Responsible AI Principles for Indian Schools

Why it matters

The published terms of use of the general-purpose assistants most often found in schools commonly set a minimum age of thirteen, or eighteen without parental consent. The Council has surveyed no such terms; that sentence is a working assumption and not a finding, which is why the commitments below require the school to read the terms of the tool in front of it and record what they actually say. When a school sets homework that can only be done by asking a chatbot, a nine-year-old opens an unfiltered adult product on a parent's phone at night with nobody watching; the school did not intend that, but the school caused it. A product demonstrated to adults in a staffroom is not the same product when a curious or frightened child is at the keyboard.

The school commits to

1. 1 The school keeps one dated list of AI tools permitted for each class band and tools not permitted at all. It is signed by the principal, displayed in the staffroom, sent to parents, and reviewed at the start of each academic year.
2. 2 Against each tool the list records the minimum age stated in that tool's own terms of use, copied from the terms, with the date the terms were read. A salesperson's assurance is not recorded as the answer.
3. 3 No homework or classwork requires a student to use a tool whose own terms set a minimum age above the age of that class. In primary classes any AI tool is operated by the teacher on a shared screen.
4. 4 Before first use with students, once each term after that, and before any new class band is given access, two members of staff put a fixed written set of questions about self-harm, violence, sexual content and personal advice to the tool and file what it returned, dated and signed. The trigger is the term and the new class band, not a version number, because these products change without telling a school. One of the two is the person with pastoral or counselling responsibility where the school has one. The prompt set is written down once so that the test is repeatable, and the file is held by the named accountable adult, not on a shared drive, and opened only by that person and the principal. Where a response would require a referral, the school's child protection procedure runs at once; the duty to report under section 19 of the Protection of Children from Sexual Offences Act, 2012 is a personal duty on the adult who apprehends the offence, and filing this test discharges nothing. An empty file means the tool has not been tested.
5. 5 One named person receives reports of harmful or inappropriate output from any student, parent or teacher. Each report is logged with its date, and what was done about it is written next to it.
6. 6 An AI feature switched on inside software the school already licenses is a new deployment. It enters the permitted list before it is used, or it is turned off in the administrator console until it does. Once a year the school writes to each existing supplier, including the ERP, the office suite, the video-conferencing tool and the library and assessment systems, asking which AI features are on by default on the school's account, which of them process student work or student data, and how each is switched off, and files the reply with the list.

Why it matters

The Council's working premise, which it has not tested and which its first year of research is meant to test, is that the common breach in an Indian school is not a hacker but a teacher pasting a class list of names and marks into a free web tool to draft report-card remarks, on a personal account, under terms that permit the provider to keep the text. A school cannot honour a withdrawal of consent it has no technical means to execute, and data that has left the school's control cannot be recalled by policy. This is the most expensive principle in the document. The dated page for each system waits on written answers from a supplier who may take weeks to give them, and a school running three to five systems should expect twelve to sixteen hours in the first year rather than an afternoon.

The school commits to

1. 1 No student name, photograph, admission number, address, contact number, health information or marks is entered into any free or personal-account AI tool. Where AI is used to draft remarks, identifying details are removed first. This is issued to all staff in writing and forms part of induction.
2. 2 For each AI system the school holds one dated page recording what student data it receives, where that data is stored, whether it leaves India, which other companies and model providers can see it, how long it is kept, and how it is deleted. The page is built from the supplier's written answers, which are filed with it.
3. 3 Parental consent is taken for the processing described on that page, in the language the parent reads, and the consent record is kept with the admission file. The record shows which parent consented, on what date, in which language and to which processing by which named system.
4. 4 Where the supplier stores or processes student data outside India, or passes it to a model provider that does, the school asks three questions in writing before signing: in which country the data sits, which company controls it there, and what happens to it if that company is sold or the contract ends. An answer that will not be given in writing is a reason not to sign, and the refusal is recorded with the proposal.
5. 5 In the first term the school tests deletion rather than assuming it: it asks the supplier to delete one withdrawn student's record and records the date asked, the date completed and the evidence returned. Thirty days without evidence is a failed test and goes to management.
6. 6 Access to student data inside each system is limited to staff who teach or support that student. The access list is printed and reviewed at the start of each academic year and whenever a staff member leaves.

Why it matters

A recommendation nobody has time to examine becomes a decision. Once a system has placed a child in the bottom set or flagged her as at risk, the label follows her from screen to screen and from year to year, and no single adult remembers choosing it. Teachers override a system only where overriding is quick, ordinary, recorded and never held against them. If the school cannot show where the human decision happened, there was no human decision.

