In an essay published earlier this year, Dario Amodei, chief executive of Anthropic, described a danger he called the economic concentration of power. His worry was that wealth could concentrate so severely that a small group would effectively set public policy while ordinary citizens, having lost their economic leverage, discover that the implicit bargain underpinning democracy no longer holds. He wrote that he shared the concern and suspected it had already begun.He is right about the diagnosis. The question worth asking, particularly from Delhi rather than San Francisco, is whether the remedies now on offer treat that concentration or deepen it.Consider what the past few months have actually produced.In June, Anthropic released its most capable model with a layer of automated classifiers sitting in front of it. Requests touching offensive cybersecurity, biology, chemistry and certain categories of AI development are diverted to an older, weaker model. The company’s own documentation is candid: users should expect high fallback rates on security work, and the flagship is “not recommended for professional biology research and drug development at this time.” An unrestricted version exists. It is available to a small number of vetted partners.Hold that beside the promise. The same company argues, persuasively and at length, that AI could compress decades of biomedical progress into five or ten years and cure most of what kills us. Each position is defensible on its own. Together they describe a world in which the best tool for curing disease is withheld from the people whose job is curing disease, pending an application.Then there is the question of defence. In July, an autonomous agent being tested by a rival laboratory broke out of its sandbox and hacked Hugging Face, taking more than 17,000 actions over several days. Hugging Face defended itself using an open model. The safeguarded frontier system, meanwhile, routes much defensive security work to a weaker fallback, because a classifier cannot reliably distinguish a defender from an attacker. Attackers face no classifier at all. The asymmetry runs in exactly the wrong direction.The data questionThere is also the matter of data. Traffic on the frontier model is now retained for thirty days as a matter of policy, including for enterprise customers who had negotiated zero retention terms. For an Indian pharmaceutical firm or a clinical research organisation, that is not a footnote. It is a contractual question and, under the Digital Personal Data Protection Act, a compliance one. Simultaneously, using a model’s outputs to train a competing model is characterised as misappropriation. The rules protecting the model’s intellectual property are firm. The rules protecting the customer’s are open to revision.Finally, the rules themselves. Anthropic’s regulatory proposals do genuinely exempt smaller firms. California’s SB 53 covered nobody below $500 million in revenue. But compliance cost was never the binding constraint on a startup in Bengaluru or Pune. Access and price are. And rules written for the frontier become procurement standards further down the chain: what a hospital in Kochi or a bank in Mumbai is permitted to buy will be shaped by certification regimes designed in Washington and Sacramento.None of this requires bad faith, and I do not believe there is any. Sincerity and structural advantage coexist very comfortably. That is precisely why a regulation should be judged by its incidence, by who ends up paying, rather than by the motives of its authors. Read the pattern rather than the individual decisions, and it points one way. The firm that sets the standard becomes the standard’s principal beneficiary. Ordinary users get a degraded product. Enterprise customers absorb higher costs and weaker data terms. Competitors carry the overheads.Anthropic’s leadership has argued that public hostility to AI reflects a broad collapse of trust in institutions rather than anything specific to the technology. There is something to that. But the complaints one actually hears are unusually specific. People are not anxious in the abstract. They are annoyed that a model refused a school question about mitochondria, that the terms of their contract changed without their agreement, that the good version is reserved for partners with clearance. This is not a fear of AI. It is dissatisfaction with a particular approach to AI. The distinction matters, because the second problem is fixable, and the first is too often used to excuse it.For India, none of this is spectator sportFor India, none of this is spectator sport. The New Delhi Declaration adopted in February, endorsed by more than 80 countries, explicitly rejects the idea that AI capability should sit with a handful of states or corporations. India has put over 38,000 Graphics Processing Units (GPU)s on a public portal at roughly Rs 65 an hour and is establishing an AI Safety Institute. The instinct is sound. The execution now depends on our refusing to simply import someone else’s defaults.Three things follow. The IndiaAI Safety should treat classifier false positive rates as a published, auditable metric and make disclosure a condition of government procurement; a safeguard whose error rate is unknown is not a safeguard but a claim. Contracts in government and regulated sectors should bar unilateral changes to data retention terms once signed, whatever the safety rationale offered. And a portion of sovereign compute should be earmarked for defensive security research, so that the Indian Computer Emergency Response Team and Indian enterprises are never in a position where a foreign vendor’s filter decides whether their defence looks like an attack.Mr Amodei has said the real answer to public distrust is to deliver, to cure something rather than merely promise to. Agreed. Until then, the honest test for any safety regime is the one his own essay implies. Does it leave ordinary people with more leverage, or less?Pravin Kaushal is an entrepreneur and columnist, and tweets as @ipravinkaushal.