How the register works
What counts as an incident, how a record is written and checked, what we leave out, and what the register can and can’t tell you.
What goes in.
An AI system was involved
The case involves a system that generates content, recognises people, or makes or recommends decisions: a chatbot, a deepfake tool, face recognition, a scoring algorithm. Where sources disagree about whether AI was really involved, or the system is rule-based matching and not machine learning, the record says so.
It happened in India
Or it affected people in India, even if the system was built or run elsewhere.
Someone was harmed, or could have been
An incident is a case where harm followed. A hazard is a failure, a misuse or a practice that could plausibly lead to harm, where none has been shown yet. Both go in, and each record says which it is.
A reader can check it
Every record rests on something public: a court order, an official statement, the organisation’s own account, or reporting by an established outlet. A report sent to us is published only once we’ve found that kind of support for it.
How a record is made.
Found or reported
We add cases from public reporting, and anyone can send one through the report form.
Checked
We read the sources themselves and keep only what they state. Where accounts differ, the record gives both.
Written
A short, neutral summary in a fixed format: what happened, the system, who was affected, what followed, and the sources.
Kept up to date
When a case moves on, with an arrest, a ruling or a fix, the record is updated and the date of the change is shown.
What we leave out.
We don’t name private individuals, whether they were harmed or accused. Public figures are named only when their identity is what the case is about and it has already been widely reported. We never publish who sent us a report.
We describe what was done without the detail that would help someone repeat it. And we leave out claims we can’t support: a record is a summary of what its sources report, not a finding that anyone was at fault.
How records are classified.
Each record carries the same few labels, so the register can be searched and counted. They are our reading of the sources, and reasonable people could classify some cases differently.
Sector
Where in Indian life the system was being used. The values are: Finance and payments; Welfare and public services; Courts, tribunals and policing; Elections and public information; Online platforms and media; Health; Education; Work and hiring; Transport; Other.
System
The kind of AI involved, in plain terms. The values are: Synthetic media; Language model or chatbot; Face or voice recognition; Automated decision system; Camera-based detection; AI agent; Other or unclear.
How AI was involved
Whether someone used the system to cause harm, the system failed, or it worked as built and the way it was used caused harm. The values are: Used deliberately to cause harm; Failed or made an error; Worked as built, and the way it was used caused harm.
Harm
What kind of damage was done, from the OECD’s list. A record can carry more than one. The values are: Physical; Psychological; Reputational; Economic or property; Environmental; Public interest; Human or fundamental rights.
Based on
The strongest thing the record rests on. A finding means a court, tribunal or regulator has ruled on the facts. A statement means the police, a ministry or the organisation involved has said what happened, which is their account and has not been tested. Independent reporting means journalists or researchers documented it. The values are: A court or regulator’s finding; An official statement; Independent reporting; A report to the register, checked.
The standards it follows.
The register uses the OECD’s terms. An AI incident is an event where the development, use or malfunction of an AI system leads to harm: to people’s health, to critical infrastructure, to human rights or rights protected by law, or to property, communities or the environment. An AI hazard is an event that could plausibly lead to one. The types of harm on each record are the OECD’s list too.
A record also covers six of the seven items that the OECD’s common reporting framework treats as essential: a title, a description, how the AI system was related to what happened, whether it was an incident or a hazard, the type of harm, and the supporting material. The seventh is who submitted the report, which we keep private.
India’s AI Governance Guidelines, published by the IT ministry in November 2025, recommend a national database of AI incidents, and the Telecommunication Engineering Centre has published a schema for recording them. This register is not that database. It is a community effort that keeps the same basic fields, so its records could be mapped onto a national one later.
Corrections and replies.
If a record is wrong, incomplete or out of date, write to hello@indiaaisafety.org with the record number. Organisations and people named in a record can send a reply, and we’ll add it or link to it.
Changes are made in the open: the record shows when it was last updated, and a record that turns out to be unsupported is withdrawn with a note saying why.
What the register can’t tell you.
It can’t tell you how often AI causes harm in India. Cases get here because someone reported them and we could check them, which so far means mostly English-language coverage of cases that reached the press, the police or a court. What you see reflects what gets reported at least as much as what happens.
It is kept by volunteers and it is new. There will be gaps and there may be mistakes, which is why every record links to its sources and why the corrections address above matters.
Using the data.
The records can be downloaded as CSV or JSON and reused under a Creative Commons Attribution 4.0 licence, which asks only that you credit India AI Safety Network and link back. The articles and documents the records link to belong to their publishers.