Project Ascend — Opportunity through Education
Project Ascend
{{ item.label }}

The gap isn't access to knowledge. It's access to teaching.

Education is due its next chapter, and we think it starts in the schools with the least. Project Ascend develops classroom pedagogy that uses AI to deepen student enquiry rather than replace it. We work with teachers to put it into practice, researching our implementation so that what reaches a classroom is what actually works.

Our pedagogy

The Student

We put AI learning in the hands of every student, so they can ask questions and explore a subject without the limitations they are normally confined to. The AI meets them where they are. A student who loves sport can learn momentum through football, engaging students who struggle to concentrate in a classroom.

The Teacher

Teachers are the heart of our pedagogy, not a step in it. We train them to adopt AI so they can support students long after we've gone. There is one path to academic growth here and it starts with them: teacher adoption, then student proficiency, then student growth.

The Safeguards

AI should make students think more, not less. Our platform won't answer for a student. It asks them to show they understand before moving on. Teachers can see how their students are using it, and they know when to step in.

The Change

We've seen teachers build their own routines and become ready to train other teachers themselves. Students have become more active in their learning, asking more questions inside and outside of AI classes, and telling us they feel more confident in their own ability to learn. We measured all of it in our paper.

"AI gives me the courage to ask questions I'm afraid to ask in class."

{{ fig0 }}

students

{{ fig1 }}

weeks

{{ fig2 }}

felt more comfortable asking questions

{{ fig3 }}

introduced their own physics ideas into conversation

{{ fig4 }}

litres of water used across the pilot

Key findings

Teachers went beyond adoption

Both teachers began by treating the platform as something handed to them. Eight weeks later they needed no support from us, had built their own classroom routines, and said they were ready to train other teachers. Resources we gave directly to students failed without a teacher to interpret them. The teacher is the mechanism.

Students became more active in their learning

Students told us they felt more comfortable asking the AI questions, the strongest response in our whole study. Teachers saw the same habit spreading into their non-AI classes. That curiosity showed in how students used the tutor, by the final lesson, 34 of 48 had used it to test ideas of their own.

We detected and measured new avoidance techniques

We classified all 5,264 student messages, so we saw students getting better at avoiding work as well as doing it. The fastest-growing method was handing the tutor's own question back to it. Well formed, on topic, and difficult for AI to notice. Measurement caught it. That tells us what to build next.

Read the pilot in full

Publication

{{ paperTitle }}

{{ paperMeta }}

{{ paperCta }}

Where we are

{{ stage.label }}

{{ stage.tag }}

{{ stage.note }}

Core principles

Open Access

The platform and its integration documentation stay open source for non-commercial use.

Data Privacy

We never sell or trade student data. It is collected only to improve the platform, support schools and measure impact, and it is held securely.

Unconditional Inclusion

We do not limit ourselves to schools that already have technology. Democratising education is the goal, which means providing resources to enable AI learning is included.

Holistic Education

Curriculum we build will cover mental wellbeing and ethics, teaching students to care for themselves and to do good through understanding rather than compliance or social pressure.

Sustainable Autonomy

Every integration is built to be owned and run by the school itself. A key factor of success is whether a school can operate independently after we leave.

Environmental Awareness

We measure and report the environmental cost of what we build, and let it shape design decisions as we scale.

Follow the work as it develops.

Social

Contact

{{ homeStatus }}

Mission

Opportunity through Education.

The problem, the conditions that have to change, and how we change them.

Two Problems, One Classroom

A teacher with fifty students and no time to spare teaches for the exam, because that is what they are judged on. Content delivered, content memorised, content tested. Students learn to reproduce answers well enough to pass, and never find out what the answers mean. Nobody in that class is asking why, because the environment isn't built for it.

The textbook as the sole resource compounds the issue. A student who doesn't learn well from a page has nowhere else to go. No second explanation, no different example, nothing that connects the subject to anything they already care about. They fall behind quietly, and the class moves on.

These two problems feed each other. A teacher stretched thin leans harder on the textbook, and a classroom built around the textbook gives students nothing to be curious about. Which is why the obvious fixes don't hold. Better books don't cultivate curiosity if nobody has capacity to teach beyond them. More training doesn't help if the only channel still fits a fraction of the class.

Both have to change at once.

Conditions for Change

Teachers need time freed up to individually support students. They need a method for supporting enquiry, not just encouragement to try. Students need somewhere their questions actually get answered, and content that meets them where they are.

And everyone needs to see it working, because a teacher judged on exam results will not adopt something they cannot verify, and a student won't put in effort to learn without evidence they're getting better.

How We Meet Them

Time to support individual students → The AI tutor takes on the individual questions and explanations a teacher of fifty cannot reach, which gives the teacher their time back for mentorship, class dynamics, and guiding how students use the AI.

A method for supporting enquiry → We provide teachers with training, routines and a framework to reliably determine where a student is at and how to support them in becoming more active in their learning.

Questions answered → The AI tutor, low stakes and always available.

Content that meets them → The AI adapts to how a student learns, keeping delivery to the examples and methods that suit them best.

Everyone sees it working → We provide teachers with summaries and analytics that show how their students are learning, and students test their conceptual knowledge on end of lesson tests, showing both themselves and their teachers that their learning is improving.

Pilot

Eight weeks in a Cambodian classroom.