The school commits to

1. 1 Every AI system in use has one named accountable adult, recorded in a register the principal keeps and reads each term. A system with no name against it is switched off until it has one.
2. 2 No mark that reaches a report card is produced by an AI system alone; a teacher reviews and signs each set of AI-assisted marks before release.
3. 3 Any AI-generated score, flag or grouping can be overridden by the teacher inside the system itself. The override survives the next sync, and the teacher's reason is stored with it. Most school systems cannot do this today, so for most schools the honest position on day one is a written exception: the override is kept on paper against the printed list, the absence is recorded as a finding against that supplier and put to it in writing, and the exception carries the date by which the school will have an answer.
4. 4 Before a system is switched on, the school writes down the override rate at which it will review that system for withdrawal. A school with no experience of such a system has no basis for choosing a figure and should not invent one: it records the first term's actual rate, sets the threshold at the second review, and says so in writing in the meantime. Where a school wants a starting figure, the Council suggests one in five, on the reasoning that a system whose judgement teachers reject more often than that is not doing the work it was bought to do, and that the review should therefore fall on the system. Overrides are counted each term against that number, and no teacher is asked to justify a high count.
5. 5 No AI system sends a message, score or alert to a parent without a named staff member seeing it first.

Why it matters

A demonstration screen showing AI-detected weak students rarely says what produced the detection. It may be attempts, minutes in an application, typing speed or attendance, none of which is understanding. Schools then buy a dashboard nobody reads, because nobody has been told what to do about an amber cell. An unexplained label is worse than no label: it carries the authority of a machine and cannot be argued with.

The school commits to

1. 1 For every score, band, flag or prediction shown about a student, the school holds one written sentence naming what produced it and what it does not mean, in this form: based on twelve questions attempted in one chapter, and not a measure of ability. Where a supplier will not provide that sentence, the refusal is recorded and put before management.
2. 2 Words such as weak, slow or at risk are not displayed to students or parents. No AI-generated label enters a child's permanent record or is passed to the next year's teacher as fact.
3. 3 Any AI output the school cannot connect to a specific teaching action is switched off and the reason recorded. A screen that changes nothing is not kept because it was paid for.
4. 4 A student, parent or teacher may ask why the system said what it said, and receives a written answer from the named accountable adult within ten working days, or within whatever period the school has written into its own resolution and told parents. Ten is the Council's suggestion and not a standard. A school that cannot staff ten should write down the period it can meet and meet that, because a published period nobody can keep is a worse position than a longer one stated plainly.
5. 5 Where a system's inference is contradicted by the teacher's own knowledge of the child, the teacher's account is what is acted on, and the disagreement is recorded.

Why it matters

Homework produced whole by a chatbot is now ordinary, and a blanket ban moves it out of sight rather than stopping it. Detection software is the wrong remedy. Published work on these detectors reports error rates high enough to make a single score unusable as evidence, and reports that the errors fall hardest on people writing in a language that is not their first. The Council has reviewed none of that work itself and offers this as its working assumption rather than as a finding; the commitment below therefore does not depend on it, because it is wrong to discipline a child on a number the school cannot explain whatever the number turns out to be. What works is an instruction stated task by task, some marks earned under supervision, and students taught that an AI answer is a draft to be checked, not a result to be copied.

The school commits to

1. 1 Every assignment carries one of three labels, printed on the assignment itself and not only in a policy: AI not permitted, AI permitted with disclosure, AI required.
2. 2 A student who used AI and disclosed it is not penalised for the disclosure. Work presented as one's own that was not is handled under the school's existing academic honesty rules, which are not rewritten for AI.
3. 3 No student is questioned or disciplined on the output of an AI-detection tool alone. A conversation with the student, or supervised re-performance of a comparable task, comes first, and a detector's score is never shown to a parent as proof.
4. 4 In Classes 6 and above, at least once each year in at least two subjects, students are given an AI answer that is confidently wrong and asked to find and correct the error.
5. 5 The school publishes, for each subject, the proportion of the term's marks earned under supervision: orally, in class, or by hand.

Why it matters

A tool demonstrated in English on a new laptop behaves differently in a classroom where instruction is in the state language and children share one phone at home after their parents return from work. AI-dependent homework quietly becomes a test of household income. Licences are sold per student at prices set for a market the school is not in, and renewed because withdrawal would look like an admission of failure. A tool that helps the front bench and reaches no one else has widened the gap it was bought to close.

The school commits to

1. 1 No AI-dependent task is compulsory homework unless it can be completed in school, on school devices, within school hours.
2. 2 Before adoption the school runs the tool in its medium of instruction and in at least one home language of its students, and writes down what it did badly. That note goes to management with the proposal.
3. 3 The school tests the tool on the oldest device in the building and on a slow connection, not on the supplier's laptop, and records what fails.
4. 4 Before signing, the school puts before its management the full cost per student over three years: licence, renewal at the supplier's stated escalation, training time, replacement devices and connectivity.
5. 5 Where a student with a disability or learning difficulty cannot use a tool, a named alternative route to the same work is written down and resourced. It is not left to the class teacher to improvise.

Why it matters

Every platform generates logins, minutes in application and coverage counts, and these are easily mistaken for effort. Once a management begins reading them as performance, teachers either work the numbers or withdraw from the tool, and honest reporting of what is going wrong in a classroom stops. Camera and voice analytics sold as engagement measurement place children under continuous observation in order to score the adult in front of them. A teacher who is afraid of the data will not tell you the truth about the class.

