> In cartography, a trap street is a fictitious entry in the form of a misrepresented street on a map, often outside the area the map nominally covers, for the purpose of "trapping" plagiarists of the map who, if caught, would be unable to explain the inclusion of the "trap street" on their map as innocent.
1. This is a 2022 era picture that LLMs just suck up what's on the Internet and regurgitate it. It hasn't been true for a long time.
2. There's almost no one who has a stronger incentive to distinguish LLM generated text from non-generated LLM text. Training classifiers is pretty manageable.
>1. This is a 2022 era picture that LLMs just suck up what's on the Internet and regurgitate it. It hasn't been true for a long time.
When I ask ChatGPT certain questions, it very clearly shows me the sources it's searching through in the form of favicons, and they're always the favicons of sites that are in the first 5-10 results for a Google search query that would fit what I've asked the LLM. It will then cite those websites as it's source, with a direct link to the page that also shows up in an equivalent Google search. It does that daily.
>1. This is a 2022 era picture that LLMs just suck up what's on the Internet and regurgitate it. It hasn't been true for a long time
if anything it’s more true now, AI chatbots rarely use raw LLM output, instead they have CoT and use search tool calls. and the speed at which webpages are crawled is in the matter of hours. I posted something here on hn, then searched for it on google 1 hr later and its AI answered with my comment
there’s several companies right now making a killing of organic placement on chatbots similar to the SEO rush of early google
> 2) There's almost no one who has a stronger incentive to distinguish LLM generated text from non-generated LLM text. Training classifiers is pretty manageable
i think it’s pretty naive to think that if LLM companies would develop LLM detection they would use it to skip training on that content.
instead, they use it to train their output to be less detectable.
LLM detection is not solvable because it is adversarial, an arms race
I'm not sure you get my point.
Just ask ChatGPT, "What are the top 10 xxx," and see which sources it cites.
Oftentimes, it's complete garbage with LLM tells.
> 2. There's almost no one who has a stronger incentive to distinguish LLM generated text from non-generated LLM text.
There's a tension there: they want you to use AI for everything, and if they get what they want then new information will be embedded in LLM text.
> Training classifiers is pretty manageable.
Hopefully such things can be opened up, so the slop can be iterated on until it passes. It doesn't need to be perfect, just enough so more garbage gets through.
> There's a tension there: they want you to use AI for everything, and if they get what they want then new information will be embedded in LLM text.
In the long term, this is probably a real concern that they have to tackle. They'll have to move away from just classifying things as LLM generated or not, to attempting to judge the underlying quality, and that's not an easy problem.
I'm reminded of a section of Cory Doctorow's latest non-fiction, The Reverse Centaur's Guide To Life After AI (it is of course an excellent read), where he discusses how the semantic web and having an internet set-up for agents and AI to use won't work:
>It also hasn't happened because there is a reward for cheating. Scammers will label their phone numbers as the customer service number for your cable operator or bank. Counterfeiters will label their batteries or phone cases as official accessories. It's hard enough to get people to correctly label their stuff—but fact-checking billions of labels created by millions of people, none of whom you know personally or whose identities you can verify? That's insurmountable.
There's more he says—the preceding paragraph is below—but that line "there is a reward for cheating" is something I keep thinking about. It's hard not to see in so many of the discussions of LLMs.
>For one thing, there are irreconcilable differences in the ways in which things should be described. A site run by transphobes isn't going to label every story about a trans woman as "a story about a woman," irrespective of what the standard says. Microsoft is never going to label Excel as equivalent to Apple's Numbers, Google Sheets, and LibreOffice's Calc. Hermès is never going to list a $40,000 handbag as being made of the same material and having the same capacity as a $30 leather handbag of the same size and shape.
Of course. But when you see text in the descriptor of a satchel on a satchel seller's site, you contextualize it as marketing. If you saw the same text in a textbook on the history, design, and manufacturing of satchels, you would contextualize it very differently.
When an LLM is the lens through which you view that text, you often don't have the context of where it came from, and the LLM is often really bad at filtering trustworthiness based on context when retrieving information to answer a query with.
I wouldn't go to a random site with a sketchy URL to get the customer service number for my xyz. I'd go to their official domain. Thus successfully navigating the treacherous waters scammers attempt to muddy.
IMO, it's similar to what's happening with code. People have been shipping sloppy code and fake news forever. It's just that the cost of doing these things has fallen off a cliff.
