Research Library
The Evidence on AI in K‑12
Evidence Base & Landscape
Featured Review · Stanford SCALE
The Evidence Base on AI in K‑12: A 2026 Review
Screened 800+ papers on AI in K‑12 to find out what’s actually been proven.
Key Findings
- Only 20 of 800+ papers met the bar for strong causal evidence
- The clearest available answer to whether AI tools actually work
Think Tank Report · CRPE
Getting Beyond the Lightbulb Stage
A CRPE brief on why AI hasn’t transformed schools yet.
Key Findings
- Based on 50+ interviews with funders, developers, and district leaders
- Market, vision, and infrastructure gaps are all still unresolved
Think Tank Report · Bellwether
Productive Struggle
Bellwether’s framework for when AI helps learning versus when it’s a shortcut.
Key Findings
- Looks at memory, attention, motivation, and self-regulation
- Ease isn’t automatically good — sometimes it hides a cost
National Survey · RAND
Using Artificial Intelligence Tools in K‑12 Classrooms
RAND’s first national look at how AI was actually used in classrooms.
Key Findings
- 18% of teachers used AI for teaching as of fall 2023
- Warns AI adoption could deepen inequality without intervention
Meta‑Analysis · Johns Hopkins CRRE
A Quantitative Synthesis of Outcomes of Ed Tech in K‑12 Math
A meta-analysis of ed-tech’s real effect on math learning.
Key Findings
- 74 studies, 56,886 students: small but real positive effect
- Smaller effect size for students from low-income backgrounds
Qualitative Study · Friday Institute
Educators’ Perspectives on Generative AI in K‑12
NC State’s Friday Institute on how educators are actually experiencing AI.
Key Findings
- Compares educator experience against 12 states’ official guidance
- Flags equity, workload, and privacy as top concerns
Mixed Methods · Research in Learning Tech.
Perceptions and Preparedness of K‑12 Educators
What’s standing in the way of educators feeling ready for AI.
Key Findings
- Links AI familiarity directly to perceived readiness
- Insufficient PD is the most-cited barrier
Report · University of Sydney
Young people, learning, and generative AI: a rapid review for PreK-12
A learning-sciences read on what generative AI actually does to young people’s learning
Key Findings
- Affective gains are common, but weak indicators of durable learning
- Effects on higher-order reasoning, transfer, and self-regulated learning remain uneven
No named publisher
National Survey · Walton Family Foundation & Gallup
Voices of Gen Z: Year 4 annual survey
Gallup’s fourth annual national read on how Gen Z students feel about school
Key Findings
- Just 48% of Gen Z students look forward to school most days
- 33% do not feel they belong; four in ten feel unprepared
Qualitative Study · Discover Education
K‑12 leaders’ perspectives on the implications of artificial intelligence for educators’ work
Interviews with 36 school and district leaders about how AI changes educators’ jobs
Key Findings
- Leaders sorted AI uses into routine, instructional, operational, and decision-making categories
- Hands-on experience using AI predicted more balanced, better-informed leader views
Interviews conducted in late 2023
Open paper →Empirical Study · Computers and Education: AI
Integrating generative AI in K‑12: teachers’ preparedness, practices, and barriers
Survey of 89 rural Idaho K‑12 teachers on whether they are ready for AI
Key Findings
- Fewer than half of the 89 teachers used generative AI in practice
- Users applied it to lesson prep and admin, not live teaching
Single-state sample
Open paper →Meta‑Analysis · Humanities and Social Sciences Communications
The effect of ChatGPT on learning performance, perception, and higher-order thinking
Pooled 51 experiments to size up what ChatGPT does to learning outcomes
Key Findings
- 51 studies: large effect on learning performance, Hedges’ g = 0.867
- Smaller gain for higher-order thinking (g = 0.457); 4-8 weeks most stable
Cognitive Offloading & Cognitive Cost
Field Experiment · PNAS
Generative AI Without Guardrails Can Harm Learning
A high school field experiment on AI tutoring with and without guardrails.
Key Findings
- AI tutoring boosted practice scores, then hurt exam scores once removed
- Withholding direct answers erased the harm entirely
RCT · Computers & Education
Effects of LLM Use & Note-Taking on Reading Comprehension
A randomized experiment inside real secondary-school classrooms.
Key Findings
- Tests how note-taking habits interact with LLM use
- Conducted in actual classrooms, not a lab
Panel Study · Working Paper
The generative AI learning penalty: Evidence from Chinese secondary education
Thirty months of data on 26,811 secondary students before and after they adopted AI tools.
Key Findings
- Homework scores rose 18% while closed-book exam scores fell 20% within six months
- Entrance exam scores dropped 18% and 24%, the full penalty taking about two years
Unpublished manuscript, not peer reviewed
A ten-year panel showing study time collapsed on exactly the math problems AI can do.
