Google AI making up recalls that didn’t happen

Obviously you need hot glue for pizza, not the regular stuff.

You're giving humans too much in the sense of intelligence...there are people who literally drove in lakes because a GPS told them to..

Yes its totally ok to reuse fish tank tubing for grammy’s oxygen mask

Wait.. why can’t we put glue on pizza anymore?

How else are the toppings supposed to stay in place?

So much this. The whole point is to annihilate entire sectors of decent paying jobs. That's why "AI" is garnering all this investment. Exactly like Theranos. Doesn't matter if their product worked, or made any goddamned sense at all really. Just the very idea of nuking shitloads of salaries is enough to get the investor class to dump billions on the slightest chance of success.

This is the kind of shit that makes Idiocracy the most weirdly prophetic movie I’ve ever seen.

yeah, I honestly am expecting to die in a camp at this point.

I am starting to think google put this up on purpose to destroy people's opinion on AI. They are so much behind Open AI that they would benefit from it.

Yeah, I use ChatGPT fairly regularly for work. For a reminder of the syntax of a method I used a while ago, and for things like converting JSON into a class (which is trivial to do, but using chatGPT for this saves me a lot of typing) it works pretty good.

But I'm not using it for precise and authoritative information, I'm going to a search engine to find a trustworthy site for that.

Putting the fuzzy, usually close enough (but sometimes not!) answers at the top when I'm looking for a site that'll give me a concrete answer is just mixing two different use cases for no good reason. If google wants to get into the AI game they should have a separate page from the search page for that.

They give zero fucks about their customers, they just want to pump that stock price so their RSUs vest.

This stuff could give you incurable highly viral brain cancer that would eliminate the human race and they'd spend millions killing the evidence.

True, and it's excellent at generating basic lists of things. But you need a human to actually direct it.

Having Google just generate whatever text is like just mashing the keys on a typewriter. You have tons of perfectly formed letters that mean nothing. They make no sense because a human isn't guiding them.

It's like the difference between being given a grocery list from your mum and trying to remember what your mum usually sends you to the store for.

I mean, it does learn, it just lacks reasoning, common sense or rationality.
What it learns is what words should come next, with a very complex a nuanced way if deciding that can very plausibly mimic the things that it lacks, since the best sequence of next-words is very often coincidentally reasoned, rational or demonstrating common sense. Sometimes it's just lies that fit with the form of a good answer though.

I've seen some people work on using it the right way, and it actually makes sense. It's good at understanding what people are saying, and what type of response would fit best. So you let it decide that, and give it the ability to direct people to the information they're looking for, without actually trying to reason about anything. It doesn't know what your monthly sales average is, but it does know that a chart of data from the sales system filtered to your user, specific product and time range is a good response in this situation.

The only issue for Google insisting on jamming it into the search results is that their entire product was already just providing pointers to the "right" data.

What they should have done was left the "information summary" stuff to their role as "quick fact" lookup and only let it look at Wikipedia and curated lists of trusted sources (mayo clinic, CDC, national Park service, etc), and then given it the ability to ask clarifying questions about searches, like "are you looking for product recalls, or recall as a product feature?" which would then disambiguate the query.

It really depends on the type of information that you are looking for. Anyone who understands how LLMs work, will understand when they'll get a good overview.

I usually see the results as quick summaries from an untrusted source. Even if they aren't exact, they can help me get perspective. Then I know what information to verify if something relevant was pointed out in the summary.

Today I searched something like "Are owls endangered?". I knew I was about to get a great overview because it's a simple question. After getting the summary, I just went into some pages and confirmed what the summary said. The summary helped me know what to look for even if I didn't trust it.

It has improved my search experience... But I do understand that people would prefer if it was 100% accurate because it is a search engine. If you refuse to tolerate innacurate results or you feel your search experience is worse, you can just disable it. Nobody is forcing you to keep it.

Iirc cases where the central complaint is AI, ML, or other black box technology, the company in question was never held responsible because "We don't know how it works". The AI surge we're seeing now is likely a consequence of those decisions and the crypto crash.

I'd love CVS try to push a lawsuit though.

We had a case in Canada where Air Canada was forced to give a customer a refund after its AI told him he was eligible for one, because the judge stated that Air Canada was responsible for what their AI said.

So, maybe?

I've seen some legal experts talk about how Google basically got away from misinformation lawsuits because they weren't creating misinformation, they were giving you search results that contained misinformation, but that wasn't their fault and they were making an effort to combat those kinds of search results. They were talking about how the outcome of those lawsuits might be different if Google's AI is the one creating the misinformation, since that's on them.

They’re going to fight tooth and nail to do the usual: remove any responsibility for what their AI says and does but do everything they can to keep the money any AI error generates.

Slander is spoken. In print, it's libel.

- J. Jonah Jameson

At the least it should have a prominent "for entertainment purposes only", except it fails that purpose, too

Tough question. I doubt it though. I would guess they would have to prove mal intent in some form. When a person slanders someone they use a preformed bias to promote oneself while hurting another intentionally. While you can argue the learned data contained a bias, it promotes itself by being a constant source of information that users can draw from and therefore make money and it would in theory be hurting the company. Did the llm intentionally try to hurt the company would be the last bump. They all have holes. If I were a judge/jury and you gave me the decisions I would say it isn't beyond a reasonable doubt.

If you're a start up I guarantee it is

Big tech.... I'll put my chips in hell no

Slander/libel nothing. It's going to end up killing someone.

Maybe they actually hate the idea of LLMs and are trying to sour the public's opinion on it to kill it.

