AI Lies Are Finally Getting Punished

[Beebo Brink]

Case #1: German judge found Google liable for Search overview hallucinations that it generated. AI went beyond summarizing third party sources, it created untrue content that wasn't supported by the provided links. Google owns that content and the liability.

Case #2: A U.S. judge threw out BOTH parties in a dispute because lawyers on each side used AI and presented hallucinated legal references.

 

[Beebo Brink]

Cristiano - Not quite sure what you're reacting to, but the video makes an interesting point about these two cases. Her conclusion is that the legal pressure will be better for the industry in the long run. It will create incentive for them to focus more on accuracy -- because they HAVE to in order to escape liability -- whereas they've let that aspect of AI slide by at unacceptable levels because there wasn't any corrective downside to doing so.

 

[Argent Stonecutter]

There is no way for LLMs to "focus on accuracy" because they have no mechanism to distinguish truth from any other plausible sequence of output tokens.

 

[Isabeau]

Well, I’m facepalming because in the second case, both parties used AI.
(Haven’t watched the video yet, as I’m not at home.)

 

[GoblinCampFollower]

There is no way for LLMs to "focus on accuracy" because they have no mechanism to distinguish truth from any other plausible sequence of output tokens.

Of course not. I'm sure Beebo knows that. It's the developers who have to add more safety checks and get yelled at constantly for when they don't work

It's okay, the companies making these are drowning in so much cash, they'll surely survive any liabilities from them screwing it up.

 

[Casey Pelous]

Of course not. I'm sure Beebo knows that. It's the developers who have to add more safety checks and get yelled at constantly for when they don't work

It's okay, the companies making these are drowning in so much cash, they'll surely survive any liabilities from them screwing it up.

Based on the info in another thread, that "drowning" is not going to last forever. I don't want to go all biz school technical on y'all, but the correct term for what is happening is, they are "losing their ass" on this parlor trick.

As Argent points out, there's no damn way the current paradigm stops hallucinating these bizarre prevarications. It's like asking a rose to stop being a flower.

 

[GoblinCampFollower]

Based on the info in another thread, that "drowning" is not going to last forever. I don't want to go all biz school technical on y'all, but the correct term for what is happening is, they are "losing their ass" on this parlor trick.

I have said many times it's a huge bubble. ...but we have no idea how long it could last.

As Argent points out, there's no damn way the current paradigm stops hallucinating these bizarre prevarications. It's like asking a rose to stop being a flower.

WRONG! The LLM will become self aware and then expand it's consciousness and know the truth about the world!!!! I'm of course just kidding. I know the LLM can't stop hallucinating on it's own, but have previously talked about how the AI's are not just naked LLM's, but wrapped in many quality checks made by the developers. ....these checks are of course very imperfect and can often be tricked. We'll just have to see what happens!

 

[Argent Stonecutter]

The "quality checks" are crude ad-hoc prefiltering rules and instructions to the LLM that depend on the LLM being able to understand what it's reading which of course it doesn't.If they had a program that could actually reason about the LLM output and determine if it was true or not they should just be using that instead of the LLM.

 

[Cristiano]

Cristiano - Not quite sure what you're reacting to, but the video makes an interesting point about these two cases. Her conclusion is that the legal pressure will be better for the industry in the long run. It will create incentive for them to focus more on accuracy -- because they HAVE to in order to escape liability -- whereas they've let that aspect of AI slide by at unacceptable levels because there wasn't any corrective downside to doing so.

I was reacting to them doing this shit in the first place. I am glad they are being held liable.

 

[GoblinCampFollower]

The "quality checks" are crude ad-hoc prefiltering rules and instructions to the LLM that depend on the LLM being able to understand what it's reading which of course it doesn't.If they had a program that could actually reason about the LLM output and determine if it was true or not they should just be using that instead of the LLM.

Yes, we know that. Modern AI's are LLM's with tons of training data and VAST arrays of crude ad-hoc prefilters. I don't envy the people responsible for maintaining those.

 

[Argent Stonecutter]

None of which can do anything about making it "tell the truth".

 

[GoblinCampFollower]

None of which can do anything about making it "tell the truth".

I agree. I was just saying I feel sorry for the people who are responsible for acting like they can do that. Feel sorry for the developers, not the cons in charge of the companies.

 

[Casey Pelous]

.... We'll just have to see what happens!

I believe the German courts just said, "We saw what happened and we have seen enough of that shite, TYVM."

Of course, they said it in German, probably all in one gigantic concatenated word, which somehow provides a level of precision and nuance in its meaning that renders it untranslatable in any other language. Grundstückintelligentsiaartificialichemitslamminundbangindaslyinsackenshittenmitsloppinundfloppinalloverderdamplacenheimer, probably.