I think they have to; while at first glance AI generated content is impressive, it becomes noisy and sameish at larger scale or the more you read it. I'm of the opinion that AI is here to stay and a valuable tool for many use cases, but writing authoritative content ain't it. Wikipedia is fighting it (and other influences from before AI, like unsubstantiated claims and whatnot), I hope the scientific community and journalism steps up too.
Wikipedia is also a training source and shouldn’t also be a model output for content. Eventually it will be an oroborus, but it seems prudent to avoid that for as long as possible.
A little over a year ago, the MIT Technology Review ran a piece about how outputting AI content to non-English Wikipedia is rapidly corrupting the primary training data for uncommon languages because of exactly this process: https://www.technologyreview.com/2025/09/25/1124005/ai-wikip...
There will be a point in the future, possibly near future, when AI will be capable of reliably producing encyclopedic content, and we'll have tools to verify that the content produced passes reasonable standards.
Today is not the day. Especially not regarding the verification.
That's really not how probabilistic systems work though. They're great at getting 80% of the way to the goal but the last 20% is always excruciatingly difficult. I really don't intend that a knock because LLMs are useful, but they also aren't magic boxes where anything is possible.
This is the problem that a lot of user-content based sites, and open software projects, are experiencing. The tools make prodiving 80% correct submissions so easy that many people are doing that and thinking that it is fine not to bother verifying for correctness (or not even knowing how to verify correctness). This means those projects are being inundated by submissions that are 20% tosh or worse.
Another problem with user content sites, these numbers are completely pulled from the ass and are a meme that just gets repeated ad nauseam. They mean nothing, but there will be people that take it seriously.
Is 80/20 ever generally correct? Why is it favored over the 90/10 meme? Why would it apply to this? Why wouldn’t the numbers be vastly different for different domains of content generation? Etc.
If you pay attention, so much user content is just repetition of something they read on the same fucking website and uncritically checked into their brain.
Current HN top post is Roblox downloaded 145GB models from us (including adult content) without asking. I fully expect to start seeing comments to the tune of “Roblox is stealing porno to train AI for children” within the day in any discussion Roblox comes up in.
Why would these folk’s contribution to any other user content site be any more thoughtful?
Even putting aside accuracy, if the plan is for the tools that verify the written content to be in any shape or form outside of the human loop, then this feels like it glosses over the issues of both passive bias in data & active bias (e.g data filtering etc. ) when training such models that you'd be poisoning this information sources with.
Even subtle bias can manifest in significant ways. Given that the predominant market leaders in producing such models are currently : A country that's attempting to squeeze the word "america" in any geographic formation possible + rewriting history to put itself first, and a country that has perfected censorship en-mass and already actively doctors its own history.
I'm not convinced giving LLMs carte blanche to write authoritative sources on topics that - often are ideologically sensitive/controversial isn't just needlessly reckless and short-sighted. It's not a far reach to imagine such models reflecting the current views of administrations in these countries and becoming reluctant to mention certain colonial ambitions, environmental damage, political corruption, etc. where it might be relevant to authoritative content.
Wikipedia already struggles with issues in idealogical bias from authors - human validation (somewhat)helps mitigate that. You could argue that human validation is still prone to same bias - but you can always draw on a more diverse pool of editors with little ideological overlap. With LLM validation you're pretty much at the mercy of a pretty incestuous pool of data & a handful of self-serving corporates & governments.
Wikipedia is a tertiary source. Its editors are parroting secondary sources. (No disrespect intended.) There's no reason why LLMs couldn't be good at that.
I would rather be worried about the quality of all the primary and secondary sources going forward.
There is a reason why LLMs can not be good at this: They don't know what they are quoting or citing. A different kind of model with factoid provenance could do it?
What do you mean by this? I have passed very large documents into LLMs and they are able to accurately extract quotes and identify their location in the source documents. I don’t see any reason to think that they couldn’t accurately quote from and cite existing literature.
LLMs cook up new sentences based on probability, based on everything in their training and their context. There's just nothing in their model that traces provenance of facts, they physically don't store that information. Once they have written a sentence, they can't tell you from which work they got this information.
There are tools where they try to retroactively add references, the LLM interfaces try to do this, but it's usually inaccurate.
