AIKosh: the IndiaAI Datasets Platform

The IndiaAI Mission's Datasets Platform pillar has an outlay of Rs 199.55 crore, about 1.9% of the Mission's Rs 10,371.92 crore. The government said when it disclosed the allocation in December 2024 that it aimed to make "public sector" datasets AI-ready, and MediaNama noted that the statement did not explain how data quality would be verified, how datasets would be curated or anonymised, or whether data could be re-identified when combined with other data.⁶

On 9 October 2024 the National e-Governance Division (NeGD) said that it planned an open-source platform for hosting datasets by January 2025, modelled on Hugging Face, with data from central and state governments and the private sector. Abhishek Singh, MeitY Additional Secretary and CEO of the Digital India Corporation, had said in September 2024 that Data Management Offices were being set up, with a Data Management Unit in each department.⁴

Minister Ashwini Vaishnaw launched AIKosh on 6 March 2025, together with the IndiaAI Compute Portal, an AI Competency Framework for public officials and other initiatives, on the Mission's first anniversary. The platform was described as a repository of government and private-sector datasets, including census reports, geospatial data and language data, available to AI developers as training data.⁵

Under its Terms of Use, AIKosh is "designed, developed and hosted by the IndiaAI Division" and has a dataset repository, a model hub, a use-case library, a notebook workspace and toolkits.³ According to its FAQ, roles are:²

Only contributors from a registered organisation may upload; an upload is published only after approval by the organisation's admin and a platform admin. Eligible entity types are public organisations, private companies, non-profits, research organisations, universities and startups or MSMEs. Login uses National Single Sign-On, with encryption at rest and in transit, and the Terms add MeriPehchaan for individual accounts and Entity Locker for organisations.²³ The platform's notebooks run on the AIRAWAT cluster: the CPU tier is free, the basic GPU tier gives four free hours and the advanced GPU tier requires approval.¹ Dataset listings carry a 'Data Quality Score (Beta)'.¹

The Terms of Use set out the platform's legal position:³

• Content and rights. Contributors are "solely responsible for the Content" they post, must hold the rights to it and must not upload unlawful, misleading or malicious material. They grant AIKosh a worldwide, royalty-free, non-exclusive licence to host and distribute content to run the platform, but keep ownership of their artefacts.
• Licences. Contributors choose a licence from a list and may request others; the licence stays attached when others download or redistribute, and AIKosh does not override artefact licences. Restricted artefacts need the contributor's explicit consent.
• Personal data. The platform is described as serving "non-personal and anonymized datasets". Contributors can see downloader details such as name, email and usage; data handling is governed by a Privacy Policy that is referenced but not reproduced.
• Liability. Services and content are provided 'as is', with no warranty of accuracy or licence compliance, and users indemnify the platform against claims from their use.
• Takedown and law. AIKosh may remove content or accounts without notice, deletes a terminated user's content within 90 days on a commercially reasonable basis, and takes copyright complaints by email. Indian law applies and the courts in New Delhi have exclusive jurisdiction.

The Terms keep the platform's own intellectual property with the Government of India while leaving contributors' ownership untouched. The platform's open, restricted and private settings are the mechanism through which data owners, including public bodies and startups, retain control over use.²³

The platform has grown in the stated numbers, but the counts differ by source:

• As of 31 August 2025 the governance guidelines said AIKosh had onboarded 1,500 datasets and 217 AI models from 34 entities across 20 sectors, with "permission-based access, allowing contributors to retain control over data usage".⁷
• NITI Aayog's September 2025 report on AI for Viksit Bharat said that AIKosh hosts "over 350 curated, non-personal datasets", such as census data, Indian language resources and satellite imagery, citing a July 2025 IndiaAI call for expressions of interest to contribute datasets and AI artefacts.⁸
• The platform's own pages, viewed in October 2026, listed 16,072+ datasets, 362 models, 671 organisations and 22 sectors, with sections for use cases, toolkits, notebooks, competitions and a leaderboard.¹

In March 2026 the Ministry told the Lok Sabha that the sovereign models of Sarvam AI and BharatGen, both funded under the IndiaAI Mission, were now available on AIKosh, which the Ministry said would let other startups and researchers build applications on them.¹¹

The governance guidelines call for "market incentives" to encourage public and private entities to contribute to AIKosh, the Open Government Data Platform and the National Data and Analytics Platform, for data governance frameworks to support sharing of anonymised data, data stewardship and data sovereignty, and for standardised evaluation datasets so that developers can test fairness in the Indian context.⁷

Opacity on quality and usage. In September 2026 a MediaNama commentator called AIKosh "a promising start" but noted "no public data yet on how much it is actually used or how good that data really is", and urged faster work on high-quality sovereign data.¹⁰

What counts as non-personal. In March 2026 MediaNama reported that the government plans 50 "AI curation units" across ministries and departments to identify, filter and upload non-personal government datasets to AIKosh, but had not said which ministries would take part, how 'non-personal' data or 'high-value' datasets are defined, what safeguards precede upload, or whether the Ministry of Law and Justice was consulted on the legality of using government data for AI training. MediaNama filed an RTI request with MeitY on those points; no reply had been reported when the article was last updated on 4 April 2026.⁹

Re-identification and consent. The December 2024 breakdown did not address how anonymisation would be verified, and MediaNama pointed out that datasets released as anonymised can still allow re-identification when combined with other data.⁶ The governance guidelines separately raise questions about the DPDP Act's treatment of training on publicly available personal data and say they require review by the proposed AI Governance Group.⁷

Contributor responsibility. The Terms place responsibility for rights and legality on contributors and disclaim warranty by the platform, so downstream users of an AIKosh dataset bear the risk of any copyright or data-protection defect unless a contributor's licence says otherwise.³

As of 15 October 2026 AIKosh is live and growing, and it is the designated repository for models funded by the IndiaAI Mission. The sources reviewed did not show independent measurements of usage, a published methodology for the data quality score, or a response from MeitY to the March 2026 RTI request about curation units and non-personal data.¹⁹ The Datasets Platform pillar's outlay of Rs 199.55 crore remains the second smallest of the Mission's seven pillars after safe and trusted AI.