Why AI safety depends on constitutional democracy

Safety depends on who holds power

Imagine an AI system that cannot launch a cyberattack, create a biological weapon or command killer robots. Put it in a constitutional democracy, with free and fair elections, protected civil liberties and the rule of law, and it could accelerate research and medicine, broaden access to education and support economic prosperity. It might even strengthen democracy by helping citizens be heard and making policymaking and public services more effective and accountable.

Under authoritarian rule, the same system can serve surveillance, censorship, propaganda and political repression. Botvinick points to China, Russia, Iran, Venezuela, several Persian Gulf states and other countries as places where these uses are already happening.

The distinction is not whether the AI is safe in a technical sense. It is whether the people living under it have political freedom. Botvinick argues that this makes reversing two decades of democratic backsliding urgent, including the gradual weakening of constitutional constraints in the United States under administrations of both parties.

The institutional strain

AI could also put democratic institutions under pressure. Changes in the labor market may deepen economic insecurity and create openings for demagogues across the political spectrum. Cyberattacks and other national-security crises could give governments grounds to claim emergency powers that they rarely relinquish. Using AI in the executive branch, meanwhile, could upset checks and balances and expand political surveillance.

The longer-term risk is more speculative. Economists Daniel Susskind and Anton Korinek are concerned that AI could replace most jobs. If that happens, citizens may lose influence over governments that no longer depend on taxing their labor. Botvinick points to oil-producing states as a warning: governments with less need to raise revenue from citizens often pay less regard to their rights.

His proposed first step is legal reform: limit executive power, revise emergency and surveillance laws, and strengthen due process by narrowing the circumstances in which the state-secrets privilege can be used. In the United States, that would include revisiting the National Emergencies Act and the International Emergency Economic Powers Act, and closing the “data broker” loophole in intelligence gathering.

As executive agencies adopt AI, Congress and the courts should also have tools powerful enough to keep pace with their oversight. The principle is old. James Madison wrote that “ambition must be made to counteract ambition”; the task now is to apply it in a digital environment.

Longer term, avoiding a broad loss of political influence will require renewed traditions of self-government and more public involvement in policymaking and oversight. AI could help make information easier to access and let citizens reason together, Botvinick suggests, like participants in digital town halls.

A responsibility for AI companies

That work belongs to politicians, civil society and citizens. Botvinick also puts responsibility on the AI industry. Major labs have spoken on specific policy questions, and some executives have argued that militaries and intelligence services should adopt AI so democracies can compete abroad. But, he says, those leaders have had far less to say about democracy at home.

Technology companies have spoken out before. In 2013, Apple, Google, Microsoft, Facebook, Twitter, Yahoo, LinkedIn and AOL jointly told the president and Congress that government surveillance had “tipped the balance too far in favor of the state and away from the rights of the individual — rights that are enshrined in our Constitution.”

In 2018, Microsoft urged Congress to limit government use of facial recognition, warning that it should not allow 2024 to become a page from the novel “1984.” In 2021, Alphabet, Amazon, Apple, Facebook, IBM, Microsoft and Twitter signed an open letter opposing proposed changes to state election laws and declaring support for democracy.

Many technology companies have since fallen quiet. Botvinick acknowledges a commercial reason: public statements can put share prices at risk. But if leading AI labs mean it when they describe safety as central to their mission, he argues, they should defend constitutional principles despite that risk.

I think the argument’s strongest point is also its least comfortable one: model safeguards cannot compensate for institutions that let governments use technology without effective limits. The announcement Botvinick wants from AI companies is not another safety pledge, but a clear position on the political conditions under which their systems will be deployed.

Matthew Botvinick is the author of the forthcoming book “AI and Political Freedom” and a senior fellow at Yale Law School. He has held several research roles in the AI industry, including at Google DeepMind, and currently works at Anthropic. In this article, he writes solely in his capacity as a Yale employee.