The Generative AI Generation in Higher Education · Evidence dashboard, Nov 2022 to Nov 2026

Sheet 02 of 17 · Longitudinal

Adoption over time

Repeated surveys show rapid growth on every measure. Filter by country and years, switch to a table, and select any point to see its sample and caveats.

Fig. 2.1: Reported use by survey series

UK series (HEPI)

US series (Tyton, Pew)

Dash pattern distinguishes series within a country. Select any point for its source.

SeriesSource2023202420252026
Years are report years; HEPI fieldwork runs the preceding December. Sources: HEPI 2024–2026; Tyton Partners 2023–2026; Pew Research Center 2023–2026.

Selected data point

·

Source
Sample
Caveat

Fig. 2.2: Measurement ladder

The same student can be counted very differently depending on the threshold a survey uses. These are not interchangeable.

Ever tried AI broadest

Used for coursework this year 85% · IHE 2025

Monthly or more 59% · Tyton 2024

Weekly 57% · Gallup 2026

Daily 32% · Tyton 2026

Table 2.3: How "longitudinal" is the evidence?

Every trend line above comes from repeated cross-sections: new respondents each wave. They show how the population changed, not how individual students' habits developed. No national panel yet follows the same students from high school through graduation.

Repeated cross-section

HEPI / Kortext ↗

UK full-time undergraduates · ~1,050 per wave · 2024, 2025, 2026

Strongest annual series. Pollster changed in 2024; some options reworded.

Repeated cross-section

Tyton, Time for Class ↗

US students, faculty, administrators · 2023–2026

Only multi-stakeholder series. Frequency thresholds and course scope shift by wave.

Repeated cross-section

Pew Research Center ↗

US teens 13–17 · probability panel · 2023, 2024, 2025

Best pipeline trend. 2025 widened "ChatGPT" to "any chatbot".

Repeated cross-section

College Board ↗

US high-school students · ~600 per wave · Jan and May 2025

Short window, unweighted samples.

Retrospective recall

Middlebury ↗

One college · single survey, late 2024–early 2025

Students recalled when they started: under 10% before spring 2023, over 80% by fall 2024. Depends on memory.

Emerging panel

Dartmouth AI Survey ↗

Multi-year study · 2026 baseline incl. 332 incoming Class of 2030

Could become the first true panel if it re-surveys the same students; one institution, so not nationally representative.

What a true panel would add

T0

College entry baseline

T1

End of first year

T2

End of second year

T3

Graduation

T4

Post-graduation

Linking each wave to grades, persistence and career outcomes. Design recommended in the incoming-cohort report.

Reading note: These series describe different populations, technologies and frequencies and should not be merged into one trend line. HEPI's general text-generation measure fell from 64% to 56% between 2025 and 2026 while overall adoption rose. Students may be shifting toward task-specific tools.

Fig. 3.1: Exposure chart by entering class

Each row is a fall entering class: four years of high school, then four of college. Shaded portions fall after ChatGPT's public release on 30 Nov 2022. Select a row to drill in.

CHATGPT · NOV 2022

High school, pre-AI

High school, AI era

College, pre-AI

College, AI era

Fall entrants · Class of

Projection, limited evidence

years of high school with public GenAI

years of college with public GenAI

Reading note: Cohorts are defined by educational exposure, not birth year, and they overlap. They also include adult, transfer, community-college and graduate students, not only traditional-age undergraduates.

Table 3.2: Compare entering classes

Exposure follows from the calendar; adoption, instruction and competency have to be measured. Figures are the nearest available survey for each class, not measurements of that class. Selected class: Fall .

Exposure vs. evidence

Fig. 4.1: Use cases, by survey

View source →

Augmentation: supports reasoning

Either: depends on task and purpose

Automation: AI does the intended work

Classification is this dashboard's, following the augmentation / automation distinction in the source reports. Percentages are not comparable across surveys.

Table 4.2: Student AI use taxonomy

CodeCategoryExamplesMode
AConceptual understandingExplaining, rephrasing, walking through problemsAugmentation
BReading & informationSummaries, themes, study notesAugmentation
CStudy & assessment prepPractice exams, flashcards, feedbackAugmentation
DResearch developmentQuestions, hypotheses, methodsEither
EWriting & communicationGrammar, outlines, sections, full essaysEither
FTechnical workDebugging, generating code, data analysisEither
GPersonal navigationCareers, résumés, institutional processesSupport
HAgentic workflowsMultistep, delegated task sequencesEmerging

Fig. 5.1: Drill down by dimension

↗

Fig. 5.2: Six layers of the AI divide

Each layer builds on the one below. A student can clear the first and still fall behind at the next, so read from the bottom up. The top layer is about benefit, not access. Tools can be shared equally while learning outcomes stay unequal.

