Research analysis · Higher education policy
Indian students have settled the question of whether to use generative AI. Their institutions have not yet decided whether to have a position. Meanwhile a second wave, agentic AI, is already reshaping enterprise hiring while degree curricula are still catching up with the first.
Indian higher education has arrived at an unusual configuration. Adoption of generative AI among students is approaching saturation. Formal classroom rules are almost entirely absent. Public access programmes have been signed at national scale. And the technology has already moved into a second phase, agentic AI, which most degree curricula have not addressed even in its first form.
The material below draws on government surveys, parliamentary oversight documents, peer-reviewed and preprint academic work, and industry reports. Read together, they describe a system in which access is expanding much faster than governance, and in which a commercial certificate market is supplying a competence that credit-bearing degrees are not.
An important qualification applies throughout. Evidence quality in this field is uneven. Government statistics, parliamentary replies and academic studies rest on different foundations from vendor surveys and platform enrolment counts. Each claim below is labelled accordingly, and a dedicated section sets out the methodological cautions that a careful reader should carry into the numbers.
The scale problem comes first
Any assessment of AI in Indian higher education has to begin with the size of the system being asked to change. The All India Survey on Higher Education for 2023-24, released by the Ministry of Education in July 2026 alongside the 2022-23 edition, records total enrolment of 4.50 crore students, up from 4.46 crore the previous year and 3.42 crore in 2014-15, a decadal increase of 31.5 per cent. The Gross Enrolment Ratio rose from 29.5 to 30.0. Female enrolment reached 2.24 crore, and the Gender Parity Index stood at 1.08, above parity for the seventh consecutive year. STEM enrolment crossed one crore for the first time, at 1.02 crore, or 22.5 per cent of total enrolment, with women accounting for 44 per cent of STEM students against 38.4 per cent a decade earlier. A total of 59,533 institutions submitted data for 2023-24.[1]
Two features of that dataset matter for what follows. The first is simple arithmetic. Revising curriculum, assessment design and faculty capability across tens of thousands of institutions is slow work under the best conditions. Generative AI reached students at scale in roughly eighteen months. The mismatch between those two clocks explains most of the governance gap discussed below.
The second is that AISHE is self-reported by institutions, subject to validation before publication rather than independent audit.[1] It is the best national instrument available, and it is still an administrative return rather than a sample survey with sampling error. It tells us how large the system is. It does not tell us anything about what happens inside a classroom when a student opens a chatbot.
Insight
The absence of a national, annually repeated survey of AI use in Indian higher education is itself a finding. The United Kingdom has run one for three consecutive years. India has isolated institutional studies, vendor reports and platform enrolment counts. Policy is therefore being made on inference rather than measurement, in the country with the largest student cohort using these tools.
What students are already doing
The most substantial recent dataset is the Digital Education Council’s AI in Higher Education Global Survey 2026, which draws on 45,398 responses, comprising 27,284 from students and 18,114 from faculty across 35 countries.[2] It reports that 88 per cent of students now use AI in their learning and 77 per cent of faculty use it in teaching, the faculty figure having risen by 16 percentage points in a year.[3]
The gap is not in adoption. It is in the scaffolding around adoption. In the same survey, 57 per cent of students said their assessments came with inadequate guidance on AI use, only 29 per cent believed their instructors were equipped to advise them, and just 31 per cent of faculty agreed that their institution involved them meaningfully in shaping AI policy. Thirty-seven per cent of students expressed serious doubt about whether their programme remains relevant for an AI-affected labour market, while 43 per cent of faculty globally were not worried that what they teach will be outdated by graduation.[3] That divergence between student anxiety and faculty confidence is one of the more consequential findings in the dataset, because it predicts where institutional resistance to curricular change will sit.
