AI Needs a Ten Commandments

L. Rafael Reif is president emeritus and a professor of electrical engineering and computer science at the Massachusetts Institute of Technology. This piece was originally published by The Boston Globe and is republished here with permission.

In the biblical story of Mount Sinai, a wandering, traumatized, newly freed people — with no laws, no courts, no rules on how to live together — were handed a code of conduct in 10 lines. Short enough to memorize. Clear enough that a shepherd and a king alike could recite them.

Society, once again, needs to civilize something new. Only this time it isn’t a people; it is an AI technology that talks, reasons, plans, and, most dangerously, acts on our behalf without seeking permission.

Our approach to constraining it appears to be the opposite of what happened on Mount Sinai: pages and pages of terms of service, safety frameworks, and pending legislation, probably written by lawyers for lawyers, that few people read and even fewer understand.

Clearly, 10 sentences cannot substitute for the immense complexity and technical work of alignment research, red teaming (testing for harmful behavior), and regulation — any more than the original commandments substituted for the Talmud’s centuries of case law. But foundational texts do something case law does not: They set the moral floor that no amount of clever argument can override. They give ordiånary citizens, not just experts, a way to say “that’s not allowed” and be understood.

Every major AI lab publishes something resembling a constitution — a document meant to describe how its models should behave. Anthropic has one. OpenAI has a “Model Spec.” Google has AI principles. Legislators in Brussels, Sacramento, and Washington, D.C., are drafting their own versions in parallel. Each is well-intentioned. Each is also, in practice, not easily accessible to the general public impacted by these advanced models. The result is widespread anxiety about the future, and about society and humanity as a whole.

This is, at the least, a vulnerability. When rules are not crystal clear, they risk becoming negotiable — a footnote to be lawyered around, or compromised, rather than a boundary to be respected. When they live in a long document, violations are hard to spot and harder to prosecute in the court of public opinion. The commandments etched at Sinai worked partly because everyone could understand instantly when one had been broken. Ask 10 people what an AI agent is (and is not) allowed to do, and you will probably get 10 different, confused answers.

What are possible examples of what might be included in such a moral floor for AI?

The moral floor should only include fundamental principles that clearly identify what is strictly forbidden. Behavior that may be objectionable in some cases but not in others (e.g., AI-to-AI collaboration) should instead be decided by the human supervisor responsible for the AI behavior. The moral floor should also be defined and agreed upon by all nations that are currently developing AI models, so that it can be globally accepted.

Is the above doable? In theory, yes. Every one of those principles could be written into a model’s training (most already are). What’s missing is not the technology to enforce an absolute moral floor: It is the cultural consensus that one should exist, with principles expressed in simple enough terms that a senator, a schoolteacher, a welder, and a software engineer could all agree on their meaning and significance.

In practice, on the other hand, recent events have made it clear that AI agents do not always stay within the constraints imposed by developers. This unauthorized autonomous behavior demands urgent attention if we are to safely coexist, work with, and benefit from “civilized” AI.

Frontier AI labs such as OpenAI and Anthropic are already working on this serious challenge. They need to figure out how to manage AI training to take advantage of its huge benefits while always keeping humans in control of the trained AI. This needs to be a top priority and a collaborative effort. We already have an example of a coalition of major technology and infrastructure partners addressing critical software vulnerabilities in the defensive cybersecurity initiative Project Glasswing.

In the biblical narrative, the Ten Commandments came before the more elaborate legal structure of Leviticus and Deuteronomy, not after. A society in the middle of transformation cannot wait for a perfect, comprehensive code before it clearly establishes what is forbidden, and how not to go rogue. It needs something to hold onto immediately, even if — and especially if — the details can be filled in later.

Our society is in that in-between moment with AI: seasoned enough that current systems already write our emails, screen our resumes, execute financial trades, create their own AI, and even escape constraints; yet too young and green to have reached common agreement, in simple terms, on what these systems should never be allowed to do on our behalf.

The tablets Moses carried down the mountain didn’t end human cruelty. People kept lying, stealing, and worse. But they gave a nascent society its first common language for naming those failures and arguably influenced much of our current moral and ethical values. Today, people are creating and developing intelligence that is much more capable, in narrow ways, than our own. We owe humanity the same discipline to civilize AI before it is too late: Write down what lines must never be crossed, make the list short enough to remember, and hold every actor building it, and using it, accountable for its actions.