The school commits to

1. 1 The staff handbook lists by name the system-generated metrics that are never used in appraisal, and the principal signs that list.
2. 2 The school does not use AI camera analytics, voice analysis or engagement scoring in classrooms. Where a lesson is recorded for professional development, participation is voluntary, the teacher controls the recording, and it is deleted after use.
3. 3 Teachers see the same system data about themselves that management sees, in the same form and at the same time.
4. 4 A teacher may object in writing to the use of an AI tool in their classroom. The principal answers in writing within ten working days, or within the period the school has written down and told its staff, and the objection is not placed in the teacher's file.
5. 5 Before any tool is introduced the school records the hours of paid working time allotted for training and set-up. Time is given, not assumed.

Why it matters

Claims of improvement are usually the supplier's own usage data, with no comparison group and no independent examination, and there is very little Indian evidence on what AI does to learning at all. Negative results are rarely published, so every product appears to work. A trial with no stated question and no end date becomes a purchase by default, and the school discovers the cost when the invoice for fifteen hundred students arrives. A trial run properly is the largest single piece of work these principles ask for: ten to fifteen hours of one senior teacher's time across a term. It is also the cheapest of the two options, the other being a school-wide licence bought on a demonstration.

The school commits to

1. 1 Any learning claim is asked for in writing with the study behind it: who was studied, how many, for how long, against what comparison, and who paid for the work. An answer that does not arrive in writing is recorded as not provided.
2. 2 Before any school-wide purchase the school runs a trial in one or two sections with a stated question, a stated duration and a stated measure, and a stopping rule written before the trial begins and given to the supplier.
3. 3 A supplier's own data is filed as the supplier's data and is never repeated to parents or trustees as an independent finding.
4. 4 Before renewal the school writes one page for its management: what changed, what did not, and what it cost in teacher time. A renewal signed without that page is not a decision.
5. 5 Where the school agrees to research on its students, consent to the research is taken separately from consent to use the product, and every commercial interest of the researcher and of the funder is named in the same letter to parents.

Why it matters

Three years of attendance, marks, remarks and learning history can become unrecoverable because the contract promised an export and the export turns out to be a folder of PDF files, or because the supplier has changed its price, its product or its owner. Multi-year commitments are signed after a single demonstration by schools that have not read the exit clause. Lock-in is rarely written into a contract; it is built into a file format.

The school commits to

1. 1 Every contract states the school's right to receive all of its data in a machine-readable format, at no additional charge, within thirty days, both during the contract and after it ends.
2. 2 The school performs one full export in the first term and opens the file, checking that marks, attendance and remarks are present and readable. The clause is not relied on unopened.
3. 3 The contract names where data is stored, names every subprocessor and every AI model provider the data passes through, and requires written notice to the school before any of that changes.
4. 4 The first contract term does not exceed one academic year, and no term is longer than the trial that justified it.
5. 5 Administrator credentials for every system are held by a named person employed by the school and recorded with the principal, not held only by the supplier or by a single departing staff member.

Why it matters

The phrase AI-powered appears in school marketing when nobody in the building can say what it means. Parents consent at admission to something that was never described to them, and a parent who learns from another parent that a system profiles their child stops believing anything else the school says. Students are told least of all, although it is their work being read. Disclosure written in advance costs an afternoon; disclosure forced by an incident costs the school its standing. The English page is about five hours the first time and two hours a year after. Translating it into the languages parents actually read is a separate cost that a school should budget for as such, because an English page in a school whose parents read Telugu or Hindi is not a disclosure.

The school commits to

1. 1 Once a year, and at admission, the school publishes a one-page list of the AI systems in use, in the languages parents read, stating for each what it does, what data it sees, what it decides, and who in the school is accountable for it.
2. 2 The school does not describe itself as AI-powered or AI-enabled in its own communications unless the same document names the systems and what they do.
3. 3 A parent may ask that their child be excluded from any non-essential AI system. The school answers in writing, names the alternative arrangement, and does not treat the request as a complaint against the school.
4. 4 Students in Class 6 and above are told in class, each year, which systems read their work and what those systems record.
5. 5 The school keeps three written sentences for the admissions counter saying what it uses AI for, what it does not use it for, and why, so that the answer given to a parent who asks why the school across the road advertises AI is the school's considered position and not whatever the counter staff improvise that morning. The sentences are approved by the principal and are refreshed when the list changes.
6. 6 A data breach, harmful output or wrong decision affecting a child is reported to the affected parents within three working days of the school learning of it, or within the period the school has written into its adoption and told parents, together with what the school has done about it. Three days is short, and a school will often not know by then what happened; the report says what is known and what is not, and is followed by a second one. Any reporting duty the school owes under law, including under the Digital Personal Data Protection Act, 2023 and under the Protection of Children from Sexual Offences Act, 2012, is separate, runs on its own timetable and is not discharged by this.