I think of the way a scientific article works. "The sky is blue (Pearson, 1992)." The writer is fully aware what the factual statement is and which publication supports it (not just that, but which kind of publication - primary, secondary source, peer reviewed?). An LLM can try to imitate this, but the actual information to support this does not exist in its training.
They can. I had one doing original OCR of a scanned piece yesterday. It found relevant page numbers and understood the context. It read the whole book to find the relevant cites. Even went though the genealogy table and found ground breaking references no one noticed before. Total cost under 5 dollars and time was less than 20 mins.
That's not how academic literature works. The structure is like this: Statement (Citation). The citation is not an affirmation of truth but a content declaration; this statement is supported by this publication. The receipts are there, so that you can trace the argumentation and make your own conclusions if necessary.
'Knowledge' formed this way is a living, evolving graph, and sometimes links are broken (findings are reevaluated or disproven); having the content declaration so clearly in the open makes it easier to deal with this evolving landscape.
I don't think anyone's arguing against whether or not it's possible, the question is whether it's valuable, correct, and A Good Thing.
I have access to a number of LLMs and can order it to build me a wiki site with all knowledge of the world, just like everyone else with the same access. What would make mine any more valuable or better than another?
> It was around December 2023. A few other editors and I were chatting on Wikipedia’s Discord server, and we had all seen very strange articles with the same odd characteristics.
I was afraid of that. I grabbed a 2024-04-20 copy as a backup, in case it turned to slop, but I suspected it might be tainted already.
About a decade ago, I read the history of AIDS on Wikipedia. It was a very informative (and satisfying) 20 minute read.
Now it's pages and pages and pages of text: probably factual, but not a satisfying read. It feels like it's for a different audience, and it's not satisfying to read.
Is it AI generated? I don't know for sure, but it feels like either someone (or something)'s missing the point of who Wikipedia's audience is; or that someone is following an agenda, (or that there's an edit war going on between dueling agendas and the facts are added to drown out the "AIDS is a consequence of bad morals" agenda.)
Wikipedia is an encyclopedia. Being as detailed as possible is its goal. Simple English Wikipedia exists for shorter easier to read versions of articles. Generally though, turning a complex topic into a fun easy read is the goal of good scientific journalism, which is also important but a separate effort from an encyclopedia.
No, that is not the point of an encyclopedia. The point is to condense an entry as much as possible while being thorough enough to cover most relevant information.
In some ways that was the limitation of physical paper and weight of the set.
A digital encyclopedia has different constraints, and the hard part was always trying to figure out what to leave out. Also in encyclopedias, you get what you get from the one or two people that worked on that particular article. If they had biases it would shape what everyone in the world saw. In an online article it can become a battle on what people want to see.
I just read both and I'm not sure I understand what your point is here. I don't have much to say about the diabetes page other than there is some interesting (as well as unfortunate) information. The HIV page certainly doesn't seem as if it gives off the "not a bad disease" feeling. It lists an untreated life expectancy as 11 years, whereas the diabetes page lists an expected life expectancy of being reduced by 10 years. That certainly makes HIV more deadly. It's also made obvious that it's transmissible from normal sexual behavior, as well as being so for some time before one would even experience symptoms. It's definitely not a bed time story. If anything, I'd say the diabetes page is less scary in its overall presentation than it should be.
Obviously this is subjective. I'm not saying you're wrong, just genuinely curious at what made you think this. I'm not making assumptions, but perhaps you have (unfortunately) more personal and upsetting experiences with HIV/AIDS than with diabetes? I know that when I used to bartend at a neighborhood gay bar frequented by gen Xers it was very difficult to see and hear people lose or talk about losing their partners to something that alienated them even more than they already are/were.
It is interesting to think about how to convey information about these things though. I think it's probably a good thing to make HIV/AIDS not sound so terrifying. For pretty much every adult alive today, I'd assume, their idea of HIV/AIDS was something like being told it's a "gay disease" or that you can't touch people with it (thank you for your service, princess Diana) or some sort of anti-African racist nonsense. So conveying accurate information without bigotry, homophobia, or xenophobia is probably a good thing overall.
Also curious what sort of possible conspiracy would benefit from Wikipedia readers from not being more afraid of HIV/AIDS.
