Google found liable for bad AI Overview results. Let's play Truth Or Consequences

OPINION Tech companies hate liability, or at least the sort that makes them liable if something goes wrong. It doesn’t much matter if what they ship is buggy, shabby or simply blows chunks, it’s on you for using it. You fool. Corporates can get service level agreements to focus their suppliers’ minds, and life-critical …

[SVD_NL]

...until hallucinations and false equivalencies are fixed

Soo.... never?

I'm also wondering about their statement about "everyone" knowing not to trust AI. All i see around me, is people not moving past the AI overviews, because they are very convenient. Why do i see this? Because i constantly get people telling me falsehoods, and being unable to solve issues with non-existent or non-applicable fixes. I'd like to have someone subpoena the statistics on click-through rates on the AI summaries links, and comparison data between search result website visits before and after introducing AI overviews. You have to back up your claims in court, don't you? I've seen some small-scale research papers painting a very different picture from what Google describes here, showing that zero-click search results are down overall, and showing a very strong correlation between zero-click searches and AI overviews being present.

Two examples: SparkToro, PewResearch.

[MiguelC]

I've had colleagues creating some user panic because of info they got from Google's AI summaries being delivered as "the truth". While those colleagues are technically knowledgeable people, they seem to lack critical thinking.... not ideal when dealing with this new 'vibed' world

[mlvj]

As a minimum, you could EXPECT Google to run the AI generated results through an AI results checker, which could immediately reduce hallucinations by SOME percentage.

It seems they do not even do this minimum.

[SVD_NL]

I'm pretty sure they use Retrieval-Augmented Generation (RAG), which does reduce hallucinations and helps with models getting outdated, but it's not a silver bullet by a long shot. ARS Technica has a pretty good writeup on the subject, if you're interested in what it does and what the limitations are. TL;DR it still hallucinates, doesn't know it's own limitations (so doesn't know when to initiate RAG), and can't accurately determine the credibility of the source used with RAG (i.e. "Shit in = shit out").

[MonkeyJuice]

It's also extremely optimistic to use a vector embedding for similarity search to reliably pull document chunks into context in the first place. They have extremely poor retrieval accuracy, and fall flat on their face with basic things like negated clauses. Certainly, things have come a long way since the cosine similarity scores and similar of the 80s and 90s, but the way RAG is still being touted as revolutionary in some circles beggars belief. While it does improve things, as your corpus grows the gains rapidly fall into statistical noise.

It's also kinda weird- the original RAG paper specifically finetunes the embeddings AND generator end to end. Modern 'RAG' systems do not- they use a black box, (possibly) tuned embeddings model but don't touch the LLM weights directly, so is anyone really doing RAG outside of an academic setting anyway, or have the AI bros short-cutted a more nuanced concept to sound cool?

I couldn't possibly comment.

[IGotOut]

Yes!

No all we have to do is all get some no win no fee lawyers to go after them and destroy them.

I'll happily only get 50% of the payout if it means thousands of cases go after them.

No class action shit. Individual cases to tie their legal teams up for the next decade and cost them far more money, every time. No getting the usual "A settlement was agreed and Google accepted no liability" crap.

We need these companies to hurt.

[Aladdin Sane]

Time to buy popcorn

[Alumoi]

Don't forget to google for the best offer!

[Aladdin Sane]

10/10, no notes.

[theOtherJT]

It can no more compete on quality of results...

...than a dodo can enter the All Mauritius Aviad Aerobatics championship.

This line is worthy of Red Dwarf in it's prime. It was impossible not to read that in Kryten's voice. That nearly caused coffee to meet keyboard.

[Bebu sa Ware]

Re: It can no more compete on quality of results...

followed rather nicely by "the first rats of legal liability have scuttled ashore." — Presumably from the wreck of "No Liability Disclaimer."

[David 132]

Re: It can no more compete on quality of results...

That nearly caused coffee to meet keyboard.

"An excellent suggestion, sir, with just two minor drawbacks. Firstly, I don't have any coffee, and secondly, I don't have any coffee. I realize that technically that is only one drawback, but it's such a biggie that I felt it was worth mentioning twice."

[Alan Mackenzie]

This article could have been hallucinated by an LLM.

There are no details in this article: a link to court's determination, the identities of the persuers, dates, anything. Just the fact it happened in Munich.

Why? Is this real?

[Lexi Ryedale]

Re: This article could have been hallucinated by an LLM.

Do you really believe an LLM could come up with the quality of puns and witticisms contained in this article?

Unless "Rupert Goodwins" is the code name of a new and massively advanced LLM release, I sincerely doubt it.

[Jonathan Richards 1]

Guy has got a Wikipedia page...

...here

I've been reading his stuff for decades.

[Irongut]

People still use Google search?

How quaint.

[RM Myers]

It amazes me when people complain that AI has turned Google search into shit. Google search was shit long before LLM's became a thing - ads disguised as search results, totally crap sources/scams floating to the top of the results, consistently returning the negative of your question, etc. And this was before considering the lack of privacy protection.

[JoeCool]

That is gold Jerry

"lie ability trumps liability."

Rupert, you can take the rest of the day off.

[Bebu sa Ware]

AI Google's kittle cattle

Unfortunately for the Chocolate Factory one has always been responsible for the damage or harm caused by one's cattle. Even before Roman law.

Doubtless this liability will be legislated or signed [EO] away in Trumpisstan but hopefully less surreal jurisdictions will follow this European precedent.

[mlvj]

LLMinati

I love that term.