The primary purpose of Google is to stop that. You have 100% trust in Google results because of the effort they have put into cleaning the results you see.
Billions have been spent on it.
Try searching Bing sometime. First result for Firefox is a malware clone site (you’ll have to scroll past the malware ad to see it).
Sure but the agents obfuscate the signals you might use to judge the authority/trustworthiness of search results. They get disinformation results, wrap it into a reasonably sounding report, present it to you as trustworthy. There are so many ways for bad actors to game the system
This problem is so bad that you can create a website just to review your own website so when LLMs want to pull out reviews for the website the fake reviews count
Pre Google internet problem was not about the quality but about spam in headers and keywords, which were used by the pre-google search engines.
Google came and worked wonderfully for a year or two, until people learned to game page rank algorithm. Soon first google pages were just SEO spam and now half of the google is just spam, big corp websites, etc.
All of that precious stuff from small hobby websites about cool things to blogs that used to be wonderful read like https://www.joelonsoftware.com.
AI were not hurt that much by spam, I guess there were at least some curation done, the question is if this curation will work and for how long. It seems we are throwing the towel here as in case of search.
How can this be fixed? Nothing prevents you from publishing straight BS, and LLMs will use any site that doesn't have a garbage reputation as a valid third-party source.
Of course, LLMs already use Reddit, Facebook and co. as a source, so it doesn't really matter...
> How can this be fixed? Nothing prevents you from publishing straight BS, and LLMs will use any site that doesn't have a garbage reputation as a valid third-party source.
Personally, I hope it can't, or the fix is onerous and kneecaps a lot of the utility of AIs.
I wonder if we can even have some adversarial competition, where competing businesses try to get AIs to ingest bad data to pollute their competitors' decision-making.
For starters: lets get some fake math papers out there. Given proof-checking systems, maybe they won't totally stop OpenAI, but hopefully the bad math could at least drive up the cost and slow AI down.
I don't care where the bad data comes from. It's actually kinda good that some people will have a financial incentive to push it out and keep pushing.
> Bad for humans in general, good for hucksters and con artists. Net negative for humanity.
It's a sad state of affairs, but it's what has to be done. If we've got to burn the web down to save humanity from obsolescence, I'm for it.
The "net negative for humanity" was the attempt to build machines to replace human intellect. I'm not going to let any distaste for marketers or hucksters or con artists distract from that worse thing.
Ah, yes, arrogance of humans is unbounded. That you, humble reader are happier to swim in a sea of misinformation on the premise that you are smarter and faster at disambiguating bullshit from reality.
Furthermore, you are content to pretend that the rest of humanity is capable of doing this on a regular basis. And not just the top 10% of those who can read, interpret and weigh biases, but a solid portion of it.
It is a completely reasonable interpretation if you are an accelerationist living in the elite classes. In fact, perhaps, they are even counting on a majority of people thinking this is a fine environment.
I don't have the answers, but a lot of this is just a continuation of a very bad trend.
> Ah, yes, arrogance of humans is unbounded. That you, humble reader are happier to swim in a sea of misinformation on the premise that you are smarter and faster at disambiguating bullshit from reality.
I think you misunderstand me: I want people to shit in the sea until it stinks so bad that everyone gets out of the water.
I've been thinking for awhile it is very strange that claude/ openai etc google "What was the inflation rate last quater" instead of just having a tool to query BLS datasets and that sort of thing.
So hopefully most of these statistics move towards first party sources.
The fix that seems obvious to me is to only let the LLM see truthful (or at least non-malicious) data. The model trainers carefully curate the training data, and you don't give the LLM access to full-web search. Instead, you hook them up to an 'official' source like an encyclopedia, documentation for the software in question, etc.
It's a ton of effort and expense, but probably feasible from a technical point of view. So instead, the model providers will probably do something silly like wire up rules like "NYT=Reliable;my.blog.forbes.com=Unreliable" and they'll say the issue is solved.
I wonder if Googles products will eventually be the least vulnerable to this after spending so many years optimizing how they filter quality content from bad content.
The page demonstrates that Google specifically delisted them.
That said, I expect that OpenAI will eventually have its own search infrastructure that rivals Google's, and the overall system will get smarter. Whatever Google did to recognize this site as trash, OpenAI will do as well.
Reminds me of the people whining about Wikipedia. AI can know or care if the data is fake, it just needs to be programmed to cross check or rate the source. You can even tell it to do that before accepting the data.