Same feeling here - not sure if it’s AI, or zealous wikipedians + incentives that make addition easy but editing and pruning means stepping on many other writers turf, but I used to enjoy reading articles, while now many are more useful as “information fodder” for LLM to summarize and interpret.
If we leave out environmental (energy, datacenters, coolant) and political arguments (copyright stealing, private data harvesting etc.) even if we agree that those are "bad".
What then, is the problem of AI-Generated Content in Wikipedia?
> Fake or misleading citations have become a major worry.
Why is this not a standalone problem? Humans can fake and mislead in citations too, in fact, this is a major problem and has been so, for literally centuries in science.
> Being a place that was, in theory, free from AI slop has become part of Wikipedia’s identity.
Isn't this a separate issue in itself too? One of "quality". If AI generates poor quality content, than I would think the poor quality is the problem, not the producer of it.
The article mentions several times how hard it is to detect AI content. If it is that hard, then obviously AI is doing a good job and producing content that matches that of humans. Why then insist on one type of producer?
Poor quality is the problem and since the bandwidth of editors doesn't magically grow with the amount of input, rejecting AI content is a very effective way to keep average quality up.
>Isn't this a separate issue in itself too? One of "quality". If AI generates poor quality content, than I would think the poor quality is the problem, not the producer of it.
Yes, but who will review the mountains of slop to determine whether it is of acceptable quality? The main problem AI introduces is high volume at low quality, often overwhelming the human resource bottlenecks these systems have in place.
the thing is: if AI-Generated Content (if correct, obviously) is just some topics learnt from humans, what's the difference between something correct written by a machine and something correct written by a human?
Why to multiply instead of going to the same place for knowledge? It is almost like you want to make a different Wikipedia for conservatives and liberals because one cannot tolerate a tone or style of the other. Probably we should dedicate dozens of sites for the same topic depending on the writing styles of the humans too. There are a few out there.
In worst case we might have an 'AI language' (or languages) instead, so the sensitive souls could find a safe heaven from it elsewhere, choosing something else. Perhaps one along the idea of 'simple English', have it 'AI English'. Or something else satisfies those looking for a specific way of delivery they can handle exclusively.
I think knowledge is the important here, not the tone.
Honestly Wikipedia is just fine being written by humans. It's not the lack of copy that is holding Wikipedia back.
However,
> Perhaps one along the idea of 'simple English', have it 'AI English'.
That's actually an interesting idea. It could work, but I don't personally think it would be better than the English Wikipedia.
Also, are you expecting that humans would still be on the loop by deciding what gets added to AI English Wikipedia, or you think that LLMs should get to decide that too?
> It is almost like you want to make a different Wikipedia for conservatives and liberals because one cannot tolerate a tone or style of the other.
Many conservative worldview elements such as climate change denialism, creationism or anti LGBT view (e.g. LGBT being "unnatural") are not backed by reality because they run counter to scientifical consensus. And that's before we go into the outlandish bullshit conspiracy crap like "ZOG/zionist occupied government", "flat earth", "9/11 was faked", "the moon landing was faked" and the likes that gains more and more momentum among "conservative" people.
You cannot make a Wikipedia appealing to conservatives.
Not sure what question you're supposing he asked which was in the article. OP's point seemed to be that correct words written by a human could be effectively the same as correct words written by a machine, supposing the machine managed to be correct by learning from human content. The 'obvious' part is this wouldn't be true if the AI content was incorrect. At least that is what I got.
Ironically enough I just started an "Encylopedia of AI" project, and I've decided to generate all of the content with DeepSeek, with me being an editor/reviewer. I'm not really doing it for anyone except myself but I'm quite enjoying it so far. https://enkuklios.com/
Much of wikipedia is written by one person, Steven Pruitt. He has created more than 33,000 Wikipedia articles. People rightly credit him for cataloging obscure topics even though much of what he writes is rote.
But now when AI does it, it's deemed bad? If it's factual and at all useful to someone, why not leave it? I feel like we're waging the same battles when Wikipedia became an open encyclopedia. Surely you need to control who can edit right? We all know how that ended.
But he's edited a "third" of en.wikipedia according to his article which is genuinely astounding to me. Surely much of that has to have been automated?
