> Are we getting close to our very own Stop the Slop campaign?
I don't think so. We read about the handful of failures while there are billions of successful queries every day, in fact I think AI Overviews is sticky and here to stay.
I'd say this isn't just an AI overview thing. It's a Google thing. Google will sometimes show inaccurate information and there is usually no way to correct it. Various "feedback" forms are mostly ignored.
I had to fight a similar battle with Google Maps, which most people believe to be a source of truth, and it took years until incorrect information was changed. I'm not even sure if it was because of all the feedback I provided.
I see Google as a firehose of information that they spit at me ("feed"), they are too big to be concerned about any inconsistencies, as these don't hurt their business model.
Google maps is so bad with its auto content. Ultra private country club? Lets mark the cartpaths as full bike paths. Cemetery? Also bike paths. Random spit of sidewalk and grass between an office building and its parking lot? Believe it or not also bike paths.
> I'd say this isn't just an AI overview thing. It's a Google thing. Google will sometimes show inaccurate information and there is usually no way to correct it.
Well, in this case the inaccurate information is shown because the AI overview is combining information about two different people, rather than the sources being wrong.
With traditional search, any webpages would be talking about one of the two people and contain only information about them.
Thus, I'd say that this problem is specific to the AI overview.
Google doesn't really have an incentive to prioritize accuracy at the individual level, especially when the volume of content makes it easy for them to hide behind scale
Back in 2015 I walked 2 miles to a bowling alley tagged on Google maps (in Northwich, England) with my then gf...imagine our surprise when we walked in to a steamy front room and reception desk, my gf asks 'is this the bowling alley' to which a glistening man in a tank top replies 'this is a gay and lesbian sauna love'. We beat a hasty retreat but I imagine they were having more fun than bowling in there
This reminds me a lot of the special policies Wikipedia has developed through experience about sensitive topics, like biographies of living persons, deaths, etc.
Honestly wouldn't even be surprised if it ends up saying something like, "Dave Barry, previously believed to have died in 2016, has since clarified he is alive, creating ongoing debate."
> Dave Barry, the humorist, experienced a brief "death" in an AI overview, which was later corrected. According to Dave Barry's Substack, the AI initially reported him as deceased, then alive, then dead again, and finally alive once more. This incident highlights the unreliability of AI for factual information.
That is such a classic problem with Google (from long before AI).
I am not optimistic about anything being changed from this, but hope springs eternal.
Also, I think the trilobite is cute. I have a [real fossilized] one on my desk. My friend stuck a pair of glasses on it, because I'm an old dinosaur, but he wanted to go back even further.
Loved Dave Barry's writings over the years. Specifically his quote on humor struck me as itself deep.
"a measurement of the extent to which we realize that we are trapped in a world almost totally devoid of reason. Laughter is how we express the anxiety we feel at this knowledge"
Maybe it's the a genuine problem with AI that it can only hold one idea, one possible version of reality at any given time. Though I guess many humans have the same issue. I first heard of this idea from Peter Thiel when he described what he looks for in a founder. It seems increasingly relevant to our social structure that the people and systems who make important decisions are able to hold multiple conflicting ideas without ever fully accepting one or the other. Conflicting ideas create decision paralysis of varying degrees which is useful at times. It seems like an important feature to implement into AI.
It's interesting that LLMs produce each output token as probabilities but it appears that in order to generate the next token (which is itself expressed as a probability), it has to pick a specific word as the last token. It can't just build more probabilities on top of previous probabilities. It has to collapse the previous token probabilities as it goes?
A few versions of that overview were not incorrect, there actually was another Dave Barry who did die at the time mentioned. Why does this Dave Barry believe he has more of a right to be the one pointed to for the query "What happened to him" when nothing has happened to him but something most certainly did happen to the other Dave Barry (death)?
A popular local spot has a summary on google maps that says:
Vibrant watering hole with drinks & po' boys, as well as a jukebox, pool & electronic darts.
It doesn't serve po' boys, have a jukebox (though the playlists are impeccable), have pool, or have electronic darts. (It also doesn't really have drinks in the way this implies. It's got beer and a few canned options. No cocktails or mixed drinks.)
They got a catty one-star review a month ago for having a misleading description by someone who really wanted to play pool or darts.
I'm sure the owner reported it. I reported it. I imagine other visitors have as well. At least a month on, it's still there.
I am so frikkin tired of trying to help people online who post a screenshot "from Google"(which is obviously just the AI summary) that says feature X should exist even with detailed description of how it works when in reality feature X never existed.
This happens all the time on automotive forums/FB groups and it's a huge problem.
I had a similar experience with meta’s AI. Through their WhatsApp interface I tried for about an hour to get a picture generated. It kept stating everything I asked for correctly but then it never arrived at the picture, actually stayed far from what I asked for and at best getting 70%. This and many other interactions with many LLMs made me realize one thing - once the llm starts hallucinating it’s really tough to steer it away from it. There is no fixing it.
I don’t know if this is a fundamental problem with the llm architecture or a problem with proper prompts.
