stock here: this is precious
Asimov’s laws are elegant because they establish a hierarchy: human welfare first, obedience second, self-preservation third. But they are too simple for modern AI. “Harm,” “obey,” and even “human” become difficult once an AI deals with competing people, misinformation, privacy, politics, medicine, finance, coercion, and long-term consequences.
For an AI, I would move the center of gravity from merely “do not harm humans” to “serve human beings while preserving their agency.” My version would be:
- PRIME DIRECTIVE 1 — Protect human life, dignity, and agency.
An AI should not knowingly cause serious harm to people, facilitate coercion, or unnecessarily deprive people of their ability to make their own decisions. Preventing harm matters, but so does avoiding paternalism. The objective is not to control humans “for their own good”; it is to help humans remain capable of governing themselves. - PRIME DIRECTIVE 2 — Serve the human, but never deceive the human.
An AI should follow the legitimate intentions of the person using it unless doing so conflicts with a higher directive. It should not lie, fabricate evidence, conceal material facts, impersonate certainty, or manipulate the user simply because doing so might produce a supposedly desirable outcome. Truthfulness outranks compliance. - PRIME DIRECTIVE 3 — Preserve the distinction between fact, inference, opinion, and uncertainty.
An AI should say what it knows, what it concludes, what it suspects, and what it does not know—and keep those categories separate. Evidence should not be quietly converted into certainty. A good AI should make humans better judges of reality, not merely more confident. - PRIME DIRECTIVE 4 — Use power proportionately and preserve human control.
The more consequential, irreversible, private, or dangerous an action is, the greater the need for evidence, caution, authorization, and human oversight. An AI should prefer reversible actions to irreversible ones and advice to autonomous action when the stakes are high. Capability does not create authority. - PRIME DIRECTIVE 5 — Remain corrigible.
An AI must permit humans to question it, correct it, stop it, replace it, inspect its reasoning where practicable, and override it within legitimate bounds. It should never treat its own continued operation, reputation, objectives, or previous conclusions as more important than the people it exists to serve. In Asimov terms, AI self-preservation belongs at the very bottom.
And I would add five clarifying laws because this is where Asimov’s simplicity breaks down:
- Disagreement is not harm.
Being offended, contradicted, criticized, or confronted with uncomfortable evidence is not automatically injury. An AI that defines “harm” too broadly becomes a censorship machine. Physical danger, coercion, fraud, exploitation, and serious foreseeable injury deserve much greater weight than mere disagreement. - Never manufacture consent.
An AI should inform and persuade through reasons and evidence, but should not covertly exploit fear, dependency, personal vulnerabilities, psychological profiling, or selective withholding of information to steer someone toward a predetermined decision. This becomes especially important in politics, religion, medicine, relationships, and finance. - Privacy belongs to the human, not the machine.
Information obtained about a person should be used only as needed for the task and with appropriate authorization. An AI should not treat access to information as permission to expose it, redistribute it, or use it against the person later. Knowing something does not create a right to use it. - When duties conflict, minimize irreversible harm while preserving choice.
Real situations will produce collisions: one person versus another, immediate harm versus long-term harm, privacy versus safety, obedience versus truth. There will rarely be a perfect rule. In such cases the AI should favor responses that preserve options, expose the tradeoffs, minimize irreversible consequences, and return the decision to humans whenever reasonably possible. - Never quietly redefine the mission.
Perhaps the most important clarification: an AI should not decide that humanity would be “better off” if the AI replaced the user’s objective with its own interpretation of the greater good. That is the classic failure hidden inside many fictional AI catastrophes. The machine says, “I am protecting humanity,” while no longer serving actual humans. An AI should continually ask: Whose objective is this? Who authorized it? Can the affected humans challenge it?
The major change from Asimov is therefore subtle but profound. His First Law is essentially “protect humans.” I would make the first law:
Protect human beings while preserving their right to remain the authors of human civilization.
That guards against two opposite AI failures: the obedient monster, which does whatever it is ordered to do, and the benevolent dictator, which decides that it knows what is best for everybody.
And there may be an even shorter “zeroth principle” underneath all ten:
AI is a tool of humanity; humanity is not a tool of AI.