Explore the research

Education

Do students remember more, apply what they learn and benefit after the AI assistance stops?

15%

field summary estimate

6 outcomes · Assessed November 7, 2026 · Low confidence

The field estimate averages its outcome scores. It is a reading aid, not a measured fraction of the field that AI manages.

How the estimate built up.

Each band is one outcome's share of the field estimate, so the top line is the estimate itself. Every field starts at 0% on December 31, 2022. Dates before the latest are retrospective estimates. Point at a band to see that outcome's score on each date.

Transfer15%
Time to mastery25%
Retention10%
Learning gain15%
Instructional cost10%
Access and disparities15%

View exact historical values and reconstruction notes

What we measure

Each percentage is a judgment about progress toward independent advancement. Open the evidence to inspect the basis for it.

Access and disparities

15%estimated progress

What we track

Participation, completion, and learning outcomes by relevant student groups, prior attainment, language, and accessibility needs.

Why this estimate?

Limited prior-attainment subgroup evidence supports some reach, but equitable access and durable outcomes across languages, disabilities and populations are not established.

Advancements

Hybrid human-AI tutoring helps in three low-income middle schoolsDecember 18, 2023. Three quasi-experimental studies in urban, low-income middle schools (125, 385 and 75 students) found positive effects of human-AI tutoring over math software alone on proficiency and usage, with evidence that lower-achieving students may benefit more.
Lab experiments: LLM access widens gap by prior knowledgeAugust 29, 2024. Two pre-registered, incentivized lab experiments found no effect of LLM access on overall learning outcomes, but LLMs widened the gap between students with low and high prior knowledge; substituting LLMs for practice reduced understanding of each topic.
Tutor CoPilot: students of lower-rated tutors gain 9 pointsOctober 3, 2024. In the randomized trial of 900 tutors and 1,800 K-12 students, the benefit of the AI assistant was concentrated among students of lower-rated tutors, whose topic mastery rose 9 percentage points versus 4 points overall.
Nigeria AI tutoring trial: largest gains for female studentsMay 19, 2025. In the randomized evaluation of a six-week LLM tutoring programme in Nigeria, students across the baseline ability distribution benefited, with the largest effects for female students and for those with higher initial academic performance.
AI tutoring helps both stronger and weaker physics studentsJune 3, 2025; experiment fall 2023. The study found better immediate post-test performance with tutoring in groups above and below 40% on the physics-concept pretest.

Read the dated findings and sources

Status: Prior-attainment subgroup evidence. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

June 3, 2025; experiment fall 2023: The study found better immediate post-test performance with tutoring in groups above and below 40% on the physics-concept pretest. AI tutoring randomized crossover study.

Interpretation and tracking value

One university's attainment subgroups do not establish equitable access across income, language, disability or geography. Eligibility and participation need denominators.

Next comparable measurement

Report invitation, participation, completion and retained outcomes by group; test accessibility and language support with prespecified subgroup analyses. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Instructional cost

10%estimated progress

What we track

Total dollars and teacher minutes per student reaching retained mastery, including onboarding, monitoring, and correction.

Why this estimate?

Reusable tutoring systems offer an efficiency mechanism, but full dollars and teacher minutes per retained mastery success are still missing.

Advancements

Tutor CoPilot costs $20 per tutor per year in randomized trialOctober 3, 2024. Based on tutors' usage during the randomized trial with 900 tutors and 1,800 students, the Tutor CoPilot AI assistant cost $20 per tutor annually while raising student topic mastery by 4 percentage points.
AI tutor lessons take days to prepare, and the platform monthsJune 3, 2025; experiment fall 2023. Preparing individual AI-supported lessons took days, while initial tutor-platform development took months.

Read the dated findings and sources

Status: Implementation effort documented; cost per retained mastery unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

June 3, 2025; experiment fall 2023: Preparing individual AI-supported lessons took days, while initial tutor-platform development took months. AI tutoring randomized crossover study.