The Indian evidence base
Indian evidence points in the same direction but rests on much thinner foundations. A University of Delhi study of Library and Information Science students, widely reported in May 2026, found that 66 per cent of respondents would recommend ChatGPT for academic work and 67 per cent named free round-the-clock availability as its principal advantage. Nearly 90 per cent said their institutions still had no formal AI-use policy.[5]
A 2026 scoping review mapping AI adoption, perceptions and governance across Indian higher education gives a sense of the wider literature. It includes 236 sources spanning 2018 to 2026, of which 197, or 83 per cent, were published after the public release of ChatGPT in November 2022. By source type, 180 are empirical studies, 38 are conceptual papers, nine are prior reviews and nine are policy documents. By discipline, technical and engineering education dominates at 112 sources, or 47 per cent, followed by general higher education at 66, medical and health professions education at 26, and library and information science at 15. Legal, pharmacy and nursing education together account for three.[6]
Caution on the Indian survey evidence
The Delhi finding is frequently cited as though it described the national student body. It does not. The underlying comparative study of LIS research scholars at the University of Delhi and Babasaheb Bhimrao Ambedkar University collected 62 valid responses during the 2024-25 session.[6] A sample of that size, drawn from a single discipline at two universities, supports a directional observation about policy absence. It cannot carry a percentage estimate for 4.50 crore students. The disciplinary skew in the wider literature compounds this: with under 2 per cent of the reviewed sources covering law, pharmacy and nursing together, almost nothing is known empirically about AI use in the professional disciplines where assessment integrity has statutory consequences.
The comparison that India cannot yet make
The HEPI and Kortext Student Generative AI Survey 2026, conducted by Savanta in December 2025 among 1,054 full-time UK undergraduates, reported that 95 per cent of students use AI in at least one way, up from 66 per cent in 2024, and that 94 per cent use generative AI to help with assessed work, against 51 per cent the previous year. Nearly two-thirds said assessment at their institution had changed significantly in response.[4]
Two secondary findings from that survey deserve more attention in the Indian debate than they have received. First, students from households in the higher social grades were more likely to use certain AI tools than students from manual or unemployed households, a disparity the report notes is unclear in origin given that a free version of ChatGPT exists. Second, several students described anxiety about being falsely accused of AI misconduct, including changing their own vocabulary in essays to avoid suspicion.[19] Both effects would be expected to be sharper, not milder, in a system with India’s income spread and its share of students writing academic English as a second or third language.
| Instrument | Coverage | Sample | Repeated annually | Evidence status |
|---|---|---|---|---|
| AISHE 2023-24, Ministry of Education | India, all higher education institutions | 59,533 institutions reporting | Yes | Official statistic |
| DEC AI in Higher Education Global Survey 2026 | 35 countries | 45,398 responses (27,284 students, 18,114 faculty) | Series since 2024 | Sector survey |
| HEPI and Kortext Student Generative AI Survey 2026 | United Kingdom | 1,054 full-time undergraduates | Yes, third edition | Sponsored national survey |
| University of Delhi and BBAU LIS study | Two Indian universities, one discipline | 62 valid responses | No | Small-sample study |
| Comparable Indian national student survey | Does not currently exist | |||
Scroll the table sideways on narrow screens.
Sources as cited at references 1 to 6 and 19.
The governance vacuum
The 2026 scoping review found no evidence that either the University Grants Commission or the All India Council for Technical Education had issued AI-specific classroom-use guidance.[6] The instrument being applied by default to AI-generated work is the UGC’s 2018 Regulations on the Promotion of Academic Integrity and Prevention of Plagiarism, which predate generative AI entirely, contain no disclosure requirement and offer no methodology for handling detection outputs.
What has emerged instead is improvisation at the institutional level. Similarity-checking infrastructure now produces AI-detection scores alongside conventional similarity reports for doctoral theses, and many universities have adopted informal thresholds above which an enquiry is triggered, without any national standard defining what such a score means or how it should be weighed. AICTE has designated AI a strategic priority, asked affiliated institutions to prepare implementation plans, and treats undisclosed AI use in academic submissions as plagiarism.[8]
Methodological caution: detection scores
Treating a numerical AI-detection score as evidence of misconduct is not defensible on current evidence. Detection tools do not disclose validated false-positive rates for the populations to which Indian institutions apply them, and the error is not randomly distributed: writing that is formulaic, heavily edited, or produced by second-language writers is systematically more likely to be flagged. An unofficial institutional threshold applied without published validation converts a probabilistic output into a disciplinary finding. Where such enquiries proceed, the detection score should function as a trigger for a viva or a request for drafting evidence, never as the substantive finding itself.
