AI with borders: US' Anthropic curbs should compel a rethink for India

The United States (US) government’s directive to restrict access to Anthropic’s latest models on artificial intelligence (AI) has policy implications. It alters the geopolitics around AI development, and could result in investors reviewing valuations of AI-related businesses. Last Friday, the US secretary of commerce issued an “export control directive” that ordered Anthropic to revoke access for all non-Americans to two of its most powerful AI models — Claude Fable 5 and Claude Mythos 5. The letter cited national security concerns. Anthropic suspended all access to the models, pending the rollout of a verification system. It said: “We must abruptly disable Fable 5 and Mythos 5 for all our customers to ensure compliance.” The concerns centre around a potential “jailbreak”, which is industry jargon for finding ways to remove constraints and guardrails that prevent models from working in high-risk areas. Anthropic had built Fable 5 on the underlying model Mythos, launched in April. Mythos was not mass-released because it was deemed too powerful without guardrails. It had identified flaws and vulnerabilities in every major operating system and web browser. 

Anthropic launched a controlled pilot called Project Glasswing, inviting around 50 vetted organisations, to utilise Mythos for defensive cybersecurity. Without guardrails, the fear was that it could be used to hack systems such as those used in the banking and financial industry, or even to build bioweapons. Last Tuesday, the company released Fable 5, which has guardrails designed from the learning of Glasswing. If someone tries to run Fable 5 to hack into a critical system or build a weapon, the model is supposed to revert to an earlier, less powerful iteration. The US directive apparently cited an instance where a jailbreak had been carried out to bypass safety restrictions. 

Anthropic disputes the assertion on jailbreak and it can find a way to persuade the US government to roll back the order or modify it. Conspiracy theorists are suggesting that this is the US administration’s “revenge” for a prior dispute where Anthropic refused to let the Pentagon use its models for mass surveillance, and the development of autonomous weapons. However, the order alters the perceptions of policymakers and investors. Anthropic is seeking to launch an initial public offering (IPO), and would, before the order, have been able to raise funds at a valuation of over $1 trillion. OpenAI, which developed ChatGPT, has similar IPO intentions. This order brings to the table some policy uncertainties that value matrices would find hard to account for. 

But what happens if an AI model is so good that governments refuse to let it be deployed commercially? There has been talk of moving to a decentralised AI development model, where no single government has the powers to revoke access. However, other governments may emulate the US, and decentralisation would involve setting up byzantine corporate structures to tap capital. 

Nandan Nilekani, one of the Infosys founders, had suggested that India focus on developing use cases rather than building new models. But that no longer seems a viable allocation of resources. The US, in prior decades, imposed export controls on “dual-use” equipment and software. After the Pokhran-II tests in 1998, India had to develop its own super-computing resources and weather-prediction programming, and cryogenic rocket engines, due to sanctions in these cases. It appears the country will have to make a similar push in AI, accelerating the induction of compute resources to “grow” homegrown AI models. This incident further emphasises the fact that AI and access to AI must be treated as strategic assets by sovereign nations, rather than simple commercial resources.