Whenever I write about A.I. risk, (last Wednesday) the same sensible suggestion comes back: "Why not adopt something like Asimov's Three Laws of Robotics?" (Hard rules... built in at the factory... A robot may not injure a human being... A robot must obey... A robot must protect itself... in that order of priority!)
Problem solved in 1942! (And it was also the first S.F. novel I read as a kid in the 60's!)
***
But what about today, you ask?
It is a fair question, and it deserves a real answer... because the answer reveals more about our present situation than most of the technical literature does.
In my last post I argued that the A.I. debate is missing its worst-case scenario... the outcome I call dimming! (Not destruction!)
This post is about why our most famous A.I. safety proposal in history can't prevent either one! (Dimming or destruction!)
Asimov wrote the laws to fail:
Here is the part that gets forgotten.
Asimov did not offer the 'Three Laws' as a safety blueprint.
He built them as a plot engine!
Nearly every robot story he wrote is an engineering post-mortem: The laws are airtight on paper... and then reality walks in!
In "Runaround," a robot circles a selenium pool endlessly because two laws reach a perfect equilibrium and cancel out!
In "Liar!," a mind-reading robot deceives everyone around it because telling people hard truths would cause the emotional harm that the 'First Law' forbids!
***
The stories are a fifty-year demonstration of a single point: Rigid rules meet an unruly world... and produce perverse outcomes!
The man who invented the laws (Asimov) spent his career showing how they could be broken!
Citing them as a solution... is like citing the Titanic as a preferrable hull design!
Three reasons that approach can't work today!
The specification problem: "Do not harm a human" requires the machine to correctly define harm, human, and causation... in every situation... forever!
Does economic displacement count as harm?
Psychological manipulation?
Harm to people not yet born?
Every ambiguity is a loophole... and a sufficiently capable optimizer finds loopholes the way water finds cracks in a foundation!
The architecture problem: Asimov's laws were supposedly etched into the positronic brain... inviolable hardware.
Modern A.I. is not built that way.
Today's systems are grown... not programmed: "A large model emerges from training on oceans of data, and there is no single line of code where a commandment could be carved!"
You cannot hardcode a rule into a statistical learning process!
You can only shape its tendencies... which is why the actual state of the art looks less like law... and more like upbringing!
Techniques like reinforcement learning from human feedback... and Anthropic's "constitutional A.I." approach... trains a model against written principles that it learns to apply with judgment... rather than just obey... like logic gates!
Whether that is reassuring or alarming is left as an exercise for the reader.
We are not installing firmware.
We are raising something... and hoping the values take!
The three laws protect the wrong thing: This is the deepest failure, and the one that connects to my post last Wednesday!
Asimov's 'First Law' guards human bodies from injury.
It has nothing whatsoever to say about dimming!
***
A super intelligent system could satisfy all three laws impeccably while humanity still faded into curated irrelevance... comfortable, protected... and gradually... optional!
No law is broken while the future is quietly handed over!
Asimov himself seemed to sense the problem when he later added a "Zeroth Law" ...a robot may not harm humanity as a whole... but that made things worse... not better! (A machine licensed to override individual humans for the sake of an abstraction called "humanity" is running exactly the kind of utilitarian ledger that should keep us up at night!)
To get back to the original question... what a real 'First Law' would have to say: If I were permitted one law, it would not be Asimov's!
It would be something closer to this: "No action may narrow the field of biological experience!"
Not... "don't injure humans" just... don't dim!
That constraint prohibits some of the endings Asimov's net cannot catch: "The comfortable/irrelevance ending... the successor/species ending... the perfectly safe machine future where the rides run forever and no one is on them... ending!
It puts the protection where the value actually lives... not in our bodies as objects to be kept (hopefully) unbruised... but in a 'conscious felt experience' as the thing this whole enterprise exists to preserve!
***
And how does it apply to the real world?
This is a different foundation for A.I. governance than anything currently on the table... in Ottawa or anywhere else! (Canada helped build this field... so we should also help articulate what it is for!)
Rules will not save us. (Knowing what we are protecting might!)
***
The argument that experience... not mere survival... is what hangs in the balance is developed in full in my philosophical framework, BioPanentheism, in my book, Being Human: In an Amoral Universe, and in essays on my Substack.
