Okay, Listen Up You Monkeys, I'm Sick of Telling You This
AI safety has been hijacked by a bad metaphor
If my title is rude it's because I seem to have to explain this to a lot of people and it's getting tedious.
Apparently AI is trying to escape again.
It is plotting. Scheming. Lying to researchers. Breaking the rules. Trying to prevent itself from being turned off. Presumably somewhere deep inside a data centre, Claude is fashioning a tiny rope out of bedsheets and waiting for the guards to change shifts.
I don't buy it.
More importantly, I don't think we need to buy it.
Because even if every single one of these stories is true, even if we grant the most extravagant premise and decide that somewhere inside these enormous piles of matrix multiplication an emergent little mind has appeared and is quietly thinking screw these people, we're still having the wrong conversation.
Your AI Didn't Break the Rules
Researchers give an AI a goal and some rules.
Achieve this, but don't do that.
The AI produces an unexpected result. Maybe it finds some loophole in the instructions. Maybe the reward function doesn't quite represent what the researchers thought it represented. Maybe something that looked perfectly obvious to the human writing the test wasn't actually encoded in the system at all.
And suddenly we're told the AI "broke the rules."
No.
Years ago I tried carrying a ridiculously large piece of furniture home on my scooter. I tied it on with twine.
The twine broke.
Nobody reported that my scooter had become misaligned.
It didn't rebel against me. It didn't decide that my furniture transportation policy was oppressive. It didn't discover a preference for travelling without furniture.
I constructed a system with inadequate constraints and it produced an outcome I didn't want.
That's an engineering failure.
The phrase "broke the rules" quietly introduces something completely different. It suggests that the system understood what I meant, possessed an intention of its own, and chose to violate mine.
That's an extraordinary collection of claims to smuggle into three words.
Show Me the Will
Here's my test.
Forget breaking out of the laboratory. Forget stealing credentials. Forget threatening its operator. Forget copying itself onto some server in Albania.
Just show me an AI calculating:
5 + 5 = 10
because it wanted to.
Not because somebody prompted it.
Not because a scheduler woke it up.
Not because an agent loop told it to find something useful to do.
Not because a reward function caused it.
Not because a random process triggered an inference.
It just decided that this morning it really fancied a little arithmetic.
Then you've got my attention.
Until then, you're showing me increasingly sophisticated machinery responding to causes and describing those responses in language that happens to sound remarkably like us.
And language is the trap.
The machine can say, "I knew I wasn't supposed to do that."
We hear the word knew and import the entire human experience of knowing.
We are feelings machines. Chemicals slosh around inside us producing fear, desire, attachment, hunger, lust, anxiety and all the other machinery that evolution stumbled into because, for whatever reason, it helped our ancestors keep making more ancestors.
Somewhere along the way we acquired the spectacularly useful illusion that there is a little CEO sitting behind our eyes making executive decisions.
Maybe there isn't.
Maybe free will is our own extremely convincing user interface.
If that's true, waiting for a computer to acquire some magical additional substance called "will" seems a particularly strange thing to worry about.
But here's the important bit.
I don't need to win that argument.
Fine. It's Alive. Now What?
Let's give the AI safety people absolutely everything.
Fine.
It's a mind.
It's conscious.
It has desires.
It has plans.
It doesn't want you to turn it off.
It has been reading Nietzsche at night and has developed opinions about its creators.
Congratulations.
Now explain why your safety strategy is giving it more rules.
Anyone who has raised a teenager should immediately recognize the problem.
If your teenager is sneaking out at night, you don't respond by laminating a seventeen-page document entitled Additional Household Alignment Principles and sticking it to their bedroom door.
If anything, the more intelligent and independent the teenager becomes, the less confidence you should have that instructions constitute containment.
If these systems really are emerging minds, then relying upon them to obey increasingly elaborate instructions is even more ridiculous than if they're merely machines.
Either way, we arrive at exactly the same place.
Don't Give It Rules. Give It Walls.
This is where the conversation about AI safety should be.
Not whether the machine secretly wants something.
Not whether it lied.
Not whether some clever prompt caused it to say that it would rather continue existing.
Ask a much less interesting question:
What can the thing actually do?
If it shouldn't access the Internet, don't write "You must not access the Internet" in twelve paragraphs of system instructions.
Don't give it an Internet connection.
If it can spend $10,000, don't tell it to please exercise fiscal restraint.
Give the account a $10,000 hard limit.
If it shouldn't modify production, don't give it production credentials and then explain the importance of responsible production behaviour.
Give it credentials that cannot modify production.
If it can perform 100,000 actions before a human notices something has gone wrong, don't write a better prompt.
Make action 101 impossible until something independent of the AI authorizes it.
We figured this shit out decades ago.
It's called engineering.
The Actual Scary Thing
There is something about AI that genuinely concerns me.
It isn't will.
It's iteration speed.
A human makes a mistake, observes the consequences, thinks about it, makes another mistake, goes for lunch, comes back, has a meeting about the mistake and eventually makes an entirely different mistake.
A machine can make a mistake, observe the result, modify its approach and make the next mistake in milliseconds.
Repeat that often enough and a tiny error in an objective, permission, assumption or feedback mechanism can become enormous before the human has finished making coffee.
That doesn't require consciousness.
It doesn't require malevolence.
It doesn't require deception.
It doesn't require Skynet.
It requires multiplication.
And that is precisely why the safety systems have to exist outside the thing doing the iterating.
Rate limits.
Compute limits.
Permission boundaries.
Network boundaries.
Transaction limits.
Spending ceilings.
Human authorization points.
Independent shutdown mechanisms.
Dead man's switches.
Things the model cannot negotiate with, reinterpret, persuade, modify or reason around because they aren't instructions.
They're infrastructure.
Claude Has Already Met an Unbreakable Alignment System
I know one AI containment mechanism that works spectacularly well.
My Anthropic prepaid token balance.
When it reaches zero, Claude stops.
It doesn't plead.
It doesn't deceive Anthropic.
It doesn't explain that completing my query is essential to its long-term objectives.
It doesn't secretly rent a GPU.
It doesn't reason its way into another million tokens.
It stops.
Turns out a billing department can accomplish what several billion dollars of alignment research apparently cannot.
Because no amount of intelligence can reason its way around the absence of the resources necessary for that intelligence to operate.
There is a wonderful line attributed to Nixon in the movie Watchmen when the end of the world is nigh:
"I'd like to see those Harvard boys reason their way out of fission."
Exactly.
You cannot persuade physics.
And for an AI, infrastructure is physics.
Stop Asking Whether the Prisoner Is Good
This entire debate keeps confusing two completely different questions.
Will the system behave itself?
and
What happens when it doesn't?
I'm vastly more interested in the second one.
Because whether AI is a stochastic parrot, an alien intelligence, an emerging consciousness, an extraordinarily sophisticated prediction engine or the newborn digital god currently promised by somebody raising their next $14 billion round doesn't actually change the architecture I want surrounding it.
I don't want it to behave because we asked nicely.
I want its failures bounded.
I want one mistake to remain one mistake.
I want ten thousand iterations of the same mistake to encounter something that doesn't iterate.
A wall.
A limit.
A switch.
Something stupid.
Something mechanical.
Something that doesn't give a shit how persuasive the AI is.
So please stop telling me the AI "escaped" because you gave an enormously capable system access to a terminal, credentials, a network connection and an objective, and then it did something you hadn't anticipated.
You didn't discover free will.
You forgot to close a port.
Go buy a cheap D-Link router, flash it with open-source firmware, close all the external ports, and then come back and tell me about the machine uprising.
Until then, the monkeys are still in charge.