India's AI Rulebook Is Being Written by Police Complaints, Not Lab Evals

In the last week of July, India's AI story was not a model launch. It was a police complaint.

A deepfake video of Commerce Minister Piyush Goyal, falsely showing him threatening student protesters, circulated widely enough that the government started pulling links and Delhi Police registered an FIR. The Press Information Bureau's fact-check unit ruled videos of two sitting ministers fake. Pune police broke up an extortion network that used AI voice masking to demand ransoms of 2 crore to 50 crore rupees from at least four businessmen. Days earlier, the National Testing Agency called viral NEET answer sheets digitally altered or AI-generated and warned candidates that fabricated documents carry legal consequences.

Four incidents, one theme: nobody could prove what was real.

Then look at what the state funded. The IndiaAI Mission's Safe and Trusted AI pillar approved 13 research projects, aimed at deepfake detection, bias mitigation, privacy-preserving learning and explainable AI, alongside 27 India Data and AI Labs and a new IndiaAI Safety Institute. Separately, the IT minister told Parliament the government had selected 20 sovereign Indian model projects, including Sarvam AI, Gnani.ai, BharatGen and Avataar AI.

Now put that next to Washington. The administration's finalized safety guidelines exempt open-weight US models from voluntary testing and reserve federal evaluation for closed frontier systems with serious cyber capability.

Two governments, two definitions of AI risk. The US definition is capability: can this thing help someone build a weapon or breach a network. India's working definition is authenticity: can a citizen tell whether this video, this voice, this document is real.

Neither is wrong. But only one of them is being driven by an incident queue that grows every week.

Here is the part worth predicting. Capability evals are a small market by design. They apply to a handful of frontier labs, run behind NDAs, and produce reports most buyers never see. Provenance and detection are the opposite: they apply to every video platform, every bank call center, every exam board, every court filing. That is procurement at national scale, and India is the first large market whose AI institutions were built with that as the founding problem rather than an afterthought.

If the IndiaAI Safety Institute publishes detection benchmarks rather than capability evals, it will not just diverge from the US and UK institutes. It will hand a ready-made compliance template to every country that has elections, exams and phone scams but no frontier lab. That is most of the world.

The obvious objection is money, and it is a fair one. India's public AI programme is broad, but the private capital under it is thin. The largest AI deals we track are compute contracts worth billions. One of the Indian rounds in the same month was a $1 million seed for Bioscan Research, which uses near-infrared spectroscopy and AI to catch intracranial injuries early. Excellent company, and roughly one ten-thousandth of a single compute contract. A detection standard that nobody can afford to implement is a document, not a regime.

**What to do with this**

Ask your AI vendors one question: can this system prove where its output came from, in a form an auditor or a court would accept. Not watermark marketing, an actual chain of custody. If the answer is no, you are exposed in exactly the way India's last week was.

And watch the IndiaAI Safety Institute's first publication. If it leads with detection and provenance rather than frontier capability, the global AI rulebook just split into two documents, and the bigger population is reading the other one.