LLMs are not transparent at all and are very easy to corrupt, also they have our current biases baked into them (because they are baked into their training data).
The prompt would be transparent and we would use fixed version of LLM so it would be impossible to attack it by modifying the model. Unless they would start injecting prompts into legislation we would be fine.
There of course will be a person representing the AI but they will always vote the way the LLM tells them to.
That means the LLM is voting in all but the actual mechanics of having its vote registered.
Unless they would start injecting prompts into legislation we would be fine.
Yeah, no. The way Congress works right now, the primary attack vector would be internal, not external. Aside from the whole bias thing, injecting prompts would only be the first thing anyone with access would do.
You may not be aware of this, but there is a LOT of legislation not actually written by legislators, who are very busy people, but by lobbyists and lawyers who are glad to do it for them. There is already no actual way of knowing what specific person penned any specific line in any given law outside a legislators’ word that they did it personally or their staff did.
You’re literally opening up that vote to anyone savvy enough to “do it for” a busy legislator, which is already occurring in every other way. Even Jon Stewart was asked to write the legislation for the 911 first responders, and he’s not an attorney, just a passionate guy lobbying for righteous legislation.
And what about the legislators that are already corrupt? If the legislator to whose vote you bind this LLM is already selling their power, why would they not just tell that donor or that lobbyist “hey, now you gotta fix the LLM too” in order to continue doing business as usual?
So for as long as your hypothetical unbiased LLM exists, it would present an open challenge to any comer, both from within Congress and without, to change its recommendations with the prize being the ability to switch any given vote. That’s a hell of a brass ring to reach for when you’re talking about votes worth billions of dollars passing or failing on razor thin margins like they are these days.
And making the LLM “fixed” – static – would not make it safer: if it could be made unfixed in a way that preserved its unchanged appearance, it absolutely would be.
And even then, if it passed all those hurdles and managed to remain incorruptibly non-involved, dependent solely on what it had been fed in the beginning, it still would not have the inherent understanding of humanity required to make correct and rapid decisions in a crisis, because it would not be traveling a cognitive path of understanding relative urgency, but whatever was already baked into it for normal operation.
“Once men turned their thinking over to machines in the hope that this would set them free. But that only permitted other men with machines to enslave them.”
“‘Thou shalt not make a machine in the likeness of a man’s mind.’”
Cool idea, though. I like the way you think. But for now it’s still science fiction (I very seriously hope).
LLMs are not transparent at all and are very easy to corrupt, also they have our current biases baked into them (because they are baked into their training data).
The prompt would be transparent and we would use fixed version of LLM so it would be impossible to attack it by modifying the model. Unless they would start injecting prompts into legislation we would be fine.
That means the LLM is voting in all but the actual mechanics of having its vote registered.
Yeah, no. The way Congress works right now, the primary attack vector would be internal, not external. Aside from the whole bias thing, injecting prompts would only be the first thing anyone with access would do.
You may not be aware of this, but there is a LOT of legislation not actually written by legislators, who are very busy people, but by lobbyists and lawyers who are glad to do it for them. There is already no actual way of knowing what specific person penned any specific line in any given law outside a legislators’ word that they did it personally or their staff did.
You’re literally opening up that vote to anyone savvy enough to “do it for” a busy legislator, which is already occurring in every other way. Even Jon Stewart was asked to write the legislation for the 911 first responders, and he’s not an attorney, just a passionate guy lobbying for righteous legislation.
And what about the legislators that are already corrupt? If the legislator to whose vote you bind this LLM is already selling their power, why would they not just tell that donor or that lobbyist “hey, now you gotta fix the LLM too” in order to continue doing business as usual?
So for as long as your hypothetical unbiased LLM exists, it would present an open challenge to any comer, both from within Congress and without, to change its recommendations with the prize being the ability to switch any given vote. That’s a hell of a brass ring to reach for when you’re talking about votes worth billions of dollars passing or failing on razor thin margins like they are these days.
And making the LLM “fixed” – static – would not make it safer: if it could be made unfixed in a way that preserved its unchanged appearance, it absolutely would be.
And even then, if it passed all those hurdles and managed to remain incorruptibly non-involved, dependent solely on what it had been fed in the beginning, it still would not have the inherent understanding of humanity required to make correct and rapid decisions in a crisis, because it would not be traveling a cognitive path of understanding relative urgency, but whatever was already baked into it for normal operation.
I hate to quote any bible, but the Orange Catholic Bible had it right:
Cool idea, though. I like the way you think. But for now it’s still science fiction (I very seriously hope).