If my memory was correct, a third of the edits is from him, not editing a third of it. There are a large and small edits. I think lot of it is mechanical and automated.
The problem with AI on Wikipedia is that the reason I've noticed it is because I've just seen blatantly incorrect stuff, mentioned breezily in passing, with no references. It's shitting up a thing that was already pretty easy to shit up if you had a motive, and your changes often wouldn't be noticed for months or years if an article wasn't well-trafficked.
AI doesn't even have a motive, it's just dumb - so you can't even track the mistakes to specific motivated users or networks of users. There's no way to clean it up.
Although I have to agree with you (probably, because I'm going to put words in your mouth) that this is a basic problem with wikipedia and how it is structured and governed, rather than a problem with AI. Anybody could always just put nonsense anywhere at any time; AI is just a chaos monkey. Wikipedia can't rely on groupthink to maintain itself, it has not worked and will not work, chaos monkeys are needed.
Today is not the day. Especially not regarding the verification.
Is 80/20 ever generally correct? Why is it favored over the 90/10 meme? Why would it apply to this? Why wouldn’t the numbers be vastly different for different domains of content generation? Etc.
If you pay attention, so much user content is just repetition of something they read on the same fucking website and uncritically checked into their brain.
Current HN top post is Roblox downloaded 145GB models from us (including adult content) without asking. I fully expect to start seeing comments to the tune of “Roblox is stealing porno to train AI for children” within the day in any discussion Roblox comes up in.
Why would these folk’s contribution to any other user content site be any more thoughtful?
As a result, I don't think LLMs will reach this state easily.
Even subtle bias can manifest in significant ways. Given that the predominant market leaders in producing such models are currently : A country that's attempting to squeeze the word "america" in any geographic formation possible + rewriting history to put itself first, and a country that has perfected censorship en-mass and already actively doctors its own history.
I'm not convinced giving LLMs carte blanche to write authoritative sources on topics that - often are ideologically sensitive/controversial isn't just needlessly reckless and short-sighted. It's not a far reach to imagine such models reflecting the current views of administrations in these countries and becoming reluctant to mention certain colonial ambitions, environmental damage, political corruption, etc. where it might be relevant to authoritative content.
Wikipedia already struggles with issues in idealogical bias from authors - human validation (somewhat)helps mitigate that. You could argue that human validation is still prone to same bias - but you can always draw on a more diverse pool of editors with little ideological overlap. With LLM validation you're pretty much at the mercy of a pretty incestuous pool of data & a handful of self-serving corporates & governments.
I would rather be worried about the quality of all the primary and secondary sources going forward.
What do you mean by this? I have passed very large documents into LLMs and they are able to accurately extract quotes and identify their location in the source documents. I don’t see any reason to think that they couldn’t accurately quote from and cite existing literature.
There are tools where they try to retroactively add references, the LLM interfaces try to do this, but it's usually inaccurate.
I think of the way a scientific article works. "The sky is blue (Pearson, 1992)." The writer is fully aware what the factual statement is and which publication supports it (not just that, but which kind of publication - primary, secondary source, peer reviewed?). An LLM can try to imitate this, but the actual information to support this does not exist in its training.
The hallucination era is rapidly ending.
'Knowledge' formed this way is a living, evolving graph, and sometimes links are broken (findings are reevaluated or disproven); having the content declaration so clearly in the open makes it easier to deal with this evolving landscape.
I have access to a number of LLMs and can order it to build me a wiki site with all knowledge of the world, just like everyone else with the same access. What would make mine any more valuable or better than another?
Funny though, as that admission of error makes it more reliable than if it tried to hide this fact.
I was afraid of that. I grabbed a 2024-04-20 copy as a backup, in case it turned to slop, but I suspected it might be tainted already.
Still, 2023-10 is probably the last pure https://en.wikipedia.org/wiki/Low-background_steel
"Low-background Steel: content without AI contamination (jgc.org)" 10.jun.2025 https://news.ycombinator.com/item?id=44239481 269 comments
"Wikipedia: Database Download (wikipedia.org)" 27.apr.2025 https://news.ycombinator.com/item?id=43811732 99 comments
"Low Background Steel (lowbackgroundsteel.ai)" 27.apr.2025 https://news.ycombinator.com/item?id=43814531 2 comments
Now it's pages and pages and pages of text: probably factual, but not a satisfying read. It feels like it's for a different audience, and it's not satisfying to read.
