If implemented correctly, an LLM has an advantage as it can only van do literal implementation of the law, and pre existent legislature and is essentially blind to who people are.
I remain of the opinion that there is -technically- a way to make this work. The problem of implementation is maybe a human flaw, but we’ve seen in history that there can be ways to change things.
200 years ago three were few democracies, for instance.
The fundamental problem with the law is that she should be blind to the person being judged, this technology can offer that.
I’m quite sure most people arguing that implementation would be difficult, but that’s not due to technical circumstance.
And it’s more the technical aspect that interests me, not the ‘will never happen’ side. It’s a thought experiment.
The fundamental problem with using any kind of AI in decision-making is that they are neither blind to the person nor are they bias-free, nor are they unmanipulable.
In fact, the fundamental basis of any kind of machine-learning-based AI is that it replicates the training set. If there are biases in the training set (and there always are, it’s unavoidable), then you will see the exact same bias in the output.
LLMs (the kind of AI used here) are even worse than just a pure statistical model, since they are super easy to manipulate. Word your request correctly, and it will always approve whatever you put before it. Word it wrong and it will reject a good proposal.
They are also not deterministic. Every time you put something before an LLM you will get a different output, even with exactly the same input.
And they are under the control of the provider of the LLM. That means, whoever provides the LLM can easily manipulate what kind of output the LLM will give.
Effectively, they just gave a minister position to a mindless puppet controlled by an unelected contractor.
AI is not an emotion-less pure-rational arbiter of truth. That’s science-fiction.
So the argument boils down to “If AI was something entirely different than the whole realm of AI, machine learning and statistics it’s based on, then it could be suitable for the job”.
And if you leave out the bread and the patty and instead added spaghetti, cream, eggs and ham, a burger would be a pretty decent carbonara. But then it has nothing to do with a burger any more.
Playing the devil’s advocate:
If implemented correctly, an LLM has an advantage as it can only van do literal implementation of the law, and pre existent legislature and is essentially blind to who people are.
So it can be a blind and incorruptible judge.
However will it remain so?
If implemented correctly, politicians also have neither possibility nor incentive to fall to corruption.
If implemented correctly, we would live in a flawless utopia.
The problem is that nothing ever is implemented corrwctly, so the point of any system really is to have the least desastrous failure mode.
I mean several people have argued this.
I remain of the opinion that there is -technically- a way to make this work. The problem of implementation is maybe a human flaw, but we’ve seen in history that there can be ways to change things.
200 years ago three were few democracies, for instance.
The fundamental problem with the law is that she should be blind to the person being judged, this technology can offer that.
I’m quite sure most people arguing that implementation would be difficult, but that’s not due to technical circumstance.
And it’s more the technical aspect that interests me, not the ‘will never happen’ side. It’s a thought experiment.
The fundamental problem with using any kind of AI in decision-making is that they are neither blind to the person nor are they bias-free, nor are they unmanipulable.
In fact, the fundamental basis of any kind of machine-learning-based AI is that it replicates the training set. If there are biases in the training set (and there always are, it’s unavoidable), then you will see the exact same bias in the output.
LLMs (the kind of AI used here) are even worse than just a pure statistical model, since they are super easy to manipulate. Word your request correctly, and it will always approve whatever you put before it. Word it wrong and it will reject a good proposal.
They are also not deterministic. Every time you put something before an LLM you will get a different output, even with exactly the same input.
And they are under the control of the provider of the LLM. That means, whoever provides the LLM can easily manipulate what kind of output the LLM will give.
Effectively, they just gave a minister position to a mindless puppet controlled by an unelected contractor.
AI is not an emotion-less pure-rational arbiter of truth. That’s science-fiction.
So the argument boils down to “If AI was something entirely different than the whole realm of AI, machine learning and statistics it’s based on, then it could be suitable for the job”.
And if you leave out the bread and the patty and instead added spaghetti, cream, eggs and ham, a burger would be a pretty decent carbonara. But then it has nothing to do with a burger any more.
It will make up laws the minute you turn it on
“If implemented correctly” is doing a lot of heavy lifting in that statement.
If