PrepGraph Council for Responsible AI & School Innovation – India
AI arrived in Indian schools through children's phones — not through a circular, a staff meeting or a budget line. The Council exists to write down what good practice looks like, to test the claims schools are being sold, and to build Indian evidence on what this technology actually does to learning.
The Council is being constituted. It has no members, has held no events and has published no findings. Everything described as planned is an intention, recorded so it can be held against us.
The situation
AI arrived in Indian schools through children's phones. It did not arrive through a circular, a board directive, a staff meeting or a budget line. By the time most schools formed a view of it, it was already in use: in holiday homework, in project files, in Class 10 answer practice, in the essay that reads a little too evenly. The technology entered through the student, and the institution is still catching up with a decision it never made.
This is not a crisis. It is harder to correct than a crisis, because it is an unattended default. Teachers improvise, and improvise differently. One section's teacher treats any AI use as cheating; the teacher of the next section builds a worksheet with it on Sunday and does not mention it in the staff room. Principals sit through demonstrations they have no means to test, are quoted improvement figures measured in classrooms unlike theirs, and are asked to decide in a term when the timetable, the board results and the admissions cycle all have a prior claim on their attention. Parents, who will in time hold the strongest opinions of all, have been told almost nothing by anyone.
Beneath this sit a vocabulary problem and an evidence problem. Adaptive, personalised, AI tutor and learning outcomes mean whatever the speaker intends them to mean, so two principals comparing notes are often not comparing the same thing. And little of the evidence in circulation was gathered in an Indian classroom: not in a section of fifty-five, not under board examination pressure, not where the device is shared, the medium of instruction is the second language at home, and the syllabus must still be finished by January. There is no agreed floor. No minimum practice a parent can expect from any school, and nothing a principal can point to and say, this is what we do here, and this is why.
One caution about all of it. The Council has collected no data and has no findings, and nothing in this section is one. It is the working account the Council starts from, assembled from what the convener's staff have been told in schools rather than from any study, and the first year of research exists to test it rather than to confirm it. The Council asks schools to demand a denominator behind any number put in front of them; it can hardly exempt its own opening argument from that.
In the staff room
Two teachers of the same subject hold opposite rules on AI use in submitted work, and neither has been told which rule is the school's.
In the correction pile
A teacher is fairly certain a submission was machine-written, has no way to establish it, and no policy that says what happens next.
In the principal's office
A vendor quotes an improvement figure. There is no practical way to ask where it was measured, on whom, over how long, and against what comparison.
At the school gate
A parent asks whether the school uses AI. An honest answer takes ten minutes, and the school does not have one written down.
In the staff development plan
AI training, where it happened at all, was a single session on writing prompts. Nothing on assessment integrity, data, age-appropriateness, or when not to use it.
The argument
Three kinds of body already speak to schools about AI, and each is structurally unable to do this particular job. A vendor may publish a standard, but a standard written by a vendor is in the end a description of its own product, and schools discount it correctly. A circular can set a rule, but it arrives after the practice has already formed, applies uniformly to unlike schools, and states what is prohibited rather than what is good; it also cannot be argued with, which is what a school most needs to do while it is still deciding. A training programme can change one teacher, which matters, but it leaves no institutional artefact behind: when the teacher transfers, the capability transfers with her.
A convening body can do four things that none of those can. Hold a room in which principals say to one another what none of them will say from a stage. Settle vocabulary, so that a claim becomes comparable between two schools. Write a floor of practice with no commercial interest in the answer, and place the reasoning in public where it can be attacked. And frame research questions that no single school can afford to answer alone and no vendor can be trusted to mark its own answer to.
The argument for doing this now is not momentum. It is that defaults harden. The AI policies Indian schools write this session will be copied, adapted and inherited for years, and most will be written in a hurry, after an incident or before an inspection. A floor of practice will exist within a few years either way. The open question is whether it is set by people who have run a school, or by procurement pressure and the first serious incident involving a child.
Areas of work
They are not themes for a conference programme. Each exists because a school has a decision to make under it, and each is written to end in the question the decision actually turns on.
Privacy, bias, transparency, age-appropriate use, human oversight, academic integrity, child safety — and the discipline of deciding what not to deploy at all.