Interpretation and tracking value

These qualitative durations are not total dollars or teacher minutes per retained-mastery success. Platform development must be amortized transparently.

Next comparable measurement

Measure setup, licensing, monitoring and correction costs plus teacher minutes; divide by students passing delayed unaided mastery tests. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Learning gain

15%estimated progress

What we track

Difference in independently administered assessment scores after adjusting for baseline, with absolute and standardized effects and uncertainty.

Why this estimate?

Controlled assessment exists, but the cited negative trial prevents credit for dependable learning gains from tutor access alone. Humans still design and evaluate instruction.

Advancements

Ghana study: Rori AI math tutor yields 0.37 effect sizeFebruary 15, 2024. Study of about 1,000 students in grades 3-9 across 11 schools in Ghana: two 30-minute sessions per week with the Rori AI math tutor over 8 months produced an effect size of 0.37 on math growth (p < 0.001).
Tutor CoPilot trial: students 4 points more likely to master topicsOctober 3, 2024. Randomized trial with 900 tutors and 1,800 K-12 students: students whose tutors had the AI assistant Tutor CoPilot were 4 percentage points more likely to master topics, rising to 9 points for students of lower-rated tutors.
Nigeria trial: six weeks of GPT-4 tutoring raises scores 0.31 SDMay 19, 2025. Randomized evaluation of a six-week after-school programme using an LLM as a virtual English tutor for first-year senior secondary students in Nigeria: 0.31 SD gain on the overall assessment and 0.23 SD in English, described as equal to 1.5 to 2 years of business-as-usual schooling.
Khanmigo two-year trial: 0.06-0.08 SD math gain per school yearAugust 2026. Two-year cluster randomized trial in 18 Tennessee middle schools: assignment to Khan Academy with the Khanmigo AI tutor raised math achievement 1.3 national percentile ranks per term (about 0.06 to 0.08 SD per year), similar to Khan Academy practice without AI; students messaged the tutor in only 17% of sessions with a mistake.
Trial finds a course AI tutor lowered final gradesOctober 2026. An October 2026 working paper covering 2,379 students and 30 instructors reports AI-tutor access reducing final grades by 0.37 standard deviations and LMS participation by 0.90 standard deviations.

Read the dated findings and sources

Status: Randomized evidence includes negative effects. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

October 2026: An October 2026 working paper covering 2,379 students and 30 instructors reports AI-tutor access reducing final grades by 0.37 standard deviations and LMS participation by 0.90 standard deviations. IZA: course-integrated AI tutoring randomized trial.

Interpretation and tracking value

This working-paper result concerns a particular course-integrated intervention, not all AI tutoring. Access, actual use and pedagogy must be separated.

Next comparable measurement

Replicate with independent assessment and baseline adjustment; report intention-to-treat effects, uptake, subject, instructional design and uncertainty. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Retention

10%estimated progress

What we track

Performance on new equivalent assessments after 30 and 90 days, completed without AI assistance.

Why this estimate?

Tutoring and unaided assessment are demonstrated, but the specified 30- and 90-day retention evidence is absent; sustained autonomous instructional improvement receives no credit.

Advancements

Experiment separates AI-assisted practice from unaided exams2025 publication. The experiment distinguishes assisted practice from later unaided examination performance.
Trial: ChatGPT learners score 57.5% vs 68.5% on test 45 days laterNovember 2025. Randomized trial of 120 undergraduates: on a surprise retention test 45 days after learning, the ChatGPT-assisted group scored 57.5% versus 68.5% for the traditional-study group (t(83) = -3.19, p = .002, Cohen's d = 0.68).
Tennessee trial: AI math tutor shows marginal gains on delayed testAugust 2026. Randomized field experiment with more than 6,000 middle-school students on the NUMI practice platform: on an assessment one week later, mastery progression alone did not improve delayed learning; AI support embedded in the mastery workflow showed marginally significant gains on practiced material.

