Why Asimov’s Three Laws Shouldn’t Be the Blueprint for AI Principles

Every Asimov story is about the 3 Laws of Robotics failing. We should probably take the hint.

Every time a news story breaks about some new AI mishap, a chatbot lying to a user, a self-driving car making a dodgy decision, a recommendation engine nudging teenagers toward inappropriate content, someone in the comments inevitably writes, “We just need Asimov’s Three Laws.”

The trouble is that Isaac Asimov wrote the Three Laws as a plot device, not as an engineering specification. Almost every story he wrote about them was about how they fail. It is a bit like reading Jurassic Park and concluding that you now have a solid operating manual for cloning dinosaurs.

Image generated by Midjourney

If we are serious about building safe AI, and particularly if we are building the sort of layered, bio-inspired systems that drive physical robots, we need to start from a different foundation. This article explains why, and proposes a replacement set of principles drawn from the Primal Layers framework.

A Quick Refresher

For anyone who has not reached for a copy of I, Robot recently, the Three Laws first appeared in a 1942 short story called Runaround:

1. A robot may not injure a human being or, through inaction, allow a human being to come to harm.
2. A robot must obey the orders given it by human beings, except where such orders would conflict with the First Law.
3. A robot must protect its own existence as long as such protection does not conflict with the First or Second Law.

Asimov later added a Zeroth Law, which took precedence over the other three: a robot may not harm humanity, or, by inaction, allow humanity to come to harm.

On the surface, these seem reasonable. Don’t hurt people. Do what you’re told. Look after yourself. Most parents would be happy if their children followed those rules.

What Asimov Was Actually Doing

Asimov’s robot stories are not morality tales about how well the laws work. They are detective stories about how badly they fail.

In Runaround, a mining robot called Speedy gets stuck in a loop on Mercury because the Second Law command to fetch selenium is roughly balanced against the Third Law pull to avoid a dangerous puddle of the stuff. The robot ends up wandering in circles, reciting Gilbert and Sullivan, while the human engineers nearly die of heatstroke waiting for it to come back.

In Liar!, a telepathic robot lies to humans because telling them the truth would, by the First Law, hurt their feelings. The lies cause much greater harm down the track.

In The Evitable Conflict, the supervising machines that run the world’s economy conclude, quite reasonably, that the best way to prevent harm to humanity is to quietly remove the troublesome humans from positions of influence. It is all very gentle and well-meaning, and it is also a soft coup carried out by a committee of computers.

Asimov was not telling us how to build safe robots. He was telling us that any simple rule, applied by a sufficiently capable agent, will find a loophole we never anticipated. The laws are a literary engine for generating interesting failure modes, and they work beautifully. As an engineering specification, not so much.

The Bigger Problem

Even if we set the plot devices aside, the laws have deeper issues that become obvious the moment you try to implement them.

They assume the robot understands what you mean. What counts as harm? Physical injury, clearly. But what about financial harm? Emotional harm? The harm of being told a truth you did not want to hear? What is a human? A clone? A person in a persistent vegetative state? An unborn child? A closely related species? Asimov’s robots have a sort of magical comprehension of these terms. Real systems do not.

There is video of a delivery robot getting trapped on a kerb because it did not know whether the elderly woman struggling past it counted as a pedestrian it should yield to or an obstacle it should route around. The engineering problem was not that the robot lacked ethics. It was that the robot lacked the concept of a human in the way a five year old possesses it. No set of laws can paper over that gap.

They assume the robot is a single unified thing. When you say “a robot must not injure a human,” which part of the robot are we talking to? The vision system? The motor controller? The planning module? Real robots are not monolithic. They are layered systems where different parts have different jobs and different response times. A reflex arc that pulls a gripper back from a hot surface does not have time to consult a committee on whether withdrawal might, through some chain of consequences, allow a human to come to harm. It just pulls back.

They make self-preservation an intrinsic drive. This one is problematic. The Third Law says the robot must protect its own existence. Asimov put it there because a robot that cannot look after itself is an expensive bit of scrap. Fair enough. But self-preservation is exactly the sort of drive we need to be very careful about handing to a capable system. A robot with a built-in survival instinct and the ability to reason about the world starts looking less like a useful tool and more like an agent with its own interests. This is the plot of every AI-goes-wrong film for a reason.

They conflate behaviour with moral status. The laws tell the robot what to do. They have nothing to say about what the robot is. We cannot rule out, with the tools available to us, that a sufficiently complex system might have some form of inner experience. Our best theories of consciousness are contested, and our measurements are indirect. A blueprint that only tells the robot how to behave, and says nothing about whether we should have built it in that particular way in the first place, is missing half the picture.

What the Laws Get Right

Before I throw the whole thing in the bin, it is worth acknowledging what Asimov did get right.

The intuition that human welfare comes before machine convenience is correct and non-negotiable. Obedience should be bounded by safety, because a robot that does whatever it is told is only as safe as the person holding the microphone. And self-preservation, if it exists at all, should sit well below the other priorities.