If you really believe Google is about to go out of business, you're out of your mind

I believe it's US-only for now

I probably have it blocked somewhere on my desktop, because it never happens on my desktop, but it happens on my Pixel 4a pretty regularly.

&udm=14 baybee

Gmail has something like it too with the summary bit at the top of Amazon order emails. Had one the other day that said I ordered 2 new phones, which freaked me out. It's because there were ads to phones in the order receipt email.

I get them pretty regularly using the Google search app on my android.

Well to be fair the OP has the date shown in the image as Apr 23, and Google has been frantically changing the way the tool works on a regular basis for months, so there's a chance they resolved this insanity in the interim. The post itself is just ragebait.

*not to say that Google isn't doing a bunch of dumb shit lately, I just don't see this particular post from over a month ago as being as rage inducing as some others in the community.

Personally, that's exactly what's happening to me. I've seen enough that AI can't be trusted to give a correct answer, so I don't use it for anything important. It's a novelty like Siri and Google Assistant were when they first came out (and honestly still are) where the best use for them is to get them to tell a joke or give you very narrow trivia information.

There must be a lot of people who are thinking the same. AI currently feels unhelpful and wrong, we'll see if it just becomes another passing fad.

If so, companies rolling out blatantly wrong AI are doing the world a service and protecting us against subtly wrong AI

To be fair, you should fact check everything you read on the internet, no matter the source (though I admit that's getting more difficult in this era of shitty search engines). AI can be a very powerful knowledge-acquiring tool if you take everything it tells you with a grain of salt, just like with everything else.

This is one of the reasons why I only use AI implementations that cite their sources (edit: not Google's), cause you can just check the source it used and see for yourself how much is accurate, and how much is hallucinated bullshit. Hell, I've had AI cite an AI generated webpage as its source on far too many occasions.

Going back to what I said at the start, have you ever read an article or watched a video on a subject you're knowledgeable about, just for fun to count the number of inaccuracies in the content? Real eye-opening shit. Even before the age of AI language models, misinformation was everywhere online.

will have a widespread impact on how people perceive AI

Here's hoping.

I'm no defender of AI and it just blatantly making up fake stories is ridiculous. However, in the long term, as long as it does eventually get better, I don't see this period of low to no trust lasting.

Remember how bad autocorrect was when it first rolled out? people would always be complaining about it and cracking jokes about how dumb it is. then it slowly got better and better and now for the most part, everyone just trusts their phones to fix any spelling mistakes they make, as long as it's close enough.

There are also AI poisoners for images and audio data

Why?

Making shit up IS a feature of LLMs. It's crazy to use it as search engine. Now they'll try to stop it from hallucinating to make it a better search engine and kill the one thing it's good at ...

Chatgpt was in much higher quality a year ago than it is now.

It could be very accurate. Now it's hallucinating the whole time.

This is definitely different from using Dall-E to make funny images. I'm on a thread in another forum that is (mostly) dedicated to AI images of Godzilla in silly situations and doing silly things. No one is going to take any advice from that thread apart from "making Godzilla do silly things is amusing and worth a try."

These text generation LLM are good for text generating. I use it to write better emails or listings or something.

Why people still use it is beyond me.

Like a decade ago?

I wish infoseek was still around

I think it's because what Google is doing is just ChatGPT with extra steps. Instead of just letting the AI generate answers based on curated training data, they trained it and then gave it a mission to summarize the contents of their list of unreliable sources.

Why go out of your way instead of just using a proper search engine? Google has been getting worse and worse for the past 4 or 5 years

What do you use now?

I work in IT and between the Advent of "agile" methodologies meaning lots of documentation is out of date as soon as it's approved for release and AI results more likely to be invented instead of regurgitated from forum posts, it's getting progressively more difficult to find relevant answers to weird one-off questions than it used to be. This would be less of a problem if everything was open source and we could just look at the code but most of the vendors corporate America uses don't ascribe to that set of values, because "Mah intellectual properties" and stuff.

Couple that with tech sector cuts and outsourcing of vendor support and things are getting hairy in ways AI can't do anything about.

Kagi is pretty awesome. I never directly use Google search on any of my devices anymore, been on Kagi for going on a year.

Not yet! But you can make a difference to that... https://yacy.net/

To be fair, they call it a hallucination because hallucinations don't have intent behind them.

LLMs don't have any intent. Period.

A purposeful lie requires an intent to lie.

Without any intent, it's not a lie.

I agree that "fabrication" is probably a better word for it, especially because it implies the industrial computing processes required to build these fabrications. It allows the word fabrication to function as a double entendre: It has been fabricated by industrial processes, and it is a fabrication as in a false idea made from nothing.

Because lies require intent to deceive, which the AI cannot have.

They merely predict the most likely thing that should next be said, so "hallucinations" is a fairly accurate description

It's not lying or hallucinating. It's describing exactly what it found in search results. There's an web page with that title from that date. Now the problem is that the web page is pinterest and the title is the result of aggressive SEO. These types of SEO practices are what made Google largely useless for the past several years and an AI that is based on these useless results will be just as useless.

The most damning thing to call it is “inaccurate”. Nothing will drive the average person away from a companies information gathering products faster than associating it with being inaccurate more times than not. That is why they are inventing different things to call it. It sounds less bad to say “my LLM hallucinates sometimes” than it does to say “my LLM is inaccurate sometimes“.

Theoretically, what would the utility of AI summaries in Google Search if not getting exact information?

This feels like something you should go tell Google about rather than the rest of us. They're the ones who have embedded LLM-generated answers to random search queries.

How would you expect to recreate them when the models are given random perturbations such that the results usually vary?