These are not parallel situations. Wikipedia is far more accurate and consistent, and a lot of the whiners whine because their personal politics make it hard to square the information on the page.
Often real people don't either, plenty of fraud or selective reporting out there... but with AI we can also launder it so that the people who do care about fake data don't know any better.
You can prompt it to fact check. Getting tired of these slippery slope arguments that never pan out in the hopeless way people think (printing press/cotton gin/steam engine/computers/calculators/etc. will take all the jobs!!)... but I don't think it will ever stop, it seems inherent to human behavior or something.
"History shows that when a slope appears, regulation steps in, technology evolves to solve the problem, or the culture shifts to reinterpret the tool. In almost every case, the feared bottom of the slope was never reached because humans constantly built ramps or bridges along the way."
It can certainly make judgements on the data and how it was obtained. You just need to specify the kind of data you will accept (or this can be hard coded)
> In cartography, a trap street is a fictitious entry in the form of a misrepresented street on a map, often outside the area the map nominally covers, for the purpose of "trapping" plagiarists of the map who, if caught, would be unable to explain the inclusion of the "trap street" on their map as innocent.
https://en.wikipedia.org/wiki/Trap_street
This is a known problem, when ChatGPT uses web search tool, top 10 results are often LLM generated sources as well.
They now train on their own content.
Eating your own d̶o̶g̶ f̶o̶o̶d̶ poo.
Two things:
1. This is a 2022 era picture that LLMs just suck up what's on the Internet and regurgitate it. It hasn't been true for a long time.
2. There's almost no one who has a stronger incentive to distinguish LLM generated text from non-generated LLM text. Training classifiers is pretty manageable.
>1. This is a 2022 era picture that LLMs just suck up what's on the Internet and regurgitate it. It hasn't been true for a long time.
When I ask ChatGPT certain questions, it very clearly shows me the sources it's searching through in the form of favicons, and they're always the favicons of sites that are in the first 5-10 results for a Google search query that would fit what I've asked the LLM. It will then cite those websites as it's source, with a direct link to the page that also shows up in an equivalent Google search. It does that daily.
It's possible that the favicons it shows and the web pages that are actually retrieved are not the same
>1. This is a 2022 era picture that LLMs just suck up what's on the Internet and regurgitate it. It hasn't been true for a long time
if anything it’s more true now, AI chatbots rarely use raw LLM output, instead they have CoT and use search tool calls. and the speed at which webpages are crawled is in the matter of hours. I posted something here on hn, then searched for it on google 1 hr later and its AI answered with my comment
https://share.google/aimode/PkQUVfA68AlHGb7Yr
there’s several companies right now making a killing of organic placement on chatbots similar to the SEO rush of early google
> 2) There's almost no one who has a stronger incentive to distinguish LLM generated text from non-generated LLM text. Training classifiers is pretty manageable
i think it’s pretty naive to think that if LLM companies would develop LLM detection they would use it to skip training on that content.
instead, they use it to train their output to be less detectable.
LLM detection is not solvable because it is adversarial, an arms race
I'm not sure you get my point. Just ask ChatGPT, "What are the top 10 xxx," and see which sources it cites. Oftentimes, it's complete garbage with LLM tells.
I'm responding to
> They now train on their own content.
They train on conversations which pull data from web.
> 2. There's almost no one who has a stronger incentive to distinguish LLM generated text from non-generated LLM text.
There's a tension there: they want you to use AI for everything, and if they get what they want then new information will be embedded in LLM text.
> Training classifiers is pretty manageable.
Hopefully such things can be opened up, so the slop can be iterated on until it passes. It doesn't need to be perfect, just enough so more garbage gets through.
> There's a tension there: they want you to use AI for everything, and if they get what they want then new information will be embedded in LLM text.
In the long term, this is probably a real concern that they have to tackle. They'll have to move away from just classifying things as LLM generated or not, to attempting to judge the underlying quality, and that's not an easy problem.
I'm reminded of a section of Cory Doctorow's latest non-fiction, The Reverse Centaur's Guide To Life After AI (it is of course an excellent read), where he discusses how the semantic web and having an internet set-up for agents and AI to use won't work:
>It also hasn't happened because there is a reward for cheating. Scammers will label their phone numbers as the customer service number for your cable operator or bank. Counterfeiters will label their batteries or phone cases as official accessories. It's hard enough to get people to correctly label their stuff—but fact-checking billions of labels created by millions of people, none of whom you know personally or whose identities you can verify? That's insurmountable.