The scoping review’s own conclusion is worth taking seriously on its terms: expanding access without matching usage rules risks reproducing, in Tier-2 and Tier-3 institutions, precisely the integrity problems that metropolitan institutions have not solved.[6]
Access has been pushed harder than policy
On 25 August 2025 OpenAI launched an India-focused Learning Accelerator in partnership with IIT Madras, the Ministry of Education, AICTE and the schools body Arise. The programme covered the distribution of approximately 500,000 free ChatGPT licences to students and educators over six months, alongside a research collaboration with IIT Madras backed by 500,000 US dollars, reported in Indian coverage as roughly Rs 4.4 crore to Rs 4.5 crore, to study how AI affects learning outcomes and teaching methods with reference to cognitive neuroscience. A distinct India-specific subscription tier, ChatGPT Go, was launched at Rs 399 per month with UPI integration.[7]
In September 2025 AICTE signed a separate agreement with OpenAI to provide 150,000 free ChatGPT Go licences for six months to students and faculty at affiliated institutions, with a stated emphasis on strengthening digital skills and employability, and a teacher training component.[8]
These are substantial distribution commitments. They are not classroom policy, and the distinction has been blurred in public discussion. A licence determines who can use a tool. It says nothing about what a third-year student may submit, what must be declared, or how a department should respond when a submission is disputed. The training component inside the AICTE arrangement is the part of the package most directly relevant to the governance gap, and it is the part least visible in the coverage.
Insight: the sequencing question
Access-first sequencing is not automatically wrong. There is a reasonable argument that usage rules written before institutions understand actual student behaviour will be poorly specified and quickly obsolete, and that distributing tools generates the evidence base on which better rules can be built. The difficulty is that the research designed to produce that evidence, including the IIT Madras longitudinal work, will report on a timescale of years, while assessment decisions are being taken every semester in the interim. The defensible position is not access-then-policy or policy-then-access, but interim departmental rules that are explicitly provisional and dated for revision.
From generative to agentic
Generative AI produces content in response to a prompt. Agentic AI uses the same underlying models to plan, chain decisions, invoke external tools and carry out multi-step tasks with limited human direction at each step. The distinction matters for teaching for a specific reason: the second kind cannot be assessed in the way the first kind is currently being assessed. A submission produced by a system that retrieved sources, ran code, evaluated its own intermediate output and revised its plan is not usefully described by a similarity score or an authorship judgement.
The industry evidence that this shift is under way is real but should be read with care. Nasscom’s work on AI-first enterprises reports a majority of surveyed firms scaling a meaningful share of their generative AI initiatives, and a smaller but substantial minority scaling agentic proofs of concept. Its study of enterprise experiments with AI agents, drawing on a global survey of more than 100 large and mid-sized firms, reports high readiness to allocate dedicated budgets and a majority already somewhere between pilot and production.[9][10] Nasscom’s own reading is that the binding constraints on scale are now organisational rather than technical: data readiness, integration complexity and workforce adaptation, not model capability.
The hiring data corroborates the direction independently. Quess Corp’s India AI Workforce Analysis 2026 records governance, agent operations, runtime operations, evaluation and quality assurance functions together accounting for 26 per cent of AI hiring demand, and notes sharply rising demand for specialised skills including retrieval-augmented generation, orchestration frameworks, machine learning operations and AI governance as enterprises move from pilots to production.[13] That composition, weighted towards evaluation and control rather than model building, is the clearest available signal of what agentic deployment actually requires from people.
Caution on industry adoption figures
Enterprise AI adoption percentages circulate widely and should be treated as the weakest class of evidence used in this article. They rest on self-selected respondent panels, often numbering in the low hundreds; they depend on definitions of “scaling” and “in production” that vary between reports and are rarely published; and they are produced by organisations with a commercial or advocacy interest in the direction of the finding. They are useful as directional signals about where enterprise attention is moving. They should not be used as base rates, quoted to two significant figures, or presented to students as established fact.
The certificate market has moved faster than the degree
Almost every substantial agentic AI offering in India today sits outside the credit-bearing degree. It is a certificate, delivered online, aimed at working professionals, priced at professional rates, and in most cases routed through a commercial platform partner rather than the institution’s own delivery infrastructure.
| Institution | Programme | Duration | Delivery | Status of claim |
|---|---|---|---|---|
| IIT Bombay, Department of CSE | Certificate in Agentic AI, covering retrieval-augmented generation, agent protocols, orchestration frameworks, vector databases and multi-agent systems | Five months | Platform partner | Provider claim[20] |
| IIT Delhi, Continuing Education Programme | Advanced Certificate in Agentic AI, with the SeNSE Centre | Six months | Live online, platform partner | Provider claim[21] |
| IIIT Hyderabad | Engineering Agentic AI Systems | Twelve weeks | Direct institutional delivery, no platform intermediary | Provider claim[22] |
| EICTA consortium at IIT Kanpur, a MeitY initiative with the IITs, NITs and IIITs | No-code agentic AI course | Short course, of the order of twenty hours | Online | Provider claim[23] |
Scroll the table sideways on narrow screens.