Key Findings
- High school study time on AI-susceptible math problems fell 31.3%
- Proctored retention items showed a 25% cumulative decline in odds of answering correctly
Preprint, not peer reviewed
Open paper →Empirical Study · MIT Media Lab
Your brain on ChatGPT: Accumulation of cognitive debt when using an AI assistant for essay writing task
MIT used EEG to watch what happens in the brain during AI-assisted essay writing.
Key Findings
- Brain connectivity fell as tool support rose; LLM users showed the weakest coupling
- LLM users reported low essay ownership and could barely quote their own writing
Preprint, not peer reviewed; 54 participants
Open paper →RCT · British Journal of Ed. Technology
Beware of metacognitive laziness: Effects of generative artificial intelligence on learning motivation, processes, and performance
The randomised study that named metacognitive laziness, comparing ChatGPT against a human expert.
Key Findings
- The ChatGPT group improved essay scores most but gained no knowledge or transfer
- Self-regulated learning behaviours differed sharply while intrinsic motivation did not
Higher ed sample
Open paper →Randomised trials with 1,222 people testing what is left once the AI help is withdrawn.
Key Findings
- AI users performed significantly worse without the tool and gave up sooner
- The effect appeared after roughly ten minutes of AI interaction
Preprint, not peer reviewed
Open paper →An experiment separating students who had AI explain concepts from those who had it write.
Key Findings
- AI access raised immediate test scores 0.27 standard deviations, holding a week later
- Automation users lost their quality advantage as soon as the tool was removed
Preprint, not peer reviewed; undergraduate sample
Open paper →An experiment on whether AI helps or hurts critical thinking depending on when you get it.
Key Findings
- Early LLM access helped under time pressure but hurt when time was sufficient
- Starting the task independently produced the opposite pattern
Survey Study · Societies
AI tools in society: Impacts on cognitive offloading and the future of critical thinking
A 666-person study linking heavy AI use to weaker critical thinking through cognitive offloading.
Key Findings
- Frequent AI use correlated negatively with critical thinking, mediated by cognitive offloading
- Younger participants showed the most AI dependence and the lowest critical thinking scores
Correlational, not causal
Open paper →White Paper · EDSAFE AI Alliance
Future-proofing human flourishing: The case for a learning sciences benchmark for AI
Proposes a learning-sciences benchmark for judging whether an AI tool builds or erodes competence.
Key Findings
- Names a ’Zone of No Development’ and a ’Crutch Effect’ from AI overreliance
- Proposes three pillars and argues voluntary industry standards are not enough
Empirical Study · Computers in Human Behavior
AI-mediated teaching in K‑12 classrooms
Interviews with 23 high school teachers about what AI quietly takes over in their work.
Key Findings
- AI cut workload but displaced diagnostic reasoning, sequencing, and evaluative judgment
- Teacher agency persisted but became conditional on institutional pressure and algorithmic opacity
Teacher interviews, not student outcomes
Open paper →Tutoring & Personalized Learning
RCT · Stanford
Access Is Not Enough
Two RCTs on whether students actually use AI tutors when given the chance.
Key Findings
- Nearly half of students never logged into the AI tutor at all
- Pairing it with a human tutor helped, but not enough to move reading scores
Systematic Review · npj Science of Learning
AI‑Driven Intelligent Tutoring Systems in K‑12
A Nature-family review focused specifically on K‑12 tutoring systems.
Key Findings
- Covers intelligent tutoring systems built for K‑12 classrooms
- Maps out where the evidence is strong versus still thin
Meta‑Analysis · Computers & Education
Does ChatGPT Enhance Student Learning?
A meta-analysis of experimental studies, not just surveys.
Key Findings
- Positive effects found on performance and higher-order thinking
- Synthesizes results across many independent experiments
Field Study · Preprint
Learning to Prompt
An adaptive tutoring system tested with real high schoolers.
Key Findings
- Tested with 359 real students across 656 conversations
- Adaptive prompting outperformed static tutoring scripts
Bias Study · Stanford
Marked Pedagogies
Tested 4 major LLMs on real 8th-grade essays for bias.
Key Findings
- Feedback shifted based on a student’s race, gender, and learning needs
- Shows “personalized” AI feedback isn’t neutral
Framework · Preprint
Building AI Companions That Prioritise Learning Over Performance
A design case, not an outcomes study, for slower AI tutors.