I

Material access Can they get online and use the tools? Devices, connectivity, required tools. Tyton 2024.

About 1 in 4 limited in access to basic infrastructure such as reliable internet (Tyton 2024) ~25%

II

Capability access Do they get the stronger models? Free tiers have tighter limits and fewer features. 11.3% paid at Middlebury. Tyton 2024.

Regular student users who pay for AI tools 44%

III

Instructional access Have they been taught to use it well? Verification, privacy and disclosure. EDUCAUSE 2025.

Received institution-provided AI training 20%

IV

Disciplinary access Do they learn AI within their field? Field norms shape what students practise. Sciences, maths and CS students adopted at nearly twice the rate of literature students at Middlebury.

Academic AI adoption, literature vs. sciences (Middlebury) 49–91%

V

Human support Is a person there when AI falls short? A tutor, librarian, instructor or adviser. This is the layer AI is most likely to replace.

No reliable measure yet —

VI

Outcome equity Do they benefit equally? Learning, persistence and employment. Tutoring gains were largest for students with lower prior knowledge (IZA 2025), but no study yet tracks outcome gaps by group.

No reliable measure yet —

↑ Foundation at the bottom Top layer hardest to measure

Reading note: Higher adoption in a group does not establish greater access to advanced models, stronger literacy or better learning outcomes. First-generation status remains a major evidence gap and should not be inferred from race, income or school type.

Fig. 6.1: How much schoolwork gets chatbot help

US teens 13–17, n = 1,458. Source: Pew Research Center, fall 2025 (published Feb 2026).

Fig. 6.2: Do students understand school AI rules?

n = 1,017 teens. Only 52% said most teachers had similar rules. Source: Common Sense Media, Aug 2026.

Fig. 6.3: First move when stuck on schoolwork

Dark bars: AI-mediated routes, together ~22% of first help-seeking. Source: Common Sense Media, 2026.

College Board · 2025 ↗

79 → 84%

HS GenAI use for schoolwork, January to May 2025.

Details +

Common Sense · 2026 ↗

44% · 59%

had an AI tool blocked at school, among teens who use AI for schoolwork; of those, 59% switched to a personal device.

Details +

Common Sense · 2026 ↗

30%

of teens had ever discussed safe AI use with a teacher.

Details +

College Board · 2025 ↗

39%

of school or district administrators prohibited student GenAI use; 19% allowed it without a policy.

Details +

Fig. 7.1: AI use by role, year by year

Measure:

The threshold changes by wave (weekly, monthly, daily), so compare roles within a year, not bar lengths across years.

Source: Tyton Partners, Time for Class 2023–2026.

Fig. 7.2: Institutional support, student view

UK undergraduates. Source: HEPI 2026.

Fig. 7.3: Policy timeline

2022–23 Concern and uncertainty; bans and detection focus.

2023 Course-level restrictions; integrity debates.

2024 Institutional guidance, disclosure, faculty discretion.

2025 Policy clarification and assessment redesign.

2026 Competency development and institutional platforms.

Tyton 2026 ↗

32%

of administrators report an institution-wide AI policy.

Details +

Tyton 2024 → 2026 ↗

36 → 55%

of faculty cite cheating as a challenge.

Details +

HEPI 2026 ↗

65%

report assessment has changed substantially in response to AI.

Details +

Liang et al. 2023 ↗

61.3%

average detector false-positive rate on non-native English (TOEFL) essays.

Details +

Table 8.1: Four outcomes, measured separately

Better work with AI is not the same as learning. Each study below is sorted by what it actually measured. Only column D shows whether students can do the work on their own.

A Assisted performance Quality of work produced with AI Details +

B Speed Time taken to complete the task Details +

C Perceived learning How much students feel they understood or engaged Details +

D Independent competence Performance without AI, ideally after a delay Details +

StudyA · AssistedB · SpeedC · PerceivedD · Independent
Bastani et al. 2025 · RCT+48% general GPT; +127% tutor——−17% after general GPT; tutor largely avoided the loss
Kestin et al. 2025 · RCT—Less time on taskHigher engagement and motivationMore than 2× median gains on post-test
IZA 2025 · experiment———+0.23 SD exam score (check whether AI was available)
Middlebury · experiment———+0.27 SD knowledge test; more durable for augmentation
Chen & Cheung 2025 · meta-analysisMixed; not separated—g = 0.62 affective-motivationalNot separated from assisted results

— = not measured or not reported. Classification is this dashboard's reading of each study; check the originals.