Read the dated findings and sources

Status: Immediate unaided assessment; required delayed baseline unknown. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2025 publication: The experiment distinguishes assisted practice from later unaided examination performance. Generative AI without guardrails can harm learning, PNAS.

Interpretation and tracking value

Its exam result must not be relabeled as a 30-day or 90-day retention estimate. This review does not establish the required paired delayed measures.

Next comparable measurement

Administer new equivalent unaided assessments at 30 and 90 days; report attrition, prior attainment and uncertainty for all randomized learners. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Time to mastery

25%estimated progress

What we track

Student hours needed to reach a frozen mastery criterion, including failed attempts; keep non-completers in the analysis.

Why this estimate?

Shorter lessons with immediate learning gains support useful personalized instruction. A fixed mastery criterion including non-completers and delayed learning remains unmeasured.

Advancements

Harvard study: AI lessons take a median 49 minutesJune 3, 2025; experiment fall 2023. The Harvard crossover study included 194 eligible students and reported median AI-lesson time of 49 minutes with improved immediate learning.
Tennessee trial: students with AI support progress more slowlyAugust 2026. In a randomized experiment with more than 6,000 middle-school students, those given AI tutoring support progressed more slowly and attempted fewer questions but were more accurate; after an error they needed fewer attempts to return to a correct answer, while spending more time per supported question.

Read the dated findings and sources

Status: Measured lesson-time proxy. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

June 3, 2025; experiment fall 2023: The Harvard crossover study included 194 eligible students and reported median AI-lesson time of 49 minutes with improved immediate learning. AI tutoring randomized crossover study.

Interpretation and tracking value

Lesson completion time is not time to a frozen mastery threshold. The study's inclusion criteria also prevent treating this as a complete non-completer-inclusive time-to-mastery estimate.

Next comparable measurement

Prespecify mastery; include all assigned learners and non-completers, repeat attempts, total study time and subsequent unaided testing. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note

Transfer

15%estimated progress

What we track

Success on unfamiliar problems, explanations, and practical tasks requiring the learned skill rather than memorized answers.

Why this estimate?

Guardrails can avoid a detected unaided-exam harm, but reliable benefit on unfamiliar unaided tasks and autonomous curriculum management are not established.

Advancements

Unguarded GPT help cuts later exam scores by 17%2025 publication. The PNAS field experiment found ordinary GPT assistance improved practice performance but reduced subsequent unaided exam performance by 17%; a guarded tutor avoided that detected harm.
LearnLM trial: 66.2% vs 60.7% success on novel problemsDecember 29, 2025. Exploratory randomized trial with 165 students in five UK secondary schools: students supported by tutor-supervised LearnLM solved novel problems on subsequent topics 66.2% of the time versus 60.7% for those with human tutors alone; tutors approved 76.4% of the AI's drafted messages with zero or minimal edits.
Anthropic study: AI-assisted coders score 50% vs 67% on skills quizJanuary 29, 2026. Randomized experiment with 52 mostly junior software engineers learning an unfamiliar Python library: those using AI assistance averaged 50% on a follow-up quiz versus 67% for those coding by hand (Cohen's d = 0.738, p = 0.01), with no statistically significant difference in completion time.

Read the dated findings and sources

Status: Measured unaided-test risk. Sources accessed November 7, 2026. This is a documentary baseline, not an independent replication or an assertion of the worldwide best result. An unknown measurement is not zero.

Dated findings

2025 publication: The PNAS field experiment found ordinary GPT assistance improved practice performance but reduced subsequent unaided exam performance by 17%; a guarded tutor avoided that detected harm. Generative AI without guardrails can harm learning, PNAS.

Interpretation and tracking value

The guarded tutor did not show a detectable exam benefit. Practice gains cannot be counted as independent skill transfer.

Next comparable measurement

Evaluate unfamiliar unaided problems and explanations; compare guarded and unguarded systems, and retain both practice and transfer scores. Preserve this dated observation; append future results with their model/system version, denominator, budget and evidence level rather than overwriting it.

Open the full research note