Module 1 · Provisional research pilot 25 institutions · 2015–2025 · not a national ranking or capability score

Module 1 · Research map

How institutional AI research in India changed over eleven years.

Explore publication activity, institutional distribution and collaboration across a fixed cohort of 25 institutions. This research map is Module 1 of an evidence infrastructure being developed to connect Indian AI research with funding, infrastructure, models and real-world deployment.

Strict AI-core works12,758Across the full period
2015–2025 change4.55×476 to 2,167 works
Collaboration breadth170Observed institution pairs
Panel coverage275Institution-year observations

Annual activity

The strict AI core grew faster than the broader candidate set.

Annual AI publication activity from 2015 to 2025 Strict operational AI-core works rose from 476 to 2,167 while broad primary candidates rose from 1,850 to 5,427.
Unique works within the fixed cohort16.4% compound annual growth in the strict core
01

Output expanded

Among the selected institutions, strict-core activity rose from 476 works in 2015 to 2,167 in 2025.

02

More pairs participated

Active institution pairs rose from 20 to 68, even as multi-cohort works fell from 6.1% to 3.9% of annual output.

03

Concentration needs context

The aggregate increase disappears when VIT is excluded, so it is a member-specific pattern—not a cohort-wide shift.

Updated · Jev v0.2 beta

Audit how 41,172 candidate records were screened.

The public release documents the cohort, query, score-and-evidence routing, aggregate results, frozen-input hashes and internal agreement check. The detailed work-level explorer is withheld while field-level redistribution and data-minimisation review is completed.

Two views, two rules: the headline dashboard above uses the earlier strict-core rule frozen on 19 September 2026. The corrected Version 1.1 paper uses the v0.2 core-plus-adjacent primary rule and a 29 September work-ID snapshot. Their totals are not interchangeable.

Open release-method page Model-generated screening · not independently human-validated

Video overview

India’s AI research boom, explained.

This NotebookLM video introduces the pilot’s scope, methods, main findings and limits—and shows how to read the results without treating them as a ranking.

Institution explorer

Examine activity without turning it into a league table.

Counts are descriptive for this selected cohort. Alphabetical order is the default; publication volume is not research quality.

25 institutionsSelect a name for its eleven-year profile
InstitutionType2025 works2015–2025Co-authored share

Within-cohort collaboration

The network became broader, not consistently deeper.

Collaboration means a publication linked to at least two of the 25 cohort institutions. Partners outside the cohort are not counted.

Share of works with 2+ cohort institutions6.1% → 3.9%
Active institution pairs20 → 68

Most frequently observed ties

Repeated co-authorship links

Shared works are full counts. This list describes observed output, not relationship quality.

Robustness and uncertainty

The AI definition changes the result more than document type.

Sensitivity checks show which conclusions survive reasonable rule changes and which depend on a particular measurement choice.

Concentration diagnostic

One institution drives the aggregate increase.

2025 HHI0.1370.062 without VIT

The cohort-wide line should not be interpreted as evidence that every institution became more concentrated.

Data quality

Known limitations remain visible.

Core works with an abstract82.7%
Core works with a DOI98.5%

Independent human validation is pending. Classification and affiliation results on this site remain provisional.

Scope sensitivity

What changes when the rules change?

Bars show the change in unique works relative to the 12,758-work strict baseline.

Methods and downloads

Every number should be traceable to a rule.

The strict-core dashboard is frozen at 19 September 2026; the v0.2 explorer and Version 1.1 paper use a 29 September 2026 work-ID snapshot and a different operational rule. Sources and definitions are versioned so each view can be audited.

01

Sources

OpenAlex supplies publication and affiliation records. ROR supplies persistent organization identifiers.

02

AI boundary

A broad OpenAlex topic pull is narrowed by a conservative text rule for explicit AI or machine-learning methods and tasks.

03

Counting

Trends count each work once. Institution tables count one membership per directly affiliated cohort institution.

04

Interpretation

The cohort is purposive and selected on publication activity. Results cannot estimate national totals or policy effects.

Public research bundle · v1.1

Read the corrected paper or audit the aggregate results.

The public-minimal package contains aggregate outputs, methods, checksums and a data-minimised agreement subset of record IDs, labels and predictions. It excludes the full bibliographic corpus, publication titles and abstracts, author or affiliation strings, full work-level inputs, model outputs and membership rows, credentials and personal reviewer data.