A summary of our first pilot, run with a partner NGO at their school in Phnom Penh. The full paper, with method, statistics and appendices, will be published here.

48

students across two classes

8

weekly lessons, 90 minutes each

5,264

student messages classified

70.8%

tested an idea of their own with the tutor

0.4%

of messages checked the tutor's reasoning

Why Cambodia

Cambodia today puts education as a key part of their national policy. Around 95% of children complete primary school, and the ministry has made digital and technology-led learning an explicit priority. That is an ambitious position for a system that had to be rebuilt from almost nothing. 

The rebuilding began in 1979. Under the Khmer Rouge the educated class had been the target, and the country came out of it having lost 75% of its teachers and 96% of its university students.

Cambodia has come a long way since, but still has further to go. Constrained by teacher subject knowledge and a passive outlook to learning, most students fall to rote memorization for tests over conceptual understanding in subjects. In PISA 2022, 10% of Cambodian 15-year-olds reached Level 2 proficiency in science, against an OECD average of 76%.

That combination is what makes it the right place to test AI tutoring: a system already reaching for technology, high ambition for education, and a specific gap where subject knowledge runs out. An AI tutor can answer there, at no social cost to the student asking. It is also where the risk is sharpest, because introduced badly the same tool widens the gap it came to close. So we approach AI learning with research, ensuring what's integrated is what works.

SOURCES: CLAYTON 1998 VIA WORLD BANK; OECD 2023 (PISA); MOEYS.

How it ran

Two Grade 10 classes used a custom AI tutor to supplement their physics curriculum in their weekly two-hour STEM block, with a third class as a control. Eight lessons over two months. The tutor refuses answers and requires a student to show understanding throughout the lesson.

The school is urban and well resourced, so this is not a typical Cambodian classroom, but two conditions carried across. Students are accustomed to receiving content rather than interrogating it, and the two teachers running the sessions specialise in chemistry and biology, not physics. For eight weeks the subject knowledge in the room sat with the AI, which is where a great many Cambodian classrooms already are.

Teachers were trained before and throughout, and given the enquiry framework below with a set of routines: one concrete intervention for each step up the scale.

The Socratic Enquiry Level

A 0 to 5 scale for how a student is engaging with the tutor in a single message. Extractable from logs, and usable by a teacher mid-lesson.

00

Non-participation

Off task or avoiding the work. Flagged as social, punt, refusal or echo.

01

Passive-Consumption

Responding to what is asked, without moving beyond it.

02

Information-Requesting

Asking for content or explanation. The dominant interaction throughout.

03

Meaning-Shaping

Directing how the tutor teaches: a different example, level or approach.

04

Reason-Building

Bringing a hypothesis or a connection to work through.

05

Critical-Verifying

Checking the tutor's reasoning against their own.

CLASSIFIED BY LLM, VALIDATED AGAINST BLIND EXPERT HUMAN CODING. COHEN'S KAPPA = 0.91.

What we found

Teachers took ownership. Both began passively and by the final weeks needed no support beyond technical, having built routines of their own: covering the tutor's messages to test what a student retained, reading conversation logs during breaks. Both felt comfortable being able to train other teachers.

Students became more active. Students interacting with the tutor at SEL 4 rose from 12% to 32% of the class, and 71% introduced their own physics ideas at least once through the pilot. The most strongly agreed statement in the study was feeling comfortable asking the AI questions they would not ask in class, at 4.10 out of 5.

Avoidance grew in parallel, and did not plateau. Echoing the tutor's own question back at it rose from 3.5% of messages to 9.7%, still climbing in the final lesson. Echo is the hard case: the text is a well formed, on topic question the tutor just wrote, so it passes any check on a message in isolation. It replaced lower order SEL levels, not the higher-level use. Teachers separately caught students using other tabs and devices during quizzes.

Students viewed their subject differently. Belief that a difficult topic could be understood later rose 0.47, and thinking about physics outside class 0.46. Over the same period students became less likely to call physics easy, down 0.33. They did not just get more positive, they got more accurate about the work involved.

Test scores did not move. There were no significant shifts in test scores including the control. A school break forced a mid-pilot snapshot after three AI lessons, so it is exploratory at best.

Future research

Eight weeks at a single school more resourced than most of Cambodia can only be seen as suggestive data to support a larger study, by running a larger pilot we can reinforce and verify these observations as well as create new conclusions.

An expanded pilot could give us:

  • →Verification of initial pilot findings
  • →A refined SEL framework
  • →Whether AI tutoring improves test scores
  • →Whether AI tutoring supports rural and public schools
  • →How avoidance techniques can be managed
  • →Long-term shifts in how students learn once the tutor is withdrawn

Publication

{{ paperTitle }}

{{ paperMeta }}

{{ paperCta }}

About

Who We Are

The team behind Project Ascend.

Team

Steven Watson

Project Lead

Ben Watson

Project Advisor

Current status

We have finished our first pilot in Cambodia and written it up as a paper.

We are now planning the next stage: a larger pilot across multiple schools.

Support

Ways to back the work.

Segmented routes for different supporter types. [ALL CONTENT PENDING]

{{ route.title }}

[COPY PENDING]

[ACTION PENDING]

Contact

Get in touch.

Schools, funders, researchers and teachers curious about the pedagogy: we would like to hear from you. Email us directly, or use the form and we will reply as soon as we can.

Form

{{ contactStatus }}