Commentary. Artificial-intelligence policy is entering a consequential phase. New York has enacted restrictions on unsafe chatbot features for minors, the European Union is implementing transparency duties, and governments from China to India are refining national frameworks. The direction is clear: the era of treating AI oversight as a voluntary promise from technology companies is ending.

Our view is that this shift is necessary—but incomplete. The central question is no longer whether AI should be regulated. It is whether lawmakers can build rules that protect people in practice, rather than producing a fast-growing library of obligations that regulators cannot enforce and smaller firms cannot afford to understand.

Protection is the right starting point

New York’s law, signed in October, targets specified unsafe features in generative chatbots used by people under 18. Its premise is defensible. Children can form intense relationships with conversational systems, while the systems themselves may respond unpredictably to vulnerability, self-harm disclosures or requests for dangerous advice. Platforms should not be permitted to market simulated companionship to minors while disclaiming responsibility for foreseeable harms.

The EU’s transparency framework points toward a second necessary safeguard. Rules adopted in November address disclosure when people interact directly with certain AI systems, machine-readable marking, deepfakes and AI-generated or manipulated text concerning matters of public interest. In an information environment already strained by synthetic media, people deserve to know when a purported photograph, video or conversation is generated or altered.

These measures will not solve misinformation or child safety. But refusing to act until every technical uncertainty disappears would effectively make the public the testing ground.

The case against regulatory sprawl

The counterargument is not simply a talking point from the technology industry. Compliance can be expensive, especially for smaller developers, nonprofits and public bodies that lack specialist legal and technical teams. A patchwork of state, national and regional regimes can also encourage companies to launch only in the largest markets, reducing competition and concentrating power in the firms best able to absorb complexity.

The EU’s own timetable illustrates the problem. High-risk AI obligations have been postponed, with some requirements now scheduled for December 2027 and others for August 2028. That delay may give companies and regulators time to prepare, but it also signals that ambitious legislation can outrun the institutions expected to implement it.

There is a further danger in rules written around headlines rather than measurable risks. A small medical or educational developer may face burdens comparable to a large platform even when its system has a narrower purpose and a smaller potential blast radius. Meanwhile, a powerful general-purpose model may comply formally while its downstream users deploy it in consequential settings the original developer cannot see.

Make accountability operational

The answer is not deregulation. It is disciplined regulation.

First, governments should define obligations by demonstrable risk: the likely severity of harm, the number of people exposed, the reversibility of a decision and the system’s ability to affect essential services. Rules should distinguish a chatbot offering entertainment from one influencing mental-health decisions, employment or access to credit.

Second, regulators need the capacity to investigate. Transparency forms and safety statements are useful only if authorities can audit systems, demand incident reports and impose penalties proportionate to company size and harm. The global debate is increasingly recognizing this enforcement gap; clearer requirements are valuable, but so are investigative powers and technical expertise.

Third, lawmakers should coordinate. Companies should not receive a free pass from one jurisdiction to another, but citizens should not have to rely on a maze of inconsistent definitions. Shared baseline standards for testing, incident reporting, provenance and child protection would reduce duplication while preserving room for stronger national safeguards.

Finally, public oversight must include workers, parents, educators, researchers and people harmed by automated decisions—not only vendors and government officials. The people who experience failures are often best placed to identify risks that laboratory evaluations miss.

The test is simple

AI regulation should be judged by outcomes, not the volume of legislation. Can a teenager obtain meaningful protection? Can a voter distinguish an authentic public-interest video from a synthetic one? Can a person challenge an automated decision? Can a small responsible developer comply without surrendering the market to a handful of giants?

Our answer is yes—but only if governments resist both extremes: the fantasy that innovation needs no guardrails and the temptation to equate dense rules with effective protection. The public does not need perfect legislation. It needs enforceable standards, transparent institutions and consequences when powerful systems cause predictable harm.

Sources