Fig. 8.2: Effect sizes by outcome (Hedges' g)

0 0.5 1.0 1.5 2.0 2.5

Open marker = not statistically significant. 57 studies, 97 effect estimates. Source: Chen & Cheung, 2025 meta-analysis.

Fig. 8.3: Assisted gains vs. unaided performance

Change relative to control, ~1,000 high-school maths students in Turkey. Assisted practice improved; once AI was removed, the general-chatbot group did worse. Tutoring safeguards largely prevented the loss.

Source: Bastani et al., PNAS 2025.

RCT · Harvard physics · n=194 ↗

>2×

median learning gains from a pedagogically scaffolded AI tutor vs. active-learning class, in less time.

Details +

IZA · Dec 2025 · n=334 ↗

+0.23 SD

test performance from AI tutoring; strongest for low prior knowledge and strong self-regulation.

Details +

Middlebury experiment ↗

+0.27 SD

immediate knowledge scores; more durable for augmentation than automation users.

Details +

Reading note: Meta-analytic effects come mostly from structured interventions, not everyday chatbot use. Separate four outcomes: a better answer with AI, faster completion, feeling that one understands, and retained competence without assistance.

Fig. 9.1: The preparation gap

Each pair: what students believe they need (solid) against what they report receiving (hatched).

Fig. 9.2: AI and choice of major

1,433 associate and 2,368 bachelor's students. Source: Lumina Foundation–Gallup 2026.

Fig. 9.3: What employers prioritise

Sources: Digital Education Council 2025 (skills); NACE Job Outlook 2026, 183 employers (job requirements).

Fig. 10.1: What I do vs. what I think peers do

Self-reported

Believed of peers

US college students, weighted national sample n = 1,003. Source: Dartmouth AI Survey, Oct 2026.

Fig. 10.2: Concerns alongside use

Sources: Dartmouth 2026; Inside Higher Ed 2026.

Fig. 10.3: Net effect on the student experience

~15% use AI for companionship, advice or loneliness; reported effects on loneliness split almost evenly.

UK undergraduates. Source: HEPI 2026.

Fig. 10.4: Gen Z feelings about AI, year by year

US Gen Z aged 14–29, Gallup Panel; 2026 n ≈ 1,570. Source: Gallup, Walton Family Foundation & GSV Ventures.

Fig. 11.1: The student success continuum

Six stages from arrival to life after college. For each: what the evidence in this dashboard says, how strong it is, and what is missing.

Stages adapted from the student success continuum: preparation, transition, engagement, learning, persistence, completion. Strength ratings are this dashboard's judgment.

Fig. 12.1: Seven questions the evidence can't yet answer

Register 12.2: Evidence gaps

Next step

A longitudinal study of the Fall 2026 entering class

Most of these questions need the same students followed from matriculation through graduation and into work, measuring AI exposure, competency, independent learning, achievement, belonging, persistence and post-graduation outcomes. No such study exists in this evidence base.

Table 13.1: Evidence base

SourcePopulationStrengthLimitation
HEPI / Kortext 2024–26Full-time UK undergraduates, ~1,050/yrRepeated annual seriesExcludes part-time and graduate students
Tyton Partners 2023–26US students, faculty, admins; 750+ institutionsMulti-stakeholder, repeatedNot a panel; thresholds vary by wave
Lumina–Gallup 20263,800 US associate and bachelor's studentsLarge US sampleCross-sectional
Inside Higher Ed 2025–261,000–5,000 US studentsDirect student voiceFlash surveys
College Board 2025US high-school students, ~600 per wavePipeline and subgroupsUnweighted samples
Pew Research 2023–261,458 US teens (probability panel)Representative pipeline dataTeens, not college students
Middlebury 2025~20% of one liberal-arts collegeDetailed task and subgroup dataSingle selective institution
Dartmouth 20261,003 US national + 1,533 DartmouthIncoming-cohort dataInstitutional sample not representative
Experiments & meta-analysesSpecific courses and interventionsCausal mechanismsShort duration, narrow contexts

Table 13.2: How evidence is classified

Evidence strength depends on design, not on the publisher's prestige. Each reference is tagged with its design and with what kind of finding it supports: causal, association, measured, self-reported or synthesis.

