India Is Writing Its Own AI Safety Rules, and Declining to Copy Washington's

India needs an AI safety framework built around its own languages, scale, and application-driven ecosystem rather than a copy of the voluntary US pact, according to industry experts cited across eight outlets this week. The specific objection is to the agreement signed on September 29 by President Trump alongside executives from Google, OpenAI, Anthropic, Meta, Nvidia, and xAI.

The substantive question underneath is whether a voluntary framework built by the companies it governs can travel. India's answer, so far, is that it cannot.

What the US pact actually commits to

The September 29 agreement is voluntary, which is the first thing to understand about it. Voluntary frameworks work when the signatories have reputational incentives to comply and when non-compliance is visible. The companies that signed it are the companies whose models the framework covers, which creates a structural tension that no amount of good faith resolves.

Modeled on the arrangement, compliance is largely self-reported. There is no independent verification body, no publication requirement for safety evaluation results, and no consequence for a signatory that decides a particular evaluation is not applicable to its systems.

For the United States, this may be a defensible position. The argument is that mandatory safety requirements would slow domestic labs while Chinese labs face no equivalent constraint, and that a voluntary posture preserves competitive position. Whether that argument is correct is a live debate; several former employees of the signatory companies testified at a New York City Council hearing the following week that they do not believe the labs control what they are building, and none of the executives present would put a number on catastrophic risk when asked directly.

Why India's calculus is different

India's objection is practical rather than philosophical: the pact's assumptions do not hold in the Indian market.

The first is language. A safety framework built around English-language evaluation does not cover systems deployed in Hindi, Tamil, Bengali, Telugu, Marathi, and the rest of the country's scheduled languages. Safety evaluations are language-specific: a model that refuses harmful requests reliably in English may comply readily in a lower-resource language, because safety training is thinner and refusal behavior does not transfer cleanly. Testing only in English would produce a framework that certifies nothing about most of the deployments it purports to cover.

The second is scale and application profile. India's AI deployment pattern is heavy on public services, agriculture, financial inclusion, and health, sectors where errors have direct welfare consequences and where the user base is large, low-bandwidth, and often interacting through voice. A framework designed around frontier model capability evaluations does not map onto a system that triages crop disease for millions of smallholders.

The third is jurisdiction. A voluntary pact signed in Washington has no mechanism to bind a company operating in India, and India is not going to outsource its regulatory posture to an arrangement it did not negotiate.

What an Indian framework would need to include

The experts quoted in the reporting are not asking for a copy of the EU AI Act, which is the other obvious template. The EU approach is risk-tiered and mandatory, with conformity assessment, technical documentation, and post-market monitoring obligations. It is comprehensive and it is expensive to comply with, which is a real consideration for a market where most AI deployment happens through smaller vendors and public-sector integrators.

What India-specific concerns point toward is something narrower and more testable.

Multilingual evaluation as a baseline requirement, covering the major scheduled languages, rather than an English evaluation with translations added later.

Sector-specific standards rather than a single horizontal framework, since the failure modes in agricultural advisory differ fundamentally from those in credit scoring.

Government purchasing power as the enforcement mechanism. India's government is a major buyer, and government procurement conditions on AI systems (evaluation results, model documentation, incident reporting) would bind suppliers without requiring a new regulator.

And a domestic incident reporting channel, because the alternative is that safety-relevant failures show up first in the press.

Each of these carries a cost that has to be weighed honestly. Multilingual evaluation is not a translation exercise. Building a safety evaluation suite for a language means assembling adversarial prompts, recruiting native-speaker reviewers, and establishing what constitutes a harmful completion in that cultural context. The last part is genuinely hard and contested, because harm is context-dependent, and a refusal standard calibrated in one linguistic community does not transfer cleanly to another.

Sector-specific standards have a similar problem at the other end. A standard written for agricultural advisory and another written for credit scoring, and another for health triage, is a standard-setting workload that few regulators anywhere have staffed for. The EU handles this partly through harmonized standards bodies with industry participation, which is a structure India would need to build.

At procurement is the cheapest of the four and probably the most effective. It requires no new institution, only a clause in existing contracts. It is also the most unevenly applied, since government procurement reaches large vendors more reliably than the long tail of integrators who actually deliver many public-sector deployments.

The pattern this fits

India's position is part of a broader divergence in how jurisdictions are approaching AI governance, and the divergence is not along the usual developed-versus-developing lines.

Malaysia announced plans for its first dedicated AI law in early 2027. Norway is preparing restrictions on AI glasses in public places. California passed thirteen AI bills, including one that removed a million-user threshold for obligations. Australia's ABC publicly rejected a copyright carveout on the grounds that its content has already been scraped. The UK has taken a light-touch approach and is under pressure to formalize it.

The common thread is that countries are writing rules calibrated to their own exposure rather than adopting a template. For AI developers, that means the compliance surface is becoming jurisdiction-shaped: a model deployed in five markets may face five different evaluation regimes, five different disclosure standards, and five different incident reporting timelines.

What to watch

Two things will determine whether India's framework becomes consequential or stays a statement of intent.

Whether it is binding on government procurement. A voluntary code with no procurement teeth is a set of principles. A procurement condition is a market requirement, and vendors will build to it.

And whether the multilingual evaluation requirement survives contact with cost. Testing across a dozen languages is expensive, and there will be pressure to scope it down to a handful. The number that ends up in the final framework will be a fairly precise indicator of how serious the commitment is.

The larger point stands regardless of how India's process resolves. The September pact assumed that a voluntary agreement among frontier labs could serve as the global baseline. India's response is a test of that assumption, and it is not the only country running the test. The frameworks being written now, in Delhi, Kuala Lumpur, Brussels, and Sacramento, are producing a world where AI compliance is a matter of where you deploy rather than what you built, and that is a materially harder operating environment for anyone shipping internationally.