What must be in place before a single child uses an AI tool in this school’s name?
What a teacher should understand about how these systems work and where they fail, and what productive use looks like in preparation, feedback and assessment design.
What should a teacher here be able to do with AI by the end of this year, and how would we know that she can?
Not the ability to prompt a chatbot, which students already have and were never taught. The ability to judge: to notice when an answer is fluent and wrong.
What must a student be able to judge about a machine’s answer, and at which stage do we teach it?
Learning gaps found early, formative assessment that informs teaching, differentiated work, personalised practice — where they are genuinely better rather than merely newer.
Which teaching problem here is worth pointing this technology at first, and what result would count as a real improvement?
Adoption is a governance decision before it is an academic one. Most of governance is the capacity to say no with reasons, in a season when saying yes is easier.
What should this school adopt, what should it refuse, and what must be answered before anything is signed?
An Indian evidence base, gathered under Indian conditions, published with its method attached. Negative and null results are part of the record.
Is there Indian classroom evidence for the claim in front of me, and if not, what would it take to produce it?
Pillar 6 · Research
A company that sells to schools is convening the forum in which schools decide what to buy; that conflict cannot be dissolved, only written down and constrained.
Most of what Indian schools are told about AI and learning is enthusiasm with a number attached. The number is often real and almost always measured somewhere else: a different class size, a different language of instruction, a different assessment culture, a different level of device access, frequently a different country, and usually by a party with a commercial interest in the outcome. A principal cannot responsibly act on that, and is right to hold back.
The Council adopts the opposite discipline as a rule of publication, so that it can be checked rather than believed. Research questions and the measure of success are published before data are collected, with the date visible, so that a flattering question cannot be chosen afterwards. Method is published with the finding, in enough detail for a reader to disagree with it. Sample size, comparison group and effect size are reported plainly, including when they are small. Studies are conducted in Indian classrooms with the real constraints named rather than controlled away. Research involving children is held to explicit consent, data minimisation and handling consistent with the Digital Personal Data Protection Act 2023, with the school's own ethical judgement treated as binding and its withdrawal effective at any point. Everything is published free to any school, member or not.
This includes the inconvenient results. A finding that a tool made no measurable difference, or a difference too small to justify its cost, its risk or the teacher time it consumed, is published on the same terms as a favourable one and counted in the same year-end list. That is the purpose of the strand, not an exception to it. A body that publishes only encouraging findings has told you what it is.
One conflict is stated rather than managed quietly. PrepGraph may act as a research partner and may contribute data, funding or platform access to a study. Where it does, the participation is disclosed at the head of the published work, not in a footnote, together with what PrepGraph supplied and what it could and could not see; and the method must be reviewed before publication by at least one member who has no relationship with PrepGraph, that reviewer being named in the publication. PrepGraph's own product evidence is not a Council finding and will not be presented as one.
Year one
None of the following exists. Eight instruments do: the Charter, the Independence Policy, the Membership Terms, the Office of the Founding Chair, the Responsible AI Principles, the School AI Readiness Framework, the Research Agenda and Ethics, and the Programme Calendar are all issued today as founding drafts under Article 17 of the Charter, open for comment and to be reaffirmed, amended or withdrawn by the voting members once seven are in office. They are listed, with that status against each, on the register of publications rather than here. What follows is the work that has not been written. Each is a first-term intention, set down so that it can be held against the Council when the term ends. Drafts will be circulated to members for correction and dissent before publication, and finished instruments will be published free to any school, whether or not it has any connection to the Council, in a form meant to be edited by the school rather than adopted unchanged.
On or before 31 August 2027 the Council will publish a single page listing each item below with one of three statuses: published, in draft, or abandoned, with the reason. An item that is not delivered will be recorded as not delivered rather than quietly re-scoped or carried forward without comment.
A school policy in adoptable form: permitted use, prohibited use, disclosure of AI assistance in submitted work, staff use, data rules, and a review clause with a date on it. Written to be amended by a school, not signed as it stands.
Planned. Not yet drafted.
The questions a school should be able to ask before signing anything: where the claim was measured and on whom, what data leaves the school and where it goes, what happens to that data at exit, who is accountable when an output is wrong, and what the tool costs in teacher time as well as in fees. It names no products.
Planned. Not yet drafted.
