India notified its AI governance rules on November 20, and the shape of them will surprise anyone who expected a European-style statute. There is no pre-approval for model release, no risk-tier licensing, and no separate regulator. What there is: mandatory registration above a compute threshold, labelling obligations for synthetic media, a domestic grievance officer, and a 90-day compliance window.
What the rules actually require
Three obligations matter for anyone shipping into India. Models trained above the threshold — reported at 10^25 FLOPs, roughly the frontier tier — must register with MeitY and file a model card covering training-data categories, evaluation results and known failure modes. Providers must label AI-generated images, audio and video in a machine-readable way, and expose a detection endpoint to platforms on request. And every provider above 5 million Indian monthly users must name a grievance officer resident in the country, with a 15-day response clock.
Nothing in the text requires a licence to launch. That is the headline difference from the EU AI Act, whose general-purpose AI obligations began biting in August and whose high-risk regime arrives next year.
Why light-touch won
Two arguments carried. The first is investment: state governments are competing for data-centre capacity, and a licensing regime is an easy thing for a hyperscaler to route around. The second is that India already has an enforcement hook in the IT Act intermediary rules, and a parallel AI statute would create two regulators arguing over the same takedown.
The trade-off is real. Registration without pre-clearance means the first time anyone sees a problem is after deployment. The rules compensate with a post-hoc power: MeitY can demand an audit and, in the extreme, order a model withdrawn — heavier than it sounds, because withdrawal is retroactive.
What it changes for API sellers
If you sell model access rather than a consumer app, the registration obligation sits with the model developer, not the reseller. But the labelling obligation follows the output, so an application built on a registered model still has to label generated media. Build that into the pipeline now; retrofitting provenance metadata across a year of assets is expensive.
The grievance officer requirement is the one that catches foreign companies. A named resident person with a 15-day clock is an operational commitment, not a form. Budget for it before you cross 5 million users, because the clock starts the day you do.
The comparison that matters
For a team deciding where to launch first, the three regimes now sort cleanly. The EU requires documentation and, for high-risk uses, conformity assessment before market. The US regulates by sector — courts, states and agencies rather than a single AI statute, as California's suit against OpenAI shows. India requires disclosure after the fact and a local complaints channel.
For most products the sequence is unchanged: launch in the US, register in India, and treat the EU as the market where documentation is a launch blocker rather than a compliance chore. What has changed is that India is now a market you can plan around with a date, instead of a jurisdiction you wait on.