[Lanta]

Let's do the same with self driving cars

Ok, so there aren't any actually self driving cars yet, but there are cars that pretend to be under some circumstances, and no amount of disclaimers and get out clauses should absolve their makers of responsibility for damages and deaths they cause, and more than human drivers can get off because they are drunk.

If you put a product out there you are liable for damage it causes. Simples.

[Herring`]

Re: Let's do the same with self driving cars

Also ....

Remember when Facebook showed you the posts of the people on your list in a sensible order. I gave it up because it was showing me auto-recommended bollocks rather than what I had requested. They are not the only guilty party - all the big sites recommend stuff to keep you on there. At the point where they started doing that, (IANAL) it seems to be that they became publishers. They are taking editorial decisions about what to show you. And that makes them liable for what they show to people.

[Tron]

There is another model.

Faced with Friday coming after Thursday, AI companies are looking to governments to invest in AI. There is more to this than money. Governments often act above the law and expect to get away with it. And they often do. At the very least, they can tweak the legal system so that there are enough loop holes for them to do what they want. National security, superinjunctions, D-notices etc.

So I guess the AI shovellers have a solution. Once the government are on board, AI will be able to get away with stuff much more easily.

Incidentally, Google's search AI may be the only vaguely beneficial AI out there, albeit not in every case. It costs nothing and is sometimes useful, as long as you consider it to be as bright and trustworthy as a celebrity.

Actually charging people for AI slop is far more sinister.

[Anonymous Coward]

Re: There is another model.

I'm pretty sure it costs plenty in DC construction and running, like water and energy.

So there's the question of where the money to pay for it is coming from.

Which reminds me :

https://www.mcsweeneys.net/articles/ai-economics-for-dummies

[katrinab]

I've switched to noai.duckduckgo.com because I got fed up with all the AI slop at the top of my search results, and actual search results that are now worse than AltaVista's were before Google came along and showed everyone how search should be done.

[I could be a dog really]

If you set your default search engine to be https://www.google.com/search?q=%s&udm=14&tbs=li:1 then you'll not see the AI slop. Thanks to a previous ElReg article.

[ecofeco]

Bamboozle vs quality product and service

If companies would put the same effort into offering good products and service, they would make more money than fraud, deception and the outright disdain they have for customers.

But then, I'm not some nepo-fail-baby-trustafarian with the right connections, larping as some kind of leader, so what do I know?

[Groo The Wanderer - A Canuck]

Hallucinations and errors aren't a solvable problem, though, they're endemic to the fact that it is a statistically driven summary of information that loses context and details during the distillation into its tokens and weights.

I say again: There is no "solution" for Artificial Ignorance hallucinations. None.

[Jellied Eel]

Hallucinations and errors aren't a solvable problem, though, they're endemic to the fact that it is a statistically driven summary of information that loses context and details during the distillation into its tokens and weights.

I think they are solvable, mostly by the education system. Wiki is not a reliable source, neither are Google AI or even the Bbc. So taking one recent example-

https://www.bbc.co.uk/news/articles/cevlxg72kyvo

Sunday's operation, which saw commandos fast-roping from a helicopter on to the tanker, was the first of its kind carried out by UK armed forces.

Something was wrong with these pictures. Like the camera person already being on deck to film the Marines fast-roping onto the tanker, then being ahead of the commandos to film them doing the room clearing. So a bit of security theatre that humans should be able to spot, but an AI probably can't. At the moment, Google's AI doesn't offer an opinion on "was the raid on Smyrtos staged?", just a list of news sources raising that question.

Eventually an 'AI' should be able to reach a consensus regarding what the truth might have been, but then that consensus may be biased or poisoned, if they rely on Wiki or news sources that do a spot of circular reasoning. Use Wiki or AI to write a 'news' article and the well poisons itself. Which is only going to get worse with the fixation on banning unofficial misinformation and censorship, which will make it harder for our own wetware inference engines to detect BS.

[ChoHag]

The crew certainly won't have cameras in their phones and be willing to sell the pictures for the beeb for a few quid. Must be a set-up job.

Good thing there was a human on board to spot the flaw in your piss-poor logic eh? We wouldn't want people to get taken in by bullshit would we?

Better luck next time.

[Jellied Eel]

Good thing there was a human on board to spot the flaw in your piss-poor logic eh? We wouldn't want people to get taken in by bullshit would we?

Uh.. sure. So a raid to arrest the crew and vessel allowed the crew to film the operation unimpeded?

A MoD spokesman admitted the footage had been taken after initial forces "cleared" the tanker, in order to ensure filming could be done safely, but denied it had been "staged" or contained repeated activities.

At one point in the footage, a voice can be heard saying "rolling" and asking a soldier to "hold that" as a search is being carried out.

Some humans do seem unable to spot flawed logic and are easily taken in by BS I guess..

[ChoHag]

Tell me you've never been dealt with by the police without telling me...

> So a raid to arrest the crew and vessel allowed the crew to film the operation unimpeded?

Yes. We (still, barely) have this thing called "freedom". Look it up some time.

[Jellied Eel]

Yes. We (still, barely) have this thing called "freedom". Look it up some time.

Don't kid yourself. But let me repeat what the MoD said-

A MoD spokesman admitted the footage had been taken after initial forces "cleared" the tanker, in order to ensure filming could be done safely, but denied it had been "staged" or contained repeated activities.

So they filmed themselves clearing a ship that had already been cleared. So pretty much by definition, staged. And as for freedoms, the whole raid was to hijack a ship in international waters..

[Groo The Wanderer - A Canuck]

Could be worse - they could have tried to pass off video game footage as being "coverage of the war" like Trump did. *LOL*