Programme details, durations and prices are as published by the providers and change frequently. Prospective learners should verify current terms directly. Listing here is descriptive of the market and is not an endorsement.
The demand side is not in doubt. Coursera’s 2025 Global Skills Report recorded a very large year-on-year increase in global generative AI enrolments, with India the largest single source of learners, and Coursera executives have described India as the fastest-growing market for AI and generative AI course enrolment worldwide.[11][12]
Two conclusions follow, and they point in different directions.
The first is an equity conclusion. If agentic competence is available principally through paid certificates, then that competence becomes a function of household income and of whether a student’s family can absorb a fee that sits well above most public degree tuition. The socio-economic gradient that HEPI observed in tool use in the United Kingdom would be expected to appear in India at the level of skill acquisition rather than mere access.[19]
The second is a strategic conclusion for universities. Institutions that read the certificate market as competition are misreading it. A market that sells a skill at a premium is a signal about what the degree does not contain. The correct institutional response is curricular, not defensive.
Competing view
There is a serious argument against absorbing agentic AI into the undergraduate curriculum at all. On this view, tooling in this field turns over faster than any credit-bearing syllabus can be revised, so teaching specific frameworks guarantees teaching obsolete ones; the certificate market exists precisely because it can revise in months rather than years, and is therefore the appropriate delivery vehicle. The university’s comparative advantage lies in the durable substrate, which is probability, linear algebra, systems design, epistemology and professional ethics, and diluting that substrate to chase tooling would be a net loss. The counter-argument is that this reasoning proves too much, since it would equally have excluded databases, statistical software and version control from the curriculum, all of which are now taught without controversy. The workable resolution is to teach the invariants of agent design, which are decomposition, tool invocation, evaluation, failure modes and human checkpoints, and to treat named frameworks as replaceable illustrations rather than as content.
What the labour market is asking for
The single most instructive dataset on Indian AI hiring is Quess Corp’s India AI Workforce Analysis 2026, built from around 3.5 lakh job postings. It counts approximately 9.2 lakh AI professionals in India, of whom about 2.57 lakh sit in core, dedicated AI roles and 6.63 lakh hold AI-embedded roles, meaning they added AI skills on top of a different job. More than 70 per cent of the AI workforce therefore sits outside traditional AI specialist roles.[13]
A second finding in the same analysis sharpens the picture and is easy to miss. Between 66 and 68 per cent of active postings are for core AI roles, while 72 to 74 per cent of the existing workforce sits in AI-embedded roles.[13] Demand and supply are pointed in opposite directions. The stock of talent is broad and shallow; current hiring is asking for depth. Non-technical business functions nonetheless account for roughly 1.2 lakh postings citing AI skills, led by operations.
On readiness, the thirteenth India Skills Report, released by ETS in collaboration with CII, AICTE, AIU and Taggd and drawing on over 100,000 candidates through the Global Employability Test alongside responses from more than 1,000 organisations across seven industries, places national employability at 56.35 per cent, up from 54.81 per cent in 2025 and 46.2 per cent in 2022. It reports that more than 90 per cent of Indian employees already work with generative AI tools and that project-based hiring rose 38 per cent in a year. India is described as holding about 16 per cent of the global AI talent pool, projected to reach 1.25 million professionals by 2027.[14]
The disciplinary breakdown in the same report is where the curricular implication becomes concrete. Engineering employability slipped slightly to 70.15 per cent, with computer science at 80 per cent and IT at 78 per cent. Commerce rose sharply to 62.81 per cent from 55 per cent, science reached 61 per cent and arts 55.55 per cent, with the report attributing the commerce increase to BFSI and fintech hiring. MBA employability fell to 72.76 per cent from 78 per cent.[15]
Insight: AI literacy is not a computer science elective
The combination of two findings settles a question that Indian departments are still debating. Employers are hiring AI-fluent people into existing functions at roughly the same rate as they hire AI specialists, and the fastest employability gains are appearing in commerce rather than engineering. AI literacy therefore belongs in commerce, law, design, management, journalism and the health sciences as a matter of demand evidence, not as a matter of institutional fashion. A law faculty that treats AI as the computer science department’s problem is making a curricular error with measurable placement consequences.