DesignWhat it can showSources here

One row per statistic: source, sample, design, limitations and verification status. Open questions are on Open questions.

stories · newest first · links open the original publisher

Stories are journalism, opinion and institutional news, selected for relevance. They are not part of the dashboard's evidence base. Summaries are paraphrased; read the originals for full context.

QUESTION ONE

How has AI use changed since November 2022?

When ChatGPT went public on 30 November 2022, higher education reacted with alarm about cheating. I wanted to see what actually happened next: whether use kept growing, what students used it for, and how colleges responded.

02 Adoption over time04 Use cases07 Institutions

QUESTION TWO

What does the incoming class bring with them?

Students entering college in fall 2026 started ninth grade just weeks before ChatGPT launched. They are the first class to spend nearly all of high school with it. I wanted to know what habits, expectations and gaps they arrive with.

03 Cohorts06 High-school pipeline

The story in four years

2022–23 Surprise Students experiment; campuses debate bans and detection. Class of 2030 is in ninth grade.

2024 Habit Regular use becomes mainstream. Rules move to the course level. Class of 2030 meets AI in sophomore and junior year.

2025 Normal Near-universal use in the UK; tutoring and study prep lead. 84% of US high schoolers use it for schoolwork.

2026 Arrival The question shifts from whether students use AI to whether they can use it well. Class of 2030 arrives on campus.

Key takeaways

01

It's mostly about learning, not cheating. Explaining concepts, summarizing and brainstorming top almost every survey. Students say they ask AI questions they'd hesitate to ask a person.

02

Students think everyone else is cheating. Few admit to copying AI text, but half believe their peers do. That belief creates its own pressure to use it.

03

The incoming class is familiar, not fluent. Most taught themselves. Their high schools offered a patchwork of bans, blocks and unclear rules, and few teachers discussed how AI works or how to use it safely.

04

Students are more cautious than the hype suggests. They worry about their writing skills, about fairness and about jobs, and they still want teachers, peers and advisers alongside AI.

The question I'm left with

Almost every student uses AI now. What can they still do without it?

Not formal research

This site is a brief, informal overview. It is not meant to be in-depth or complete: it highlights selected findings rather than covering every study, population or perspective. It has not been peer reviewed, it does not collect new data, and it does not follow a preregistered systematic-review protocol. It brings together published industry, nonprofit and academic research on how students use generative AI, and presents it in one place for exploration.

Step 1

Gather industry research

Sources are existing surveys and studies from HEPI, Tyton Partners, Gallup, Pew, College Board, EDUCAUSE, Inside Higher Ed, Common Sense Media, Dartmouth and others, plus a small set of published experiments and meta-analyses.

Step 2

Analyze with AI tools

The research was collected and analyzed with three AI research assistants (Perplexity, Grok and ChatGPT), which produced five synthesis reports in November 2026.

Perplexity Grok ChatGPT

Step 3

Visualize and cross-check

Figures from those reports were charted here. Where reports disagreed, or a figure differed from the original publisher, the discrepancy is noted on the page. A retracted meta-analysis was removed.

How to use it responsibly

Verify before citing AI tools can misread or misattribute figures. Follow each source link to the original publication.

Don't merge series Surveys differ in population, wording and timing. Compare directions, not exact percentages.

Read the caveats Most evidence is self-reported and cross-sectional. Reading notes and pop-ups flag key limits.

Use for orientation Good for briefings and discussion. Don't base formal claims on it alone.

Evidence cutoff: 8 November 2026. The five source reports are listed under References → Source documents.

← Previous sheet

Title GenAI Generation, 2022–26

Evidence cutoff 08 Nov 2026

Next sheet →

Measure
Sample
Caveat
Evidence

View source

Survey series, by year

SeriesSource2023202420252026
Any AI useHEPI / Kortext Student Generative AI Surveyn/a66%92%95%
GenAI for assessed workHEPI / Kortext Student Generative AI Surveyn/a53%88%94%
AI text directly in assessed workHEPI / Kortext Student Generative AI Surveyn/a3%8%12%
Institution provides AI toolsHEPI / Kortext Student Generative AI Surveyn/a9%23%38%
Students, monthly or moreTyton Partners, Time for Class43%59%n/an/a
Students, weekly or dailyTyton Partners, Time for Class14%36%42%61%
Instructors, weekly or dailyTyton Partners, Time for Class4%n/a30%n/a
Teens, chatbot for schoolworkPew Research Center13%26%54%n/a