About this report
This report is an evidence digest. It summarises findings published by researchers and institutions in 2025 and 2026 on three questions: does AI tutoring improve learning, does AI reduce teacher workload, and who is actually using these tools. Zunkiree Labs did not conduct any of the surveys or trials. We read the published sources, extracted the figures, and added notes on what each can and cannot support.
Where we offer a view of our own, it is marked as ours. A fuller narrative is in our ebook, AI in Education: What the Trials and Surveys Show.
Key findings
- Tutor design decides the outcome. A Harvard randomized crossover trial (194 students analysed) found a carefully designed AI tutor produced higher post-test scores than an active-learning lesson, with a median 4.5 against 3.5. A PNAS field experiment with nearly 1,000 students found students who practised with a standard chatbot interface scored 17% lower on a later exam without it than students who never had access.
- Practice performance can mislead. The OECD reports that AI access improved short-term performance in the Turkish experiment by 48% with the standard interface and 127% with a tutoring version, yet students performed 17% worse once access was removed.
- Teachers report real time savings. Weekly AI users in a Gallup survey of 2,232 US public school teachers estimate saving 5.9 hours a week. The figure is self-reported and applies to the 32% who use AI at least weekly.
- Use is uneven. About half of US adults use AI chatbots according to Pew, while Microsoft estimates 18.8% of people aged 15 to 64 worldwide used a generative AI product in Q2 2026, with 28.8% in the Global North and 16.2% in the Global South.
- Integrity is a live concern. The OECD reports that seven in ten teachers on average believe AI could let students present others’ work as their own.
Finding 1. AI tutoring: promising under specific conditions
| Source | Design | Reported result | Main limit |
|---|---|---|---|
| Kestin et al., Scientific Reports, 2025 | Randomized crossover; 194 analysed of 233 enrolled; two topics, two weeks | Median post-test 4.5 vs 3.5; estimated effect about 0.63 standard deviations; median time with tutor 49 minutes; engagement rated 4.1 vs 3.6 on a five-point scale | One course, short topics, high-achieving students, expert-written prompts |
| Bastani et al., PNAS, 2025 | Field experiment, nearly 1,000 high school math students | 17% lower grades after removal of unguarded access; guardrailed tutor largely mitigated the harm | Short period; one subject and country |
| World Bank, January 2025 | Six-week after-school pilot, Benin City, Nigeria, with teacher support | About 0.3 standard deviations of learning gain reported by the authors | Authors’ own summary; no sample sizes given; not peer-reviewed |
The Harvard authors state that they do not presume structured AI tutoring will always beat in-class active learning, for example where complex synthesis and higher-order critical thinking are needed. They list conditions that may matter, including a heterogeneous class, a capable model, expert-written prompts and a carefully structured framework.
Finding 2. Teacher workload: savings reported, student impact unmeasured
The Gallup report (June 24, 2025) surveyed 2,232 US public K-12 teachers between March 18 and April 11, 2025, using the RAND American Teacher Panel. Gallup reports a margin of error of plus or minus 2.5 percentage points.
| Measure | Reported figure |
|---|---|
| Teachers who used AI for work | 60% (32% weekly or more, 28% monthly or less) |
| Estimated weekly time saved, weekly users | 5.9 hours, about six weeks over a 37.4-week school year |
| Top uses (monthly or more) | Preparing to teach 37%, worksheets and activities 33%, modifying materials 28% |
| Teachers saying quality improved at least somewhat | 57% (grading and feedback) to 74% (administrative work) |
| Teachers saying quality got worse | Fewer than 16% |
The OECD adds that in its 2024 TALIS survey 37% of teachers used generative AI for work tasks, and it cites a 31% reduction in lesson and resource planning time for secondary science teachers in England and a 9 percentage-point increase in student pass rates when low-experience tutors used AI support.
None of these figures show that students learned more.
Finding 3. Adoption: widely used, unevenly spread
Pew Research Center surveyed 5,119 US adults in February 2026: about half use AI chatbots, roughly one in four daily, and 38% of employed adults use them for work tasks. Microsoft estimates 18.8% worldwide, and says the Global North and South gap is widening. The Stanford 2026 AI Index reports 53% population adoption of generative AI within three years and a strong link between adoption and GDP per capita.
These numbers use different definitions and populations, so they should not be placed side by side. We found no Nepal-specific figure in them.
What the evidence cannot tell you
- Whether gains last. The OECD notes most evidence is short-term; the World Bank team lists long-term effects as unknown.
- Whether results travel. No cited trial or survey was run in Nepal.
- Whether teacher time savings reach students. Gallup notes teachers reinvest saved time, but the survey does not test the effect on learning.
- Whether quality ratings are accurate. They are teachers’ own judgements.
- How privacy and equity play out. The tutoring trials do not examine student data handling or who lacks access.
This report should not be read as showing that AI improves learning in general. It shows that some designs helped, some harmed, and that measurement matters.
What to do with this
Our own suggestions, not findings from the studies:
- If you run a school or training program, start with a teacher-facing task and a recorded baseline. Our step-by-step guide to introducing AI safely sets out a pilot plan and checklist.
- If you assess an AI tutor, ask whether it gives hints or answers, what it was compared against, and what happens to student performance when it is switched off.
- If you handle student inquiries at an education consultancy, the lower-risk pattern is fast replies on maintained information with a handoff to a counsellor. See our whitepaper on student inquiries.
- Plan for uneven access and skills among students and staff, per the adoption data above.
To discuss a pilot, schedule a call or read about our custom AI agent development.
Sources
- Kestin et al., AI tutoring outperforms in-class active learning: an RCT introducing a novel research-based design in an authentic educational setting, Scientific Reports, June 2025
- Bastani et al., Generative AI without guardrails can harm learning: Evidence from high school mathematics, PNAS, vol. 122, no. 26, 2025
- World Bank Blogs, From chalkboards to chatbots: Transforming learning in Nigeria, one prompt at a time, January 9, 2025
- Gallup and Walton Family Foundation, report on teacher AI use and time savings, June 24, 2025
- OECD, Digital Education Outlook 2026: Exploring effective uses of generative AI in education, February 3, 2026
- Pew Research Center, Americans and AI 2026, June 17, 2026
- Microsoft On the Issues, The continued state of global AI diffusion in 2026, September 21, 2026
- Stanford HAI, 2026 AI Index Report, 2026
Frequently asked questions
Is this original research?
No. This is a digest of published studies and surveys. Zunkiree Labs did not run any of the surveys or trials cited, and every figure is attributed to its source in the Sources section.
What is the main finding?
The design of the AI tool decides the outcome. A carefully designed tutor beat an active-learning lesson in one Harvard trial, while unguarded chatbot access led to lower later scores in a PNAS field experiment.
How much time do teachers say AI saves?
In a Gallup survey of 2,232 US public school teachers, weekly AI users estimated 5.9 hours saved per week. The 5.9-hour figure applies to the 32% who use AI at least weekly and is self-reported.
Do these results apply to Nepal?
Not directly. The trials were run in the United States, Turkey and Nigeria, and the teacher survey covers US teachers. We did not find comparable Nepal-specific studies in these sources, so local pilots are needed.
What you get:
- Key findings from tutoring trials, a teacher survey and an OECD review
- Source-by-source tables with design, result and main limit
- Teacher time-saving figures with the Gallup methodology
- A clear list of what the evidence cannot tell you
- Labelled suggestions for what to do next