Money, policy and the delivery gap
The IndiaAI Mission was approved in March 2024 with an outlay of Rs 10,371.92 crore over five years across seven pillars. Execution has fallen well short of announcement. Revised expenditure estimates for 2025-26 came down from a budgeted Rs 2,000 crore to Rs 800 crore, and the 2026-27 allocation was set at Rs 1,000 crore, roughly half the original 2025-26 proposal.[16][17]
The gap between allocation and disbursal is wider still. In a Rajya Sabha reply, the Ministry of Electronics and Information Technology stated that Rs 21.79 crore had been released in 2024-25 against revised estimates of Rs 173 crore, and Rs 379.15 crore in 2025-26 against revised estimates of Rs 800 crore, as of 9 February 2026, with nothing yet released for 2026-27.[18] The Standing Committee on Communications and Information Technology recorded that the mission had spent 32 per cent of its 2025-26 revised estimate funds as on 31 December 2025.[24]
The skilling shortfall is the part of this record that bears most directly on higher education. The parliamentary panel documented that against an undergraduate fellowship target of 5,000 in 2024, 150 fellows were selected; when the target was revised upward to 8,000 in 2025, 159 were selected. Postgraduate selections reached 101 against a target of 5,000. Only the doctoral stream exceeded its target, with 199 scholars against 167. The ministry attributed the shortfall to visibility.[24] Separately, analysis of the compute pillar indicates that over 85 per cent of the subsidy disbursed has gone towards indigenous generative model development rather than skilling.[16]
Distinguishing announcement from delivery
The IndiaAI record illustrates a general problem in reading Indian technology policy. Four distinct quantities are routinely conflated in public discussion: the five-year approved outlay, the annual budget estimate, the revised estimate, and the sum actually released. For 2025-26 these were Rs 10,371.92 crore across the mission, Rs 2,000 crore budgeted, Rs 800 crore revised, and Rs 379.15 crore released as of February 2026. Any analysis that quotes the first figure as though it described current capacity is describing an intention, not a resource. The same discipline should be applied to summit pledges and to institutional targets throughout this article.
Budget 2025-26 announced a Centre of Excellence in AI for Education with an outlay of Rs 500 crore. Budget 2026 did not re-announce it, proposing instead a high-powered Education to Employment and Enterprise Standing Committee, whose terms of reference include assessing the impact of emerging technologies on jobs and skill requirements, embedding AI in school curricula, upgrading State Councils of Educational Research and Training for teacher training, and recommending upskilling measures for technology professionals.[17]
The summit and what it committed
India hosted the India AI Impact Summit at Bharat Mandapam in New Delhi from 16 to 20 February 2026, the first global AI summit convened in the Global South, with delegations from 118 countries. The New Delhi Declaration on AI Impact was endorsed by 88 countries and international organisations, framed around three guiding principles described as People, Planet and Progress.[25] The declaration calls for international cooperation across seven themes: human capital, inclusion for social empowerment, safe and trusted AI, science, democratising AI resources, resilience and innovation, and AI for economic development and social good.[26] Official outcome documentation records a set of casebooks, startup ecosystem publications and investment commitments across infrastructure, foundation models, hardware and applications.[27]
The framing shift from earlier summits at Bletchley, Seoul and Paris, which centred on risk, to a summit centred on impact is substantive rather than cosmetic. It also has a cost. A risk-centred agenda generates commitments that are auditable, because they concern what must not happen. An impact-centred agenda generates commitments that are aspirational, because they concern what should. Human capital appears as the first of the seven declaration themes, which is the right placement. None of it reaches the level of a rule about what a third-year student may submit in an assessed piece of work.
