Concerns Over Agentic AI and Liability

Uh,

but the idea is that now

they're talking about agentic AI, which I think is a huge mistake

that they're going to be discovering very soonly.

You should not brand this thing as your agent because if the

government uses it, you are an agent of government, and

that means the Constitution applies to you.

And if they-- the government is using your model

running on your server, uh, where you are submitting queries,

they are then intercepting your query, applying their own

system prompt to it, which recomposes your query running

through

their systems,

and it makes a mistake, and people die as a result.

Or even if it's correct, people die as a result,

as they have in Iran, if

reporting is to believe. They believe where we're bombing the

wrong, uh,

buildings, you know, because, oopsie, the LLM made a mistake.

Uh, and tremendous numbers of, you know, children,

the, the most innocent that we can imagine, are being killed.

Well, who owns the liability there?

And of course, naturally, they would say the government owns the

liability because

you're never going to successfully, uh, sue the government for this, even though

they, they should heed it. Uh, but maybe that doesn't

work anymore because maybe they are an agent of government.

And I, I don't know where we're ever going to see

standing to sue on this, but

we, we should. And people in this room, I think, are frankly going to

be called upon in the course of their careers, in the course of their studies,

in

the course of whatever you do to contend with these kind of issues.

Where are we going to draw a line? Because an algorithm cannot make

mistakes in the way that we understand it as, uh, you know,

you making error. Because the errors in its outputs are not

the computer's. It's the slave to its instructions.

The responsibility lies with its developer, its programmer, its creator,

its

architect. Which is to say responsibility lies with a

person.

And we cannot hide that. We cannot abrade that away, right?

A, a gun maker

is not liable, it's true, for the misuse of a gun.

Uh, but if during the testing of their latest

piece of artillery, they

mistakenly obliterate, you know, a nearby village, it's

not the cannon that should go to jail.

And, you know, just because you put out a press release

describing how surprised

you were that your new model, you know, hacked the neighbor, uh,

or d- you know, how impressed you are by this de-destructive

power of your missile, uh, that doesn't release you from the

liability of launching

the missile that launched itself. You pushed it.

Uh, and, and so th-there's a lot we can get into,

but the idea that I want people

to contend with is, is how we deal with this.

And right now, these closed models are saying AI safety, AI

safety, safetyism, as they call it, uh, where the idea is these things are so

dangerous. Oh gosh, they're so dangerous, but they're so dazzling.

They're so important. They're so valuable. We can't get rid of them.

But, you know, oh, it's going to kill some people,

but it's all right.

You know, that just happens. Oh, it's going to hack some companies.

Oops, you know, whatever. Uh, we'll invest in them.

The idea here is

who operates this? And this should apply to our devices more

broadly around the

world. Right now, you buy a cell phone from Apple, and

before you're allowed to use it, uh, you have to click Okay to continue.

Uh, and they go, "Well, you're assuming liability for that.

You, you clicked Okay to continue." You know,

Google sells you an Android phone.

Samsung sells you an Android phone. You click Okay to continue.

And now that's all you. But that can't be possibly how it works

because you can't

inspect the system. You can't understand how it's used.

And particularly in the context of LLMs, where you are submitting queries to a

model that they are running, uh, where they are for your

safety and everyone else's safety, intercepting the prompt and

recomposing the prompt, saying, "We're doing, uh,

applying safety rules and

transformations to it."

Uh, that's on you,

but it's not. I can't even see what you did.

I asked you a question.

You gave me something back in return.

And if it's an agentic AI where I'm putting this thing in a harness, right?

The harness being just a little program that basically uses these

large models to feed it instructions that it can then

provide to tools that are existing on your local

system. So now your computer that is connected to the

internet is being, frankly, remotely controlled, uh, by a model

that's far away that you don't understand how it works.

But, you know, you hope it works pretty well.

And frankly, they do work reasonably well.

But every now and then, it might hack, you know, some random company or whatever

out there.

Uh,

I think we need to put Sam Altman in jail.

Uh, and I think it'll be an instructive

experience.