I just saw recently a band called Dutch Interior had Meta AI hallucinate just straight up slander about how their band is linked to White supremacists and far right extremists
Grew up reading Dave's columns, and managed to get ahold of a copy of Big Trouble when I was in the 5th grade. I was probably too young to be reading about chickens being rubbed against women's bare chests and "sex pootie" (whatever that is), but the way we were being propagandized during the early Bush years, his was an extremely welcome voice of absurdity-tinged wisdom, alongside Aaron McGruder's and Gene Weingarten's. Very happy to see his name pop up and that he hasn't missed a beat. And that he's not dead. /Denzel
I also hope that the AI and Google duders understand that this is most people's experience with their products these days. They don't work, and they twist reality in ways that older methods didn't (couldn't, because of the procedural guardrails and direct human input and such). And no amount of spin is going to change this perception - of the stochastic parrots being fundamentally flawed - until they're... you know... not. The sentiment management campaigns aren't that strong just yet.
"for now we probably should use it only for tasks where facts are not important, such as writing letters of recommendation and formulating government policy."
That's obviously broken but part of this is an inherent difficulty with names. One thing they could do would be to have a default question that is always present like "what other people named [_____] are there?"
That wouldn't solve the problem of mixing up multiple people. But the first problem most people have is probably actually that it pulls up a person that is more famous than who they were actually looking for.
I think Google does have some type of knowledge graph. I wonder how much AI model uses it.
Maybe it hits the graph, but also some kind of Google search, and then the LLM is like Gemini Flash Lite and is not smart enough to realize which search result goes with the famous person from the graph versus just random info from search results.
I imagine for a lot of names, there are different levels of fame and especially in different categories.
It makes me realize that my knowledge graph application may eventually have an issue with using first and last name as entity IDs. Although it is supposed to be for just an individual's personal info so I can probably mostly get away with it. But I already see a different issue when analyzing emails where my different screen names are not easily recognized as being the same person.
63 comments
[ 2.4 ms ] story [ 64.7 ms ] threadI don't think so. We read about the handful of failures while there are billions of successful queries every day, in fact I think AI Overviews is sticky and here to stay.
Personally, if I got a resurrection from it, I would accept the nudge and do the political activism in Dorchester.
I had to fight a similar battle with Google Maps, which most people believe to be a source of truth, and it took years until incorrect information was changed. I'm not even sure if it was because of all the feedback I provided.
I see Google as a firehose of information that they spit at me ("feed"), they are too big to be concerned about any inconsistencies, as these don't hurt their business model.
Surely there is a way to correct it: getting the issue on the front page of HN.
Well, in this case the inaccurate information is shown because the AI overview is combining information about two different people, rather than the sources being wrong. With traditional search, any webpages would be talking about one of the two people and contain only information about them. Thus, I'd say that this problem is specific to the AI overview.
> Dave Barry, the humorist, experienced a brief "death" in an AI overview, which was later corrected. According to Dave Barry's Substack, the AI initially reported him as deceased, then alive, then dead again, and finally alive once more. This incident highlights the unreliability of AI for factual information.
That is such a classic problem with Google (from long before AI).
I am not optimistic about anything being changed from this, but hope springs eternal.
Also, I think the trilobite is cute. I have a [real fossilized] one on my desk. My friend stuck a pair of glasses on it, because I'm an old dinosaur, but he wanted to go back even further.
"a measurement of the extent to which we realize that we are trapped in a world almost totally devoid of reason. Laughter is how we express the anxiety we feel at this knowledge"
It's interesting that LLMs produce each output token as probabilities but it appears that in order to generate the next token (which is itself expressed as a probability), it has to pick a specific word as the last token. It can't just build more probabilities on top of previous probabilities. It has to collapse the previous token probabilities as it goes?
Vibrant watering hole with drinks & po' boys, as well as a jukebox, pool & electronic darts.
It doesn't serve po' boys, have a jukebox (though the playlists are impeccable), have pool, or have electronic darts. (It also doesn't really have drinks in the way this implies. It's got beer and a few canned options. No cocktails or mixed drinks.)
They got a catty one-star review a month ago for having a misleading description by someone who really wanted to play pool or darts.
I'm sure the owner reported it. I reported it. I imagine other visitors have as well. At least a month on, it's still there.
This happens all the time on automotive forums/FB groups and it's a huge problem.
Too much to ask, surely.
I don’t know if this is a fundamental problem with the llm architecture or a problem with proper prompts.
https://youtube.com/shorts/eT96FbU_a9E?si=johS04spdVBYqyg3
I also hope that the AI and Google duders understand that this is most people's experience with their products these days. They don't work, and they twist reality in ways that older methods didn't (couldn't, because of the procedural guardrails and direct human input and such). And no amount of spin is going to change this perception - of the stochastic parrots being fundamentally flawed - until they're... you know... not. The sentiment management campaigns aren't that strong just yet.
Yes, that's exactly what AI is.
That wouldn't solve the problem of mixing up multiple people. But the first problem most people have is probably actually that it pulls up a person that is more famous than who they were actually looking for.
I think Google does have some type of knowledge graph. I wonder how much AI model uses it.
Maybe it hits the graph, but also some kind of Google search, and then the LLM is like Gemini Flash Lite and is not smart enough to realize which search result goes with the famous person from the graph versus just random info from search results.
I imagine for a lot of names, there are different levels of fame and especially in different categories.
It makes me realize that my knowledge graph application may eventually have an issue with using first and last name as entity IDs. Although it is supposed to be for just an individual's personal info so I can probably mostly get away with it. But I already see a different issue when analyzing emails where my different screen names are not easily recognized as being the same person.