appearance of words like safety or fairness in a vendor presentation.

Design a school pilot that can be stopped

Compare with the best feasible alternative · Protect the baseline · Set review dates

Pick a narrow classroom problem

State the objective before choosing the tool: improve algebra explanations, reduce time spent formatting a worksheet, or provide accessible feedback to students who currently lack it. Define independent achievement, time, quality and safety measures. Identify whether the intervention is teacher-facing or learner-facing; the consent, oversight, data and testing burden may differ substantially. [69] [89]

Balance benefits and burdens

Predefine local decision rules and the control or comparison practice. Include no-AI baseline tasks, delayed tests and notes about age, accommodations and actual usage. Track tool cost, review time, training, accessibility issues, corrective work, classroom displacement and disciplinary disputes. Use comparable denominators and don't assign implementation failures to the learner.

Verify failure modes before wider exposure

Test with simulated inputs before asking students to submit work. Ask whether the tutor gives answers too early, fabricates a source, mishandles an accessibility request, or wanders beyond the allowed topic. Establish a staffed escalation path and a method to disable the feature promptly; a pilot that cannot be stopped is an operational dependency, not merely an experiment. Keep students' welfare and work available through the fallback method.

Review evidence not product enthusiasm

Give decision-makers a short report showing assignment rules, participant groups, full costs, tradeoffs and adverse events. Separate actual service use from favorable opinions. Request review by teaching staff and relevant privacy and accessibility specialists, and a public-facing summary of what changes. If evidence is inconsistent, limit the rollout and replicate before replacing a proven teaching method. [69] [89]

Exhibit 12 · AI in class: keep people in control

This proposed sequence puts teachers in charge of the objective and assessment. Student work and independent transfer checks help distinguish learning from AI-assisted task completion. Use approved tools and applicable student-data safeguards.

48PEGASUS Research conceptual exhibit. Not empirical data.

Adult learning and job transitions need different measures

Reskilling · Participation constraints · Employment is a downstream measure

Start with the worker's actual transition

Someone facing changed tasks in an existing job needs refreshed domain practice and feedback, while a displaced worker needs access to paid work, income stability and help demonstrating skills to new employers. A common tool-usage badge is not a sufficient response to either. Assess foundational literacy, numeracy and problem solving separately from product orientation; people may need support in one dimension and already perform strongly in another. [79]

Reduce opportunity costs

Offer accessible instruction schedule options, paid learning time when feasible, and reasonable device access. In delivery roles, focus on supervised verification and handoffs rather than asking workers to generate polished text they will not be paid to use. Training is most credible when tied to tasks for which employers need verified competence.

Measure outcomes over enough time

Record who was eligible, enrolled, completed and was placed or promoted; document paid hours, wages and job quality before and after. A worker may complete training and still lack an opening in the relevant region. Measure six- and twelve-month outcomes where feasible and distinguish program selection effects from causal impacts. Payment and mentorship can be tracked directly for programs modeled on apprenticeship. [80]

Preserve options across employers

Incentivize portable evidence and practical experience, not dependence on one brand or account. Have workers verify outputs against records, explain choices to supervisors, and practice using a fallback when tools fail. Include non-AI routes for situations where privacy or safety precludes delegation. Skill metrics should tell people what they can do independently and under supervision; employment metrics should tell them where that ability translates into a real opportunity.

A school-to-work evidence dashboard

Suggested operating model · Use measure-specific denominators · Audit failure and benefit

Learning and access

A useful dashboard reports attendance and participation, assistive-versus-substitutive use, reach among students needing extra support, unaided skill, delayed transfer, and valid gains for learners otherwise lacking access. Show each metric's definition, timeframe and population. Put no-use results on the same dashboard as use results, but do not equate nonuse with a deficit. Higher participation can coexist with poorer independent outcomes. [64] [68]

Time and capacity

For educators, report preparation time net of verification; material quality; instructional contact; incident response burden; and tool spending. Count time actually reallocated, not just perceived savings. For student transitions, track real paid learning placements, completed mentorship, attainment of transferable competence, and subsequent job outcomes. [66] [67] [80]

Tradeoffs and distribution

Display privacy incidents, access failures, unresolved appeals, and opportunity gaps with suitable protection for small groups. Highlight gaps rather than smoothing them into averages. Do not reward institutions simply for using a product more often. Protect valid accommodations and allow students, workers and teachers to share practical failures without fear of punitive review.

Review the policy periodically

Identify what data the leadership team is allowed to collect, who can review it, and which thresholds trigger a redesign, halt or independent review. Reevaluate after major changes in vendor, model, task mix or student population. Keep a plain-language record that spells out what was measured and what did not follow from those measurements. The decision is not whether AI should exist in schools. It is whether a defined use produces enough learning or access to justify its costs and risks. [69]

Previous chapter

Next chapter

Back to the report

Browse: White Papers · Reports · Downloads

bottom of page