Reading the evidence: a note on categories
Much of the confusion in this debate comes from mixing evidence types that carry different weight. The following distinctions are used throughout this article and are worth carrying into any other reading on the subject.
| Category | What it is | How much weight it carries | Example in this article |
|---|---|---|---|
| Official statistic | Administrative return or parliamentary record published by government | High for what it measures; silent on what it does not | AISHE enrolment; IndiaAI funds released |
| Institutional target | A stated intention with an allocated or announced figure | Describes intent, not capacity or delivery | Fellowship targets; licence distribution volumes |
| Commercial claim | Vendor, platform or staffing firm survey, or advertised course terms | Directional only; definitions rarely published | Enterprise agent adoption shares; course fees |
| Projection | A modelled or extrapolated future figure | Sensitive to assumptions; should never be quoted as fact | Talent pool projected to 2027 |
Scroll the table sideways on narrow screens.
Claims omitted from this analysis
Several figures that circulate widely in Indian coverage of this subject have been left out because they could not be traced to a primary institutional source with a stable published methodology. These include specific percentages for the share of Indian institutions with deep AI integration, certain graduate employability indices reported at variance with the India Skills Report figures cited above, and a frequently repeated growth statistic for AI-detection flags in Indian higher education. Their absence here is not a judgement that they are false. It reflects the position that a claim which cannot be checked should not be used to support a policy argument.
An evaluation framework departments can apply this session
Very little of the useful work requires central clearance. The framework below is structured so that a department can score itself, act, and re-score after one academic session. Each criterion is written to be observable rather than aspirational.
| Criterion | Observable test | Minimum acceptable standard | Evidence basis |
|---|---|---|---|
| 1. Written policy | Can a student open the course handbook and find, for each assessment, what AI use is permitted and what must be disclosed? | A dated, one-page departmental statement covering permitted use, disclosure format and the consequence of non-disclosure | Near-total absence of institutional policy in surveyed Indian cohorts[5] |
| 2. Assessment redesign | What proportion of marks in each paper depends on evidence that cannot be produced without the student present? | At least one component per paper carrying viva, in-class writing, annotated drafts, process logs or defence of method | Detection scores are contested and error-prone; 65 per cent of UK students report assessment has already changed[4] |
| 3. Due process on suspicion | If a detection score triggers an enquiry, what happens next, and is it written down? | Score functions as trigger only; finding rests on drafting evidence or oral defence; the student sees the basis of the allegation | No official national threshold exists; false positives fall hardest on second-language writers[6] |
| 4. Faculty capability | Could a randomly chosen member of staff advise a student on permitted AI use in their own paper? | Every teacher of an assessed paper has completed one structured session and holds the departmental policy | Only 29 per cent of students believe instructors can guide them; 31 per cent of faculty feel consulted on policy[3] |
| 5. Agentic concepts taught | Does any paper require a student to specify where an autonomous step is acceptable and where a human checkpoint is mandatory? | One taught unit covering decomposition, tool invocation, evaluation, failure modes and human-in-the-loop design, framework-agnostic | Governance, evaluation and agent operations account for 26 per cent of AI hiring demand[13] |
Scroll the table sideways on narrow screens.
Notes on applying the framework
Write the policy before redesigning the assessment. The policy is a week’s work and removes most of the ambiguity that currently gets resolved as suspicion. Assessment redesign takes a session and requires board approval. Sequencing them the other way leaves students exposed for a full year.
Teach agent design as system design, not as tooling. The concepts appearing across the IIT certificate syllabi, including retrieval-augmented generation, agent protocols, multi-agent orchestration, tool use and evaluation, are teachable without heavy compute. What students most need is judgement about where an autonomous step is acceptable and where it is not, which is a design and ethics question before it is a technical one.
Invest in faculty before infrastructure. The binding constraint indicated by the survey evidence is teacher confidence rather than tool access, and train-the-trainer provision already exists inside the AICTE arrangement.[8] Departments that buy licences before training staff are solving the constraint they do not have.
Track uptake by gender and by household background. Women constitute 44 per cent of Indian STEM enrolment[1] and, in the UK data, tool use varies significantly with household social grade.[19] If either gap holds as agentic skills become a hiring filter, it will surface in placement outcomes within a few admission cycles, by which point it is expensive to correct.
Research opportunities for Indian universities
The evidentiary gaps identified above are, from a research standpoint, an opportunity. Several are addressable with modest resources and would produce work that is genuinely novel rather than derivative of Western datasets.
