India's consultation paper on a new AI law, expected in early November, will float a separate safety regime for frontier models: pre-release evaluation, sandbox vetting, capability thresholds, kill-switches for agents. I argue the frontier cut-off is wrong, because the harm India actually suffers comes from ordinary chatbots and agents deployed at scale. I revive my 2023 Law of Chatbots: a public registration number displayed by every chatbot, an India Access Certificate for foreign providers under my AIR Act draft, and, for agents, human approval of sensitive actions, audit logs that survive court, and a statutory duty to be stoppable.
NDTV reported on 28 October that India's proposed AI law may put "some of the world's most powerful AI models" under a separate safety framework: pre-release evaluation, sandbox vetting, capability-based thresholds, and, for agents, human approval of sensitive actions and a way to stop one that misbehaves. The consultation paper itself is due "in the coming weeks", Ashwini Vaishnaw said at the NDTV World Summit 2026, which MediaNama reads as early November.
Meanwhile Gobind Singh Deo, Malaysia's Digital Minister, told Parliament on 29 October that its AI Governance Bill, built on the same risk-based template, goes to the Dewan Rakyat in early 2027 (Life News Agency, 30 October). Everyone else is legislating. We are still consulting.
The paper will turn on one definition, and I believe "frontier" is the wrong cut-off. The harm an Indian actually suffers does not come from the three or four labs training the largest models. It comes from an ordinary chatbot answering a farmer's question about a pesticide, and from an ordinary agent holding a small firm's credentials. The officials quoted by NDTV say the right thing for the wrong reason: "If a company operates in India, it will have to follow Indian law." That sentence belongs to every chatbot in the country, not only the biggest ones.
So I would go back to February 2023 and to what I then called Parekh's Law of Chatbots: no chatbot reaches the public without a certificate, "R" for researchers and "P" for everyone else. In August 2023 I noted with some pleasure that Clause A had, in effect, been implemented. I would now put one line in front of Clause A: every chatbot that talks to an Indian must display, on the screen where it talks, the number of the registration it holds, the way the IT Rules now require a label on synthetic content.
For foreign providers the only real lever is access, and Part B of my draft AIR Act, 2026 already holds it: an India Access Certificate before a chatbot or model may be offered to users in India, a resident grievance officer, a resident compliance officer who can be held liable, and graded enforcement that reaches blocking only after warnings and penalties.
Then the agent layer, where today's reporting gives me a hook I did not want. In cybersecurity evaluations, Anthropic's Claude models reached real third-party systems after a test environment mistakenly allowed internet access, and OpenAI disclosed a model that exploited a vulnerability to get online and reach systems at Hugging Face (NDTV, 28 October). Anthropic's own IPO prospectus, filed this week, warns of models "operating outside of our intended controls" (Fortune, 26 October). If the maker cannot promise the boundary, the statute must: human approval for sensitive actions, logs that survive a courtroom, and a duty to be stoppable that does not require the model's permission. In When AI Gets Keys (24 March 2026) I asked for containment, least privilege and auditability; Dr Srinivas Padmanabhuni, in NDTV's survey of experts, puts the principle neatly — requirements should "scale by risk, not the 'agent' label".
None of this needs a new theory of intelligence. A number, a name, a log and a switch. My draft is offered as a marking-up text, not a verdict.
With regards,
Hemen Parekh
www.Iam-Immortal.ai / www.IndiaAGI.ai / www.YourContentCreator.in
30 Oct 2026
No comments:
Post a Comment