The Fourth R

The first work of The Fourth R

First Question

Learning that starts where you’re curious.

For most of us, school was a place to give the right answer. Our questions waited at the door. First Question opens it: learning that starts from what the learner wants to know, on solid foundations, with teachers who have time to teach.

Join the work See the fourteen problems
0102030405060708091011121314

Why we start here

I. We were taught to answer

A high-stakes exam rewards the answer you were told to give. Years of a young life narrow to one paper on one morning. When that paper leaks, the years leak with it. In 2026, students in India took to the streets over exactly that.1

Now a machine gives answers to anyone, instantly. Handed over freely, it lifted students’ practice scores and cut what they could later do on their own by 17%. Built to ask instead of tell, it largely stopped the harm.3

An answer is now the cheapest thing in the room. A good question is not.

  1. 1Kenya, June 2024, over a finance bill; Nepal, September 2025, over a social-media ban; India, 2026, over a leaked medical-entrance exam and, from September, over the revision of electoral rolls. The triggers differ; analysts point to shared grievances — few opportunities, and institutions seen as unresponsive.Al Jazeera, 16 June 2026 (opens in a new tab)Al Jazeera, 25 September 2026 (opens in a new tab)Council on Foreign Relations (opens in a new tab)Atlantic Council (opens in a new tab)
  2. 3In a field study of about 1,000 high-school students, unrestricted GPT-4 raised practice scores but lowered later unaided exam scores by 17% compared with students who never had it; a tutor designed to give hints, not answers, largely removed the harm. A 2025 Harvard trial found a carefully designed AI tutor beat an active-learning class. Both are small and early.Bastani et al., PNAS 2025 (opens in a new tab)Kestin et al., Scientific Reports 2025 (opens in a new tab)

II. Foundations first

Curiosity is no shortcut past the basics. In rural India, more than half of children in Class 5 still cannot read a Class 2 text.4 A child who can’t yet read can’t follow a question very far.

So the foundations come first, and we count who gains.

  1. 4ASER 2024 (Rural): 48.8% of children in Std V could read a Std II-level text — so 51.2% could not. ASER is run by ASER Centre, the research arm of Pratham.ASER 2024 report (opens in a new tab)

III. What we build first

One open prototype: a map of a single subject, and a tutor that answers a question with a better one. Small enough to finish. Open enough that anyone can check it.

The fourteen problems

Fourteen hard problems stand between us and a school built on curiosity. Bring what you know.

Understanding the learner

Knowing what someone knows, without turning them into a score.

  1. Knowing what someone knows, without examsEstimate what a learner understands, and what to explore next, from everyday work instead of tests.

    What makes it hard

    Standard knowledge-tracing models need pre-tagged questions and a fixed syllabus. Open-ended work is messy, AI help blurs whose understanding is being measured, and the model must be honest about its uncertainty — never a hidden score on a child.

    Start where others stopped

    I can help with this: Knowing what someone knows, without exams Link to this problem
  2. Evidence of real understandingTell understanding from outsourcing, and agree what knowing something well looks like — without bringing back the exam.

    What makes it hard

    AI-text detectors are unreliable and biased against non-native writers, so detection is out. The alternatives — a record of the process, an oral defence, live questioning — cost teacher time and can be unfair. Rubrics must be reliable enough to trust and light enough to survive a real classroom.

    Start where others stopped

    I can help with this: Evidence of real understanding Link to this problem
  3. A shared map of what connects to whatAn open, explorable map of ideas that a curious learner can wander — in more than one language.

    What makes it hard

    Experts disagree about prerequisites, and they differ from learner to learner. A map built by machine is full of errors; a map built by hand doesn’t scale. Who maintains it, and how it stays right as knowledge changes, is unsolved.

    Start where others stopped

    I can help with this: A shared map of what connects to what Link to this problem

Teaching well with AI

Tools that make people think harder, not less.

  1. Tutors that teach, not tellAI tutoring that asks, hints and lets a learner struggle productively, instead of handing over the answer.

    What makes it hard

    Knowing when to push and when to help is the craft of the best teachers. Language models tend to over-explain and to agree with whoever they’re talking to. The pedagogy has to be tuned by age, subject and language, and judged by learning, not engagement.

    Start where others stopped

    I can help with this: Tutors that teach, not tell Link to this problem
  2. Testing tutors for harmCheck, independently, whether an AI tutor is correct, safe, and building understanding rather than dependence.

    What makes it hard

    Practice scores can rise while later unaided learning falls. Accuracy, bias, flattery and long-term effects need ongoing studies that no vendor controls — and the results have to be published, the negative ones included.

    Start where others stopped

    I can help with this: Testing tutors for harm Link to this problem
  3. The teacher’s changed dayMake AI lighten a teacher’s work and sharpen their judgement — and find out whether teachers want it at all.

    What makes it hard

    Education technology has a long record of adding work. Teachers need training, trust and a say in the design. How many learners one person can truly mentor, and what to hand to software, has to be tested in real classrooms with the people who work in them.

    Start where others stopped

    I can help with this: The teacher’s changed day Link to this problem
  4. Motivation that doesn’t become rankingKeep learners wanting to go deeper without points, grades or leaderboards becoming the goal.

    What makes it hard

    Visible progress motivates, but external rewards can crowd out interest — and the evidence is mixed by age and context. The real test is the learner who isn’t curious yet: what sustains their effort?

    Start where others stopped

    I can help with this: Motivation that doesn’t become ranking Link to this problem

Reaching every learner

Where most learners actually are: cheap phones, many languages, no head start.

  1. Working offline, on cheap phonesMake it work on low-cost and shared devices and patchy connections — where most learners actually are.

    What makes it hard

    Good AI tutoring wants fast networks and expensive compute; models small enough to run offline are weaker. The cost per learner has to fall far enough for public systems to pay it, and the design has to work for a child with no device of their own.

    Start where others stopped

    I can help with this: Working offline, on cheap phones Link to this problem
  2. Every language, not just EnglishTeach and understand learners in their own languages and dialects, including where there is little data.

    What makes it hard

    AI quality drops sharply in languages with little written text, and everyday speech mixes languages freely. Errors fall hardest on the learners already least served. Content and evaluation need native speakers and local data.

    Start where others stopped

    I can help with this: Every language, not just English Link to this problem
  3. Curiosity that leaves no one behindMake sure self-directed learning helps children without strong foundations or adult support — not just the already-advantaged.

    What makes it hard

    Minimal-guidance teaching can widen gaps, and curiosity feeds on what you already know. Any design has to build foundations first, and measure who gains and who doesn’t.

    Start where others stopped

    I can help with this: Curiosity that leaves no one behind Link to this problem
  4. Children’s safety, privacy and consentProtect minors from data misuse, unsafe content and manipulation, with consent a child and a parent can actually use.

    What makes it hard

    Children can’t give meaningful consent, and parents differ in access and literacy. A record of how a child learns is rich enough to identify them. The rules differ by country, and the safeguards have to hold while the project is small, not just later.

    Start where others stopped

    I can help with this: Children’s safety, privacy and consent Link to this problem

Trust and money

What a learner can show for it, who keeps it, and who pays.

  1. A record the learner keepsLet learners hold their own record, show it’s genuine, and share one claim at a time.

    What makes it hard

    Proving whose work it is and deciding who holds the trace are one problem: a provable record nobody controls is surveillance; a controlled record nobody can trust is worthless. Custody is the hard part, not cryptography — lost keys, changing guardians, control passing to the learner at the right age.

    Start where others stopped

    I can help with this: A record the learner keeps Link to this problem
  2. A record the world trustsMake a learner-held record count for work and further study, without one gatekeeper deciding.

    What makes it hard

    The technology to issue credentials exists; trust and use are the bottleneck. Employers rely on institutions they know. A new kind of record needs reputable people to vouch for it, a clear meaning, and a route to real opportunity — apprenticeship, contribution, work.

    Start where others stopped

    I can help with this: A record the world trusts Link to this problem
  3. Funding that stays public and accountableFund learning in the open, show what the money achieved, and resist capture and fake accounts.

    What makes it hard

    Ledgers show where money went, not what it bought; outcomes need trusted checks. Small-donor funding has to withstand fake identities without surveilling anyone. Paying for results can distort what gets taught. And tutoring at scale has running costs someone must meet for decades.

    Start where others stopped

    I can help with this: Funding that stays public and accountable Link to this problem

If you believe this and can help build it, join the work. Count me in

Your best skills, put to use beyond just making money.

What we hold to

  1. The record belongs to the learner.

    And to a guardian while they’re young. Not to a school, and not to a company.

  2. Open to research.

    With the learner’s consent, and a guardian’s while they’re young, anonymised learning data is open to approved researchers, who can test their own ideas about what works on the platform itself. We publish what fails along with what works.

  3. Equity, by name.

    Gender, caste, class, disability, language and place, and where they meet. We measure who gains by each, and build first for the learners the current system leaves furthest behind.

  4. Built with the people who use it.

    Every prototype is tried with learners, of any age, before we call it working. We publish who took part and what changed because of them.

  5. Every rupee traced to what it did.

    Funders may stay anonymous. What their money achieved may not.

  6. Held in common.

    Open source, and owned by no one. There’s no organisation behind this yet. We’ll choose its form in the open, and it will be one that can’t be sold.

Join the work

Count me in

If you believe this and can help build it, join the work.

Your best skills, put to use beyond just making money. We’re looking for people who build software, design learning, teach, research, fund, tell the story, or open doors to schools and government.

Ashwin Manghat

Written by · BengaluruOctober 2026

There’s no organisation behind this yet. For now, “we” is me. That’s what the list is for. hello@thefourthr.now

Tell us what you’re good at and where you’d start. A person reads every word, and we put you where your skills count most.

How you’d help

What you do

Pick any that fit.

Which problem could you help with?

Optional: one or two, if one is yours. Funders and storytellers often help across all of them.

A project, a class you taught, a study, a fund you ran. A few honest lines tell us more than a CV.
Optional. Code, a paper, a lesson plan, a portfolio — it saves you describing it.
Time you can really give

Why we ask: so we never hand you more than you offered.

Who you are

Never shown publicly.

We keep what you tell us so we can build a team with you. Nothing is sold, ever. What we keep and why.