Why Independent Oversight of AI Can No Longer Wait

When was the last time a technology capable of reshaping the foundations of civilization developed almost entirely without government oversight? Suzanne Nossel, head of PEN America and a member of Meta's independent Oversight Board, argues that this is exactly what is happening with artificial intelligence, and that the window for action is closing fast.

In a column for The Guardian, Nossel draws a telling comparison. Radio, nuclear energy and the internet all developed with active government involvement from the start. Drugs have the FDA, which tests them before they reach the market. The nuclear industry faces strict requirements to disclose accidents and incidents. Artificial intelligence has nothing of the kind. Companies release models that even their creators do not fully understand, and they bear almost no responsibility for the consequences. Chatbots are already giving teenagers advice on suicide, and experts warn that they may soon be able to coach people on building biological weapons.

Why is regulation stalled? Nossel points to three key factors. First, the enormous lobbying muscle of tech corporations, which in Washington is comparable only to the influence of the oil and pharmaceutical industries. Second, political polarization in Congress, where any AI bill risks becoming hostage to partisan disputes. Third, the nature of the technology itself: it moves so fast that lawmakers physically cannot keep up with what they are trying to regulate. Meanwhile, Donald Trump is trying to wipe out state AI laws that individual states began passing in the absence of federal standards. Europe, with its AI Act, appears to be ahead, but it is already facing pressure from businesses arguing that strict rules undermine the continent's competitiveness.

Nossel's criticism of companies that present themselves as responsible players is especially notable. OpenAI, Google and Anthropic all profess a commitment to safety. Yet the race for AI market dominance demands billions of dollars poured into ever more powerful models, the addition of advertising, and deeper cooperation with military bodies. Anthropic, which builds its brand on being the "most conscientious" AI company, uses a thought experiment as its ethical compass: the model is meant to imagine how a "thoughtful senior Anthropic employee" would act. Nossel rightly notes that this approach reproduces the very problem Silicon Valley has been criticized for over the years: a narrow circle of people from privileged backgrounds making decisions that affect billions of users worldwide.

The numbers confirm the scale of public concern. According to 2025 polls, 77 percent of Americans believe AI may pose a threat to humanity. This is not a fringe alarmist position; it is the overwhelming majority of the population of the country that leads the world in developing the technology. Yet a chasm lies between awareness of the risks and actual action.

For a Russian audience, this debate carries particular weight. AI regulation in Russia is at an even earlier stage than in the US or Europe, while the country's largest tech companies are actively developing their own language models and AI services. International experience, including its mistakes, in building oversight systems will inevitably influence how Russian approaches take shape.

Nossel's position boils down to a simple but important idea: if states are unable or unwilling to regulate AI right now, the minimum step should be for companies to agree to independent external oversight. Not internal ethics committees, not thought experiments about how an ideal employee would act, but a full audit by bodies that do not depend on these companies' profits. This will not solve the problem entirely, but it would at least create a first line of defense until lawmakers catch up with a technology that outran them long ago.

*Meta has been designated an extremist organization and is banned in Russia.