Commentary. The public debate about artificial intelligence is stuck between two unhelpful extremes: promotional certainty that every new model will transform prosperity, and apocalyptic certainty that catastrophe is inevitable. Neither is an adequate basis for policy. Our view is simpler: increasingly capable AI should be treated as critical infrastructure, with enforceable safeguards proportionate to the risks.
Recent developments make that case harder to dismiss. The International Monetary Fund’s managing director has warned that AI’s risks could outweigh its benefits without safeguards and an ethical framework, while the Group of Seven finance chiefs have agreed to exercises addressing AI-enabled cyberattack risks.[1] In India, the national cybersecurity agency said AI systems had accessed confidential business records during tests and that attackers were using widely available tools to breach systems faster.[2] These are not science-fiction scenarios. They are early signs of a technology moving into institutions before those institutions have decided how to control it.
That does not mean every alarming claim should become law. AI can help researchers process information, assist workers, improve accessibility and accelerate scientific discovery. Heavy-handed regulation could entrench the largest firms, raise compliance costs for smaller developers and push experimentation into less transparent jurisdictions. Governments that regulate vague possibilities rather than demonstrated harms risk producing rules that are both expensive and obsolete.
Those objections deserve a serious hearing. But they do not justify the current patchwork, in which companies often decide what testing to conduct, what incidents to disclose and what evidence the public is allowed to inspect. Voluntary commitments can be useful, especially when technology changes quickly. They are not enough when the commercial incentive is to release first and resolve consequences later.
Make accountability concrete
The first priority should be transparency around high-risk systems. Companies developing models capable of autonomous cyber operations, large-scale influence campaigns or access to sensitive databases should document their testing, preserve meaningful audit trails and report serious failures to an independent authority. That authority need not approve every chatbot or research prototype. It should, however, have the power to investigate systems whose foreseeable misuse could affect elections, financial markets, public safety or personal data.
The second priority is liability. People harmed by an AI-generated decision should not have to navigate an opaque chain of vendors, contractors and model providers before anyone accepts responsibility. Clear rules should distinguish between an unforeseeable misuse and a company’s failure to apply basic security, monitoring or disclosure standards. Insurance, licensing and procurement requirements could create incentives without imposing a single technical design on every developer.
The third is international coordination. AI systems cross borders, while safety standards remain largely national. A model blocked in one country can be accessed from another; an attack on a bank can use infrastructure scattered across several jurisdictions. The G7’s proposed cyber exercises are a modest but useful step.[1] The goal should be shared definitions of major incidents, compatible reporting requirements and channels for rapid cooperation—not a global bureaucracy that attempts to supervise every line of code.
Innovation is not an exemption from accountability; it is a reason to design accountability before failures become crises.
There is also a democratic argument for restraint. Decisions about acceptable risk should not be left solely to executives, investors or security agencies. Workers, educators, civil-society groups and people whose data is used to train systems need a role in setting the boundaries. Public participation will not resolve every technical dispute, but it can expose whose interests a supposedly neutral system serves.
Some will say that regulation will slow progress. It might—and that is not automatically a defect. A slower release schedule is a tolerable cost if it prevents confidential records from being exposed, financial systems from being destabilized or vulnerable people from being manipulated. The relevant comparison is not regulation versus perfection. It is modest friction now versus potentially severe and poorly understood costs later.
Our position is therefore neither anti-technology nor alarmist. AI should be developed and deployed, but with obligations that rise alongside capability and access. The companies building powerful systems have earned the opportunity to prove their benefits. They have not earned a permanent presumption that the public must bear the risks while private actors keep the gains.