Key Findings
- Argues AI tutors should resist giving the fastest answer
- Prioritizes durable learning over the feeling of instant progress
RCT · Google LearnLM
Teaching with Gemini: measuring Guided Learning’s impact on math progress in Sierra Leone
Preregistered trial of Gemini’s Guided Learning in 48 junior secondary math classrooms
Key Findings
- 1,763 students: intent-to-treat effect of +0.258 standard deviations on math
- Full 12-hour dosage reached +0.380 SD, roughly 1.2-1.7 years of progress
Industry-authored; non-US setting
First randomized trial of an AI coach feeding live guidance to math tutors
Key Findings
- 900 tutors, 1,800 students: 4 percentage point gain in topic mastery
- Gains hit 9 points for students of lower-rated tutors, $20 per tutor
Preprint, not peer reviewed
Open paper →RCT · Annenberg Institute, Brown
One click away: AI tutoring with Khanmigo in a two-year school experiment
Two-year randomized trial of Khanmigo in 18 Tennessee middle schools’ remedial math
Key Findings
- Gains of 1.3 national percentile ranks per term, about 0.06-0.08 SD yearly
- 96% tried Khanmigo but the median student messaged it a third of days
Working paper, not peer reviewed
Open paper →RCT · Google LearnLM & Eedi
AI tutoring can safely and effectively support students: an exploratory RCT in UK classrooms
Tested supervised AI tutoring against human tutors with 165 UK secondary students
Key Findings
- Tutors approved 76.4% of LearnLM’s drafted messages with zero or minimal edits
- 66.2% versus 60.7% solved novel problems, a 5.5 point edge over humans
Industry-authored; small sample
Open paper →Empirical Study · Khan Academy
Cognitive engagement in GenAI tutor conversations: at-scale measurement and impact on learning
Scored 200,000 Khanmigo tutoring chats for how hard students were actually thinking
Key Findings
- About 9,000 students and 200,000 tutoring threads across six U.S. districts
- Higher cognitive engagement predicted better performance on the next practice item
Works-in-progress paper
Report · Stanford SCALE & NSSA
Research on AI tutoring: what the evidence says and what it means
Sorts AI tutoring into four models and says which ones the evidence supports
Key Findings
- One study found 40-47% of students never logged into the AI tutor
- Strongest results came from AI built for tutors, not for students
Summarizes a separate brief; no named publisher
Assessment, Integrity & Critical Thinking
Framing Piece · Assessment & Eval. in HE
The Wicked Problem of AI and Assessment
Argues AI exposed a problem that already existed in assessment.
Key Findings
- Traditional assessment was never a clean measure of understanding
- A foundational, widely-cited framing piece
Framing Piece · Assessment & Eval. in HE
Black Box Assessment
Makes the case for assessing visible, documented thinking.
Key Findings
- Pushes back on trying to “catch” AI use after the fact
- Proposes assessing process, not just final products
Foundational · Preprint, 2023
Chatting and Cheating
One of the earliest, most-cited looks at ChatGPT-era integrity.
Key Findings
- 244+ citations since its 2023 preprint release
- A foundational reference point for the whole field
Landmark Study · Patterns
GPT Detectors Are Biased Against Non-Native English Writers
A landmark, widely-cited bias finding in AI detection tools.
Key Findings
- Non-native English writers flagged far more often as “AI-written”
- Raises real fairness concerns for detection-based policies
Empirical Test · Intl J. Educational Integrity
Testing of Detection Tools for AI-Generated Text
A head-to-head test of popular AI-detection tools.
Key Findings
- Tested multiple detection tools side by side
- Found them unreliable across the board
Systematic Review · Soc. Sci. & Humanities Open
Reassessing Academic Integrity in the Age of AI
A 2025 systematic review of where the integrity debate stands now.
Key Findings
- Synthesizes findings across dozens of studies
- Maps how the conversation has shifted since 2023
Framework · Educational Researcher
The AI3 Model
A cross-national framework for assessment innovation.
Key Findings
- Spans K‑12 and higher education contexts
- Published in AERA’s flagship journal
Synthesis · Preprint
The Effortless Trap
Argues placement, not permission, determines AI’s effect on learning.
Key Findings
- An unguarded AI helper left students ~17% worse on an unaided exam
- A well-engineered tutor roughly doubled learning instead
Survey Study (Contested) · Societies
AI Tools in Society
Links frequent AI use to lower critical thinking scores.