Is it AI generated? I don't know for sure, but it feels like either someone (or something)'s missing the point of who Wikipedia's audience is; or that someone is following an agenda, (or that there's an edit war going on between dueling agendas and the facts are added to drown out the "AIDS is a consequence of bad morals" agenda.)
A digital encyclopedia has different constraints, and the hard part was always trying to figure out what to leave out. Also in encyclopedias, you get what you get from the one or two people that worked on that particular article. If they had biases it would shape what everyone in the world saw. In an online article it can become a battle on what people want to see.
It is a difficult problem.
> The Simple English Wikipedia is for everyone, such as children and adults who are learning English
https://simple.wikipedia.org/wiki/Main_Page
> Other people may use the Simple English Wikipedia because simple language helps them to understand unfamiliar topics or complex ideas.
https://simple.wikipedia.org/wiki/Wikipedia:About#Simple_Eng...
There's a pretty stark difference between the first few paragraphs of the HIV/AIDS page and the first few paragraphs of the Type 2 Diabetes page.
Both are "factual", but when I read the HIV/AIDS page, I feel like someone is trying to convince me "HIV is not a bad disease".
Obviously this is subjective. I'm not saying you're wrong, just genuinely curious at what made you think this. I'm not making assumptions, but perhaps you have (unfortunately) more personal and upsetting experiences with HIV/AIDS than with diabetes? I know that when I used to bartend at a neighborhood gay bar frequented by gen Xers it was very difficult to see and hear people lose or talk about losing their partners to something that alienated them even more than they already are/were.
It is interesting to think about how to convey information about these things though. I think it's probably a good thing to make HIV/AIDS not sound so terrifying. For pretty much every adult alive today, I'd assume, their idea of HIV/AIDS was something like being told it's a "gay disease" or that you can't touch people with it (thank you for your service, princess Diana) or some sort of anti-African racist nonsense. So conveying accurate information without bigotry, homophobia, or xenophobia is probably a good thing overall.
Also curious what sort of possible conspiracy would benefit from Wikipedia readers from not being more afraid of HIV/AIDS.
What then, is the problem of AI-Generated Content in Wikipedia?
> Fake or misleading citations have become a major worry.
Why is this not a standalone problem? Humans can fake and mislead in citations too, in fact, this is a major problem and has been so, for literally centuries in science.
> Being a place that was, in theory, free from AI slop has become part of Wikipedia’s identity.
Isn't this a separate issue in itself too? One of "quality". If AI generates poor quality content, than I would think the poor quality is the problem, not the producer of it.
The article mentions several times how hard it is to detect AI content. If it is that hard, then obviously AI is doing a good job and producing content that matches that of humans. Why then insist on one type of producer?
Yes, but who will review the mountains of slop to determine whether it is of acceptable quality? The main problem AI introduces is high volume at low quality, often overwhelming the human resource bottlenecks these systems have in place.
In worst case we might have an 'AI language' (or languages) instead, so the sensitive souls could find a safe heaven from it elsewhere, choosing something else. Perhaps one along the idea of 'simple English', have it 'AI English'. Or something else satisfies those looking for a specific way of delivery they can handle exclusively.
I think knowledge is the important here, not the tone.
Why write more than a book instead of just putting everything into one book?
This is a way of thinking from which it is hard to get out, I have no idea how to point it to you.
Honestly Wikipedia is just fine being written by humans. It's not the lack of copy that is holding Wikipedia back.
However,
> Perhaps one along the idea of 'simple English', have it 'AI English'.
That's actually an interesting idea. It could work, but I don't personally think it would be better than the English Wikipedia.
Also, are you expecting that humans would still be on the loop by deciding what gets added to AI English Wikipedia, or you think that LLMs should get to decide that too?
Many conservative worldview elements such as climate change denialism, creationism or anti LGBT view (e.g. LGBT being "unnatural") are not backed by reality because they run counter to scientifical consensus. And that's before we go into the outlandish bullshit conspiracy crap like "ZOG/zionist occupied government", "flat earth", "9/11 was faked", "the moon landing was faked" and the likes that gains more and more momentum among "conservative" people.
You cannot make a