From Slowing Down AI to Writing the Rules for AI

In late September, Netflix released Unabomber, a film about Ted Kaczynski. His violence cannot be justified, and it should never be romanticized as legitimate resistance to technology. But rejecting his methods does not require us to accept the opposite assumption: that every technological advance is progress, or that its direction and speed are inevitable. The pace and purpose of technology remain legitimate subjects of ethical judgment.

A voice from inside the industry made a similar point that month. On September 12, Anthropic CEO Dario Amodei published an essay, “We Must Pace the Frontier,” urging AI companies to slow the rate at which frontier models advance until safety measures can catch up. He did not call for a halt. He called for time to align and test models, and for third-party evaluators to confirm that it is happening. The call was notable because it came not from an outside critic but from the head of a leading AI company, and because it treats speed as an ethical choice, not merely a technical or commercial one.

Two days later, Microsoft AI published a draft Humanist AI Code of Conduct for its MAI models and opened it to public consultation. Its central commitments include meaningful human control, human flourishing, safety, and the protection of human judgment and agency, and it states that AI development should reflect perspectives beyond those of its developers.

The draft puts it plainly: “Any technology that cannot remain within human control should be rejected.”

Amodei’s proposal and Microsoft’s code approach the problem differently. One calls for a brake so that safety measures can catch up. The other attempts to write a rulebook defining the values and boundaries that should govern AI. Together, they raise a deeper question: Is slowing down enough, or must we also decide what kinds of AI should be created, what limits they should respect, and whose values should guide them?

Who Has the Power to Slow AI Down?

Individuals are often told to use AI responsibly. Users are expected to verify AI-generated information, protect personal data, avoid harmful applications, and decide thoughtfully when AI should or should not be used. These expectations are reasonable, but they can obscure differences in power.

The question is not whether individual users can slow AI down, but where in the chain of actors that power actually sits. A student who avoids generative AI for an assignment does not have the same influence as a company integrating AI into its products. A worker cannot always refuse a system adopted by an employer. A consumer may have little opportunity to avoid automated decisions embedded in education, employment, healthcare, banking, or public services.

Responsibility therefore cannot be assigned equally to everyone involved. Users, developers, technology companies, educational institutions, and governments possess different degrees of knowledge, authority, and control. Those differences should shape the responsibilities assigned to them.

Microsoft’s Code of Conduct recognizes some of this complexity by distinguishing among Microsoft AI, operators, users, and the wider public. Yet a corporate code raises further questions. Who gets to define human flourishing, meaningful control, and acceptable risk? Can a company adequately govern systems whose consequences reach far beyond the company itself? Is inviting public feedback equivalent to sharing decision-making power?

Writing rules for AI is itself an exercise of power.

Behind the Veil of Ignorance

John Rawls’s “veil of ignorance” offers one way to examine that power. Rawls asks us to imagine designing society’s rules without knowing what position we will occupy within it.

Suppose we do not know whether we will be AI developers, corporate executives, government officials, students, users, workers displaced by automation, or people harmed by an automated decision. We also do not know our income, education, disability status, technical knowledge, or access to advanced AI tools.

From behind that veil, what rules would we establish for AI? Would we allow companies to advance AI capabilities as quickly as possible and address harms only afterward? Would we permit AI to increase society’s overall efficiency if it substantially reduced some individuals’ autonomy? Would we rely primarily on voluntary corporate codes, or require broader public oversight?

The digital divide in the age of AI is not only about access to technology. It also separates those who can influence how AI is designed and governed from those who must live with decisions made by others. Meaningful public participation therefore requires more than commenting on rules after companies have already framed the available choices. It requires opportunities to question what AI is for, how much autonomy it should possess, which forms of development should be restricted, and whether certain systems should be built at all.

Who Should be in the Room?

Slowing down does not mean rejecting AI or halting technological development. It means creating enough social and intellectual space to ask who benefits, who bears the risks, what choices are being lost, and who should be held responsible.

That pause may give us time, but it does not answer who should set the pace or define the rules. Amodei’s proposal and Microsoft’s code are important precisely because they come from industry leaders. Yet this also reveals their limits: voluntary restraint is difficult when companies and governments view delay as surrendering a strategic advantage. It is the same logic I once described as a “dark forest” in the classroom: when no one can be sure others will hold back, restraint feels like falling behind. Only now the actors are labs and states rather than students. Decisions about when to apply the brake therefore cannot be left solely to those who benefit most from speed. Three things would make the answer less abstract.

First, verification by someone other than the developer. This institutional verification matters because, as my own interview study suggested, users differ widely in how thoroughly they verify AI outputs. System-level safety cannot depend on individual vigilance alone. Amodei’s call for independent evaluators with real access matters because it moves judgment outside the lab, and OpenAI CEO Sam Altman has said his company would do the same. The test is whether those evaluators are truly independent and whether their findings are made public.

Second, consultation that can change outcomes. Microsoft opened its draft to a six-week public consultation and asked for input on how “human flourishing” should be defined. Whether the final text shows what the company heard and what it changed will reveal whether consultation is participation or ceremony.

Third, a seat for those who will live with the consequences: educators, workers, patients, and students, not only engineers and executives. Reading the draft and submitting a comment is a modest start, but it is exactly the kind of participation we would demand from behind the veil of ignorance.

This is why technology ethics belongs in the classroom. Its purpose is not to teach students to support or oppose technology, nor to hand them a list of acceptable uses. It is to help them look beyond the language of efficiency, convenience, and innovation and ask how technology reshapes human autonomy, power, and the choices open to society. Technological development is not an autonomous force. People choose what to fund, build, deploy, regulate, and accept as inevitable.

We may need both a brake and a rulebook. What matters just as much is who is in the room when they are designed. That question should be asked in every classroom, and answered in public.