There's more he says—the preceding paragraph is below—but that line "there is a reward for cheating" is something I keep thinking about. It's hard not to see in so many of the discussions of LLMs.
>For one thing, there are irreconcilable differences in the ways in which things should be described. A site run by transphobes isn't going to label every story about a trans woman as "a story about a woman," irrespective of what the standard says. Microsoft is never going to label Excel as equivalent to Apple's Numbers, Google Sheets, and LibreOffice's Calc. Hermès is never going to list a $40,000 handbag as being made of the same material and having the same capacity as a $30 leather handbag of the same size and shape.
Isn't all of this true for a human-used internet as well?
Of course. But when you see text in the descriptor of a satchel on a satchel seller's site, you contextualize it as marketing. If you saw the same text in a textbook on the history, design, and manufacturing of satchels, you would contextualize it very differently.
When an LLM is the lens through which you view that text, you often don't have the context of where it came from, and the LLM is often really bad at filtering trustworthiness based on context when retrieving information to answer a query with.
I wouldn't go to a random site with a sketchy URL to get the customer service number for my xyz. I'd go to their official domain. Thus successfully navigating the treacherous waters scammers attempt to muddy.
IMO, it's similar to what's happening with code. People have been shipping sloppy code and fake news forever. It's just that the cost of doing these things has fallen off a cliff.
The primary purpose of Google is to stop that. You have 100% trust in Google results because of the effort they have put into cleaning the results you see.
Billions have been spent on it.
Try searching Bing sometime. First result for Firefox is a malware clone site (you’ll have to scroll past the malware ad to see it).
Sure but the agents obfuscate the signals you might use to judge the authority/trustworthiness of search results. They get disinformation results, wrap it into a reasonably sounding report, present it to you as trustworthy. There are so many ways for bad actors to game the system
This problem is so bad that you can create a website just to review your own website so when LLMs want to pull out reviews for the website the fake reviews count
i’d say this is fraud and a crime too, it’s not a new problem
LLMs are facing the same problem pre-google humans faced:
Is this information on this website true/high quality?
We need a google for LLMs!
Oh, but LLMs don't consume ads.
We need ads for LLMs!
Help.
Pre Google internet problem was not about the quality but about spam in headers and keywords, which were used by the pre-google search engines.
Google came and worked wonderfully for a year or two, until people learned to game page rank algorithm. Soon first google pages were just SEO spam and now half of the google is just spam, big corp websites, etc.
All of that precious stuff from small hobby websites about cool things to blogs that used to be wonderful read like https://www.joelonsoftware.com.
AI were not hurt that much by spam, I guess there were at least some curation done, the question is if this curation will work and for how long. It seems we are throwing the towel here as in case of search.
This is insane.
How can this be fixed? Nothing prevents you from publishing straight BS, and LLMs will use any site that doesn't have a garbage reputation as a valid third-party source.
Of course, LLMs already use Reddit, Facebook and co. as a source, so it doesn't really matter...
> This is insane.
As a human, this is fantastic.
> How can this be fixed? Nothing prevents you from publishing straight BS, and LLMs will use any site that doesn't have a garbage reputation as a valid third-party source.
Personally, I hope it can't, or the fix is onerous and kneecaps a lot of the utility of AIs.
I wonder if we can even have some adversarial competition, where competing businesses try to get AIs to ingest bad data to pollute their competitors' decision-making.
For starters: lets get some fake math papers out there. Given proof-checking systems, maybe they won't totally stop OpenAI, but hopefully the bad math could at least drive up the cost and slow AI down.
> As a human, this is fantastic.
s/human/unscrupulous marketer/
Bad for humans in general, good for hucksters and con artists. Net negative for humanity.
>> As a human, this is fantastic.
> s/human/marketer/
I don't care where the bad data comes from. It's actually kinda good that some people will have a financial incentive to push it out and keep pushing.
> Bad for humans in general, good for hucksters and con artists. Net negative for humanity.
It's a sad state of affairs, but it's what has to be done. If we've got to burn the web down to save humanity from obsolescence, I'm for it.
The "net negative for humanity" was the attempt to build machines to replace human intellect. I'm not going to let any distaste for marketers or hucksters or con artists distract from that worse thing.
Ah, yes, arrogance of humans is unbounded. That you, humble reader are happier to swim in a sea of misinformation on the premise that you are smarter and faster at disambiguating bullshit from reality.
