New Delhi: State governments rolling out Artificial Intelligence on their administrative data risk automating its errors, a report by the Tony Blair Institute for Global Change (TBI) has said, warning that AI “will expose” gaps in records “faster and at greater scale”.
The 19 October report—titled From Digital Scale to Data Power: A Data Operating Model for India’s States, and authored by Ott Velsberg and Vivek Agarwal—says that AI systems trained on incomplete datasets may reproduce the gaps in eligibility decisions “at machine speed”, leading to wrongful mass exclusions.
It says an AI system used to flag welfare fraud should not suspend a benefit on the basis of a model score alone.
Departments should test for false positives, record the data and variables used, have an official review cases before action, and give the person a route to contest the decision, it says.
Also read: AI is cracking tough maths problems. How it can bring mathematics closer to cultural studies
MCP server not proof of readiness
The report questions the use of the Ministry of Statistics and Programme Implementation’s Model Context Protocol (MCP) server as a marker of progress.
The server lets an AI system query what an indicator means and where it comes from.
The report calls it “an interface pattern, not a data-governance model”, and says states should not treat it as evidence that their data are ready for AI. It says the server does not settle legal basis, data quality, access rights or accountability.
It adds that MoSPI’s five-layer AI-readiness framework does not address records kept in different languages and scripts, where spelling and transliteration differences stop machines from matching records of the same person.
Most government AI projects, the report says, remain pilots because each new use requires its own data clean-up.
What the data show
The report lists failures in the data that AI systems would draw on. India recorded about 5.3 lakh Covid deaths, while Civil Registration System (India’s official system for recording births and deaths) data, released in May 2025 after four years, point to about 24 lakh excess deaths in 2021.
When school reporting moved to student-level records in 2022-23, 1.4 crore duplicate or “ghost” entries disappeared. Of the 3,000 to 8,000 data elements health facilities report each month into the Health Management Information System, about 10 percent become indicators and half are left blank. One state’s Health Management Information System (HMIS) recorded 728 infant deaths in April-June 2015; its surveillance system recorded 3,307.
Citing NITI Aayog, the report says data errors cause leakage of 4 to 7 percent of annual welfare spending. Welfare schemes still allocate benefits using Census 2011 figures.
The report says linking databases can cut people off faster than manual systems. In Jharkhand, linking rations to biometric checks cut benefits by 10 percent among those without Aadhaar linkage. In Odisha, the state told the High Court that database errors had cut off food benefits for over 20 lakh people.
It also flags a limit on linking records: each scheme tokenises Aadhaar separately, and only UIDAI holds the mapping back to the number, so records across schemes cannot be matched to the same person with ease.
Data still move between departments as files and emails, based on ties between officials, according to MoSPI’s consultations with states cited in the report. In 2022, 50 government websites were hacked and eight breaches were traced to government bodies.
The recommendations
The report asks states to set up a data office with powers to compel departments to share data, saying MoSPI’s proposed Data Strategy Units, housed within a department, cannot do so. “Officials resist sharing data because information is power,” it says.
It proposes named owners for each state’s top 20 to 50 data sets, an audit body outside departments that publishes findings, and a “State Data Balance Sheet” listing data assets, liabilities and risks, including “AI readiness”, to be reviewed with the budget.
It recommends that states bring in AI only after data quality, human review and appeal routes are in place, and that vendor contracts include open standards, data portability and exit clauses.
The report names Telangana’s data exchange, which released over 500 data sets for AI use from over 20 departments, along with Karnataka’s Kutumba, Rajasthan’s Pehchan and Andhra Pradesh’s Real Time Governance Society as examples, but says these remain limited to single sectors.
Under MoSPI’s roadmap, states are to set up a data body by December 2026 and run data-exchange platforms by December 2028, with “AI-driven simulations” in the final phase.
(Edited by Ajeet Tiwari)
Also read: Spotify’s big clean-up: 75 million spam songs gone, 50 million hidden