- A repeated national student survey. The single highest-value contribution available. A consortium of universities running an annually repeated, methodologically transparent instrument, stratified by discipline, institution type and state, would create the base rate that Indian policy currently lacks. The UK precedent shows the design is tractable at around a thousand respondents per wave.
- False-positive rates in AI detection for Indian academic English. A validation study using verified human-authored submissions from Indian undergraduates across disciplines and language backgrounds. This is directly actionable, it is not being done elsewhere because the population is specific to India, and its findings would bear immediately on disciplinary procedure.
- Assessment redesign as a controlled comparison. Departments switching from take-home written assessment to process-evidenced or viva-supported formats create natural experiments. Learning outcome measurement across such transitions would test the widely asserted but rarely evidenced claim that redesigned assessment produces better evidence of learning rather than merely harder-to-outsource evidence.
- The equity gradient in certificate uptake. Who buys agentic AI certificates, at what household income, from which institution type, and with what placement effect. The commercial providers hold this data and do not publish it; a university-led study using placement records could establish whether paid certification is functioning as a mobility ladder or as a stratifier.
- Discipline-specific agentic risk taxonomies. Law, medicine, accountancy and architecture each have statutory or professional constraints on delegated judgement. Mapping where an autonomous step is professionally impermissible, discipline by discipline, is scholarly work that professional councils will need and that no vendor will produce.
- Indic language performance in educational settings. Adoption programmes cite multilingual support, but independent evaluation of model performance for academic reasoning tasks in Indian languages, as distinct from translation quality, remains thin.
Insight: the comparative advantage
Indian universities cannot compete on frontier model development, where the compute constraint is decisive and where over 85 per cent of the compute subsidy has in any case been directed. They can compete decisively on deployment evidence, because India presents the largest and most linguistically varied population of student AI users anywhere. Research on how these systems behave in low-resource, multilingual, high-volume educational settings is work that only India is positioned to do well, and it is currently under-supplied.
Limitations of this analysis
Three limitations should be stated plainly.
First, the Indian evidence on student behaviour is thin. The strongest India-specific behavioural findings cited here rest on small samples in one or two disciplines. The global surveys have much larger samples but are not stratified in a way that permits reliable disaggregation for India specifically. Statements about what Indian students are doing are therefore inferences from adjacent evidence, and are presented as such.
Second, the industry and platform figures used in the sections on agentic adoption and course demand are self-reported by organisations with an interest in the finding. They have been retained because they are the only available signals in those areas and because the direction they indicate is corroborated by independent hiring data. They should not be treated as measurements.
Third, this analysis is written from the institutional side. It says relatively little about student experience beyond survey aggregates, nothing about the political economy of the platform partnerships that intermediate most certificate delivery, and nothing about school-level preparation, which determines what capability students bring into the first year. Each is a substantial subject in its own right.
Conclusion
The honest summary is that Indian higher education has bought access, declared intent and hosted the summit, but has not written the rules or changed the assessment. Students settled the question of whether to use these tools some time ago. Institutions are still deciding whether to hold a position.
The window on agentic systems is narrower than the one that preceded it. Generative AI arrived before universities had a policy, and the sector spent three years catching up. Agentic AI is arriving into the same vacuum, with enterprise hiring already weighted towards governance, evaluation and agent operations rather than model building.[13] The curriculum revision cycle at a typical affiliated college runs to several years. The distance between those two clocks is where the employability numbers are generated.
It is worth saying what the evidence does not support. It does not support the view that student AI use is primarily an integrity crisis; the DEC finding that students doubt their programme’s relevance is a curriculum finding rather than a cheating finding.[3] It does not support the view that India is behind in adoption; on the available measures India is among the fastest-adopting systems anywhere. And it does not support catastrophic readings of the skills position, given that measured employability has risen consistently since 2022.[14]
What it does support is narrower and more actionable. The deficit is in written rules, in assessment design, in due process, and in faculty confidence. All four are departmental-level problems with departmental-level solutions, and none of them requires waiting for a national framework that has not been issued in three years and may not be issued in the next three.
Reports and Research Sources
All links open in a new tab. Figures reported by commercial providers are identified as such in the text. Where reporting of a primary document differs across outlets, the government or publisher source has been preferred.