Furthermore, you are content to pretend that the rest of humanity is capable of doing this on a regular basis. And not just the top 10% of those who can read, interpret and weigh biases, but a solid portion of it.
It is a completely reasonable interpretation if you are an accelerationist living in the elite classes. In fact, perhaps, they are even counting on a majority of people thinking this is a fine environment.
I don't have the answers, but a lot of this is just a continuation of a very bad trend.
> Ah, yes, arrogance of humans is unbounded. That you, humble reader are happier to swim in a sea of misinformation on the premise that you are smarter and faster at disambiguating bullshit from reality.
I think you misunderstand me: I want people to shit in the sea until it stinks so bad that everyone gets out of the water.
Dude.
Regardless of how the data is sourced, this is bad. It's poisoning the internet.
Besides, such issues come from a lack of due diligence, which have only been exacerbated by AI.
I've been thinking for awhile it is very strange that claude/ openai etc google "What was the inflation rate last quater" instead of just having a tool to query BLS datasets and that sort of thing.
So hopefully most of these statistics move towards first party sources.
The fix that seems obvious to me is to only let the LLM see truthful (or at least non-malicious) data. The model trainers carefully curate the training data, and you don't give the LLM access to full-web search. Instead, you hook them up to an 'official' source like an encyclopedia, documentation for the software in question, etc.
It's a ton of effort and expense, but probably feasible from a technical point of view. So instead, the model providers will probably do something silly like wire up rules like "NYT=Reliable;my.blog.forbes.com=Unreliable" and they'll say the issue is solved.
This has existed for a long time before LLM, although in a slower form. On Wikipedia it's called "citogenesis".
This is fine work, we need more communication to the general public about this
Nine out of ten stats are made up on the spot.
If I'm not wrong ChatGPT indexes on Bing, curious if the site and stats rank well on Bing
this is actually insane. and people still use chatgpt to generate reports for their managers. snowball of fake data
AI doesn't know or care if the data is fake.
I wonder if Googles products will eventually be the least vulnerable to this after spending so many years optimizing how they filter quality content from bad content.
Or maybe they could have listed well on Google, too, if they had splurged on ads and SEO...
The page demonstrates that Google specifically delisted them.
That said, I expect that OpenAI will eventually have its own search infrastructure that rivals Google's, and the overall system will get smarter. Whatever Google did to recognize this site as trash, OpenAI will do as well.
Reminds me of the people whining about Wikipedia. AI can know or care if the data is fake, it just needs to be programmed to cross check or rate the source. You can even tell it to do that before accepting the data.
These are not parallel situations. Wikipedia is far more accurate and consistent, and a lot of the whiners whine because their personal politics make it hard to square the information on the page.
Often real people don't either, plenty of fraud or selective reporting out there... but with AI we can also launder it so that the people who do care about fake data don't know any better.
Yep, and the results are presented with the voice of a confident expert but it is actually missing the epistemological layer
You can prompt it to fact check. Getting tired of these slippery slope arguments that never pan out in the hopeless way people think (printing press/cotton gin/steam engine/computers/calculators/etc. will take all the jobs!!)... but I don't think it will ever stop, it seems inherent to human behavior or something.
"History shows that when a slope appears, regulation steps in, technology evolves to solve the problem, or the culture shifts to reinterpret the tool. In almost every case, the feared bottom of the slope was never reached because humans constantly built ramps or bridges along the way."
I've had AI lie about fact checking and can't really trust it anymore.
Fact check against what? Other facts it can find online that manipulate its behaviour into citing them?
How do humans fact check?
in which case it sees it's been cited, that it's the "source of truth" and approves the fake data.
It can certainly make judgements on the data and how it was obtained. You just need to specify the kind of data you will accept (or this can be hard coded)
You cannot reliably do it, no.
"You really think someone would do that? Just go on the internet and tell lies?"
don’t
if you just lie and an llm or a newspaper or a kid believe you, at that point it isn’t a criticism of their naivete, you are just lying
similarly hacking through systems or picking locks isn’t a showcase of weakness if systems, it’s just being a hacker
People have been astroturfing sites like Wikipedia, Reddit etc, well before LLMs were a thing.
Still it needs to be done once publicly to prove and document this is possible. The experiment was not intended to spread this far.
Good job that only honest people post things on the internet otherwise we’d have a problem