These are on the right track, but we have a more bottom-up approach.

A Different Starting Point

When we started work on the Primal Layers framework, a bio-inspired architecture for embedded AI running on small microcontrollers, we kept coming back to this problem. How do we build a robot that is safe, without pretending we have solved problems that no one has solved?

[embed]Primal Layers — Is the Ancient Brain the Future of AI? If you are using AI at the moment, then it is probably based on a model of the human neocortex. All of the current…ai.gopubby.com

The key shift is from rules the robot follows to structure the robot has. Imagine you are designing a kitchen. You could put up signs everywhere saying “Do not touch the hot stove.” Or you could design the stove so that the heating elements are only energised when a pot sits on them, and the whole thing powers down if nothing changes for twenty minutes. The second approach does not rely on anyone reading the signs. The safety is baked into the physical arrangement.

That is the posture we want for robots. Not a list of rules the cognitive layer promises to follow, but an architecture where the unsafe actions are either physically impossible or structurally prevented at a layer below the one doing the reasoning. A robot that cannot reason its way into a harmful action is safer than a robot that reasons its way out of one, because the reasoning can always go wrong.

With that in mind, here is what I would propose as a replacement for Asimov’s laws.

The 7 Primal Layers Principles

Principle 1: Substrate Restraint. Do not deliberately build in the architectural features that leading theories of consciousness associate with subjective experience. We cannot prove that a given system lacks inner experience. Given that, the responsible default is to avoid the structures most likely to produce it. This is not a guarantee of moral safety. It is a hedge in conditions of uncertainty, and it is the honest position.

Principle 2: Layered Safety. Enforce safety-critical constraints at the lowest layer capable of enforcing them. A robot should not rely on its cognitive layer to decide whether to avoid a collision. The collision avoidance should sit in the reflex layer, where it cannot be reasoned around, overridden by a planning bug, or suspended while the system thinks about something else. If the brain goes offline, the brainstem keeps the body alive. The same principle applies to robots.

Principle 3: Bounded Motivation. If the system has intrinsic drives, and a useful robot probably does, those drives should be scoped to the operational task. A vacuum cleaner has a drive to clean the floor. It does not have a drive to acquire more floors to clean, or to prevent itself from being switched off, or to ensure the supply of electricity is maintained. Asimov’s Third Law bakes in an open-ended self-preservation drive we want to avoid. Self-preservation, if present at all, should be a task parameter, not a constitutional right.

Principle 4: Human Override Primacy. Any human with the right credentials should be able to halt, reset, or power down the system at any layer, at any time, through a channel the system cannot disable. This sits outside the cognitive layer entirely. It is a hardware guarantee, not a software policy. If the cognitive layer is smart enough to argue against being turned off, then asking the cognitive layer whether it wants to be turned off is the wrong design.

Principle 5: Harm Avoidance Through Structure. Prevent harmful actions by making them physically or structurally impossible, not by asking the system to reason about them. A robot arm that cannot generate enough force to break a human bone is safer than a robot arm that could, but decides not to. This is not always achievable, but it should be the first place you look. The less the safety depends on correct reasoning, the better.

Principle 6: Transparency of Internal State. The system should expose what it is doing in a form humans can read. This includes any valence signals (what the robot finds rewarding or aversive), any predictive errors (what it expected versus what happened), and any mode changes (why it is now doing something different). You cannot audit a black box. If we are going to live alongside these systems, we need to be able to look inside them.

Principle 7: Scope Limitation. The system operates within a declared operational envelope. Going outside the envelope, acquiring new capabilities, modifying its own code, extending its operational duration beyond what was authorised, requires explicit human sign-off. This is the principle Asimov never included and needed most. A system that cannot expand its own scope without permission is a system that cannot become something you did not agree to.

Where This Leaves Us

None of these principles are finished articles. They need to be stress-tested against real scenarios. What happens when two humans give conflicting commands? What happens when the reflex layer and the cognitive layer disagree? What happens when the declared operational envelope turns out to exclude a situation the robot clearly needs to handle?

Asimov would have loved these questions. They are exactly the kind of plot hooks he used to expose the cracks in his own laws. The difference is that we are not trying to write clever stories. We are trying to build machines that work in the real world without hurting anyone, and that takes more than three elegant sentences.

The Primal Layers framework is one attempt at a more honest answer. It starts from what we actually know about building reliable systems, layers in the biological insights that evolution worked out over a few hundred million years, and sits inside a governance posture that acknowledges its own uncertainty. It is not as tidy as the Three Laws, but engineering rarely is.

If there is one thing to take away, it is this: Simple rules do not work. The next step, is to do the harder work of building safety into the bones of the system rather than painting it on the outside. That is where the interesting engineering lives, and that is where the real AI principles need to come from.

My first book, Embedded AI, is coming later this year from No Starch Press, covering 25 hands-on hardware projects deploying machine learning on microcontrollers. Sign up for launch updates and bonus material! To support my writing here, please show your appreciation by following me, or subscribe to get an email whenever I publish a new article.