- Ministry of Education, Government of India. All India Survey on Higher Education (AISHE) 2022-23 and 2023-24, reported by DD News, July 2026. ddnews.gov.in
- Digital Education Council. AI in Higher Education Global Survey 2026. Digital Education Council, 2026. digitaleducationcouncil.com
- EdTech Innovation Hub. AI in higher education survey 2026: student AI use hits 88%, faculty lag. ETIH EdTech News, July 2026. edtechinnovationhub.com
- Stephenson, R. and Armstrong, C. Student Generative AI Survey 2026 (HEPI Report 199). Higher Education Policy Institute, sponsored by Kortext, 2026. hepi.ac.uk
- ETV Bharat. Is a lack of policy turning Indian students into AI-dependent learners? ETV Bharat, May 2026. etvbharat.com
- Mapping artificial intelligence adoption, perceptions, and governance in Indian higher education: a scoping review. Research Square preprint, 2026. Not yet peer reviewed. researchsquare.com
- OpenAI. Introducing the OpenAI Learning Accelerator in India. OpenAI Global Affairs, August 2025. openai.com
- Shukla, S. AICTE partners with OpenAI to provide 1.5 lakh free ChatGPT Go licenses to educational institutions. Careers360, September 2025. news.careers360.com
- Nasscom. AI-First Enterprises: Trends and Emerging Directions. Nasscom Insights. community.nasscom.in
- Nasscom. Enterprise Experiments with AI Agents: 2025 Global Trends. Nasscom Knowledge Center, 2025. nasscom.in
- Coursera. 2025 Global Skills Report. Coursera Blog, 2025. blog.coursera.org
- Business Standard. Pace of AI, GenAI learning courses fastest in India, says Coursera CTO. Business Standard, March 2026. business-standard.com
- Quess Corp. India AI Workforce Analysis 2026, reported by IANS and the Free Press Journal, June 2026. freepressjournal.in
- ETS, with CII, AICTE, AIU and Taggd. India Skills Report 2026 (13th edition), reported by Careers360, November 2025. news.careers360.com
- Careers360. India Skills Report 2026: CS, IT engineers most employable; women beat men for first time in five years. Careers360, 2025. news.careers360.com
- Takshashila Institution. Decoding the budget for India’s AI priorities. Takshashila Institution, February 2026. takshashila.org.in
- ThePrint. Budget 2026 allocates Rs 1,000 crore for IndiaAI Mission, pushes data centres and AI upskilling. ThePrint, February 2026. theprint.in
- Medianama. IndiaAI Mission: only Rs 400 crore of the over Rs 10,000 crore five-year outlay has been released, reporting a Ministry of Electronics and Information Technology reply in the Rajya Sabha. Medianama, April 2026. medianama.com
- Stephenson, R. and Armstrong, C. Student Generative AI Survey 2026, full report (PDF). Higher Education Policy Institute, 2026. hepi.ac.uk (PDF)
- IIT Bombay, Department of Computer Science and Engineering. Certificate in Agentic AI, programme page hosted by the delivery partner. Accessed 2026. mygreatlearning.com
- IIT Delhi, Continuing Education Programme. Advanced Certificate in Agentic AI, programme page hosted by the delivery partner. Accessed 2026. iitdelhi.emeritus.org
- IIIT Hyderabad. Engineering Agentic AI Systems. Data Foundation and Learning, IIIT Hyderabad. Accessed 2026. dfl.iiit.ac.in
- EICTA consortium, IIT Kanpur (a MeitY initiative with the IITs, NITs and IIITs). Agentic AI course. Accessed 2026. eicta.iitk.ac.in
- ThePrint. AI Mission misses fellowship targets by miles, spends just 32% of funds, Parliament panel report reveals, reporting the Standing Committee on Communications and Information Technology. ThePrint, August 2026. theprint.in
- All India Radio News. India AI Impact Summit 2026 adopts New Delhi Declaration. NewsOnAir, Prasar Bharati, February 2026. newsonair.gov.in
- Department of Industry, Science and Resources, Government of Australia. Australia endorses India AI Impact Summit Declaration. February 2026. industry.gov.au
- Press Information Bureau, Government of India. India AI Impact Summit 2026: landmark global declaration and major AI investment commitments. PIB, February 2026. pib.gov.in
- EY India. Future-ready campuses: harnessing AI in higher education, opportunities and the road ahead. EY, undated (PDF). ey.com (PDF)