Key Findings
- A widely-debated survey of 666 participants
- Read alongside its published correction
Position Paper · RELC Journal
Tomorrowland? Technical innovation and feedback on writing
Asks whether AI writing feedback teaches writing or just acts as a marking machine
Key Findings
- Automated writing evaluation shows qualified gains: fewer errors, more drafts, greater autonomy
- Current tools still lack the contextual awareness and higher-order judgment teachers bring
Language-learning focus
Open paper →Systematic Review · Khazar Journal
AI detection tools: a systematic review of empirical evidence and implications for education
Pulled together 25 studies on whether AI text detectors are reliable enough to use
Key Findings
- 25 studies, 18 in classrooms: accuracy unstable across tools, genres, languages
- Unsuitable for high-stakes integrity decisions; authors urge assessment redesign instead
Empirical Study · Int’l Journal of Ed Tech in Higher Education
Simple techniques to bypass GenAI text detectors
Tested whether easy tweaks let AI-written text slip past six major detectors
Key Findings
- Detector accuracy dropped 17.4% once simple bypass techniques were applied
- 805 texts tested; authors say detectors cannot justify integrity violations
Higher ed focus
Open paper →News Analysis · The New York Times
Student cheating is becoming impossible to detect in an A.I. era
Reports on humanizers and autotypers that help students slip past AI detection software
Key Findings
- Humanizers rewrite AI text; autotypers fake human typing, typos and revisions
- Some firms sell both detection tools and the apps that evade them
Journalism, not research
Open paper →Systematic Review · Computers and Education: AI
The cognitive impact of ChatGPT in higher education: critical and creative thinking outcomes
Reviewed 67 studies on what ChatGPT does to critical and creative thinking
Key Findings
- 67 studies: gains appear only in inquiry-oriented, scaffolded instructional designs
- Unstructured use favored creativity over critical thinking, sometimes declining in both
Higher ed sample
Open paper →Empirical Study · Discover Education
Dedicated learning modules to inform K‑12 students about academic integrity
Built and tested video modules teaching K‑12 students six kinds of academic dishonesty
Key Findings
- 1,181 students across 10 schools improved on pre- and post-tests
- Junior high students benefited most; modules cover AI misuse and contract cheating
Non-US setting
Open paper →Empirical Study · LAK 2025
Can AI grade your essays? Large language models versus teacher ratings
Pitted five language models against 37 teachers scoring real student essays on ten criteria
Key Findings
- o1 correlated .74 with teachers and reached .80 internal consistency
- All models over-rated essays; open-source models correlated weakly with human raters
20 German essays, small set
Open paper →Agentic AI, Literacy & Policy
Foundational · Computers & Ed: AI
Conceptualizing AI Literacy
The foundational framework most later AI-literacy work builds on.
Key Findings
- Defines AI literacy as technical knowledge plus critical evaluation
- Cited across most later AI-literacy frameworks
Framework · Preprint
Addressing the Reality Gap
Three tensions schools face in adopting more autonomous AI.
Key Findings
- Names feasibility, adaptation speed, and mission alignment
- A preprint chapter for Springer’s forthcoming AI-in-education handbook
Industry Data · OpenAI
The Shift to Agentic AI: Evidence from Codex
OpenAI’s own usage data on agentic AI adoption.
Key Findings
- Agentic AI usage grew more than 5x in six months
- Adoption is spreading well beyond software developers
Perspective · Frontiers in Education
The Cognitive Mirror
A conceptual framework, not a data study, on AI and metacognition.
Key Findings
- AI’s value may be reflecting a learner’s thinking back to them
- Proposes building self-regulation, not just giving answers
Framework · OECD & European Commission
Empowering learners for the age of AI: an AI literacy framework for schools
The finalized OECD framework defining what AI literacy means for primary and secondary students
Key Findings
- 22 competences across four domains: Engage, Create, Manage, and Shape AI
- Final version strengthens metacognition, student agency, and healthy skepticism toward AI
Systematic Review · Computers and Education: AI
Unleashing human potential: an AI competency framework for K‑12 education
Scoping review of 54 studies behind a three-part AI competency model for K‑12
Key Findings
- 54 studies converge on AI knowledge, practical skills, and ethical awareness
- Existing frameworks underemphasize human values; adds an ’Unleashing’ developmental component
Position Paper · Khalid, Islam & Ajmal
Reimagining curriculum knowledge through agentic AI systems
Argues agentic AI should reshape curriculum from fixed content into a living process
Key Findings
- Proposes ’AI-augmented curriculum agency’ pairing human intentionality with machine intelligence
- Casts AI as a knowledge partner rather than a classroom tool
Conceptual, no empirical data
Systematic Review · Applied Resilience and Sustainability
Agentic AI-driven tutoring: a multi-agent cognitive architecture for personalized adaptive learning
Reviews what multi-agent AI tutoring architectures promise and where they still fall short
Key Findings
- Multi-agent systems improve behavior modeling, real-time feedback, and metacognitive support
- Privacy, algorithmic transparency, scalability, and missing evaluation standards remain unresolved
Low-profile venue
No papers match that search — try a different term or clear the filter.