What a teacher should be able to do, described in observable levels, so that a school can plan development and recognise progress instead of counting attendance at a workshop.
Planned. Not yet drafted.
Age-appropriate expectations for middle and senior school, framed around judgement, verification and academic honesty rather than tool operation.
Planned. Not yet drafted.
A plain-language note a school can adapt and send home: what the school does with AI, what it does not do, what it asks of families, and whom a parent should contact with a concern.
Planned. Not yet drafted.
A discussion paper on AI readiness in Indian schools, distinguishing plainly between what has been observed, what is inferred and what remains unknown. Where it has no evidence, it will say so.
Planned for the end of the inaugural term.
A small number of closed-door roundtables for school leaders, with a public note of what was concluded, and public sessions in which drafts are argued rather than presented.
Planned. None held to date.
The conflict
PrepGraph is an Indian education technology company that sells to schools. It founded the Council, funds it, staffs its secretariat and administers its work. The Council carries PrepGraph's name because concealing the connection would be worse than declaring it, but a disclosure settles nothing on its own; it only makes the conflict visible. The conflict is real: a company that sells to schools is convening the forum in which schools decide what to buy. What follows are the constraints that contain it. Each is a rule with a mechanism, not an assurance, and each can be tested by an outsider without the Council's cooperation.
Approval sits with the Chair, not the funder
The secretariat drafts. Nothing is published under the Council's name without the Founding Chair's written approval, and PrepGraph cannot publish, amend or withdraw a Council document on its own.
Check: every published document carries its approval date.
The funder gets one pass over a draft, and it is logged
PrepGraph has no editorial review. It receives one written factual-accuracy pass, limited to statements about its own systems, of ten working days. Silence publishes. Every comment it makes and every change that follows is logged and published with the output. It cannot delay, amend or withhold anything. Anyone who says the funder never sees a draft is describing something other than this arrangement: the secretariat that drafts is staffed by PrepGraph, and pretending otherwise would be the first thing a reader caught.
Check: the log of PrepGraph's comments, and of the changes that followed, published with each output. An output carrying no log at all is the thing to ask about.
No product is named
No Council instrument or guidance names a commercial product, PrepGraph included, outside a research study of which that product is the object.
Check: search any Council document for a product name.
One funder, and no vendor money at all
The Council takes no sponsorship, advertising, exhibition, stall or speaking payment from any technology company, including firms competing with PrepGraph. PrepGraph's annual contribution is published as a figure.
Check: the figure in the year-end statement, or its absence.
The Chair's independence is a condition of holding the chair
The office is held on the condition that its holder has no shareholding, employment, commission, consultancy or procurement relationship with PrepGraph and takes no fee for the role. The condition is evidenced by a declaration of interests, published in the register before the first meeting and dated. Until that declaration is published, a reader should treat the condition as stated and not yet evidenced, because a claim about a person's financial affairs that nobody can check is worth what any such claim is worth. If the condition ceases to be met, the Chair stands down and the reason is published.
Check: whether the declaration is in the register, and on what date it was published.
Dissent is printed, not managed
Any member may require that a dissenting note of up to three hundred words appear in the published document, under their name if they wish, unedited except for length and law. The control is real but it is not yet operative: until the first members accept, there is nobody to exercise it.
Check: a year of instruments carrying no dissent at all is evidence of a problem, not of agreement.
A stated failure condition
If the Council is used to recommend a purchase, it has failed at the thing it was made for. Members are asked to say so publicly, and the Council will carry any such statement on its own page, unedited.
Check: whether the statement appears, and how quickly.
Limits accepted at founding
Each of the following is a limit the Council accepts on itself at founding, and each can be checked by reading its published work rather than by trusting its intentions. Where a limit is breached, the evidence will be a document, and the document will be public.
The neutrality position is adopted as a standing rule and will appear in the same words on every Council document: "The Council does not endorse individual procurement decisions or technology vendors. Its role is to advance responsible AI readiness, educator capability and evidence-based innovation in schools."
An invitation
If you lead a school, what the Council asks of you is your judgement and a modest amount of your time. Not a purchase. Not a pilot. Not access to your students or their data. Not your name against a claim you have not tested, and not a public position you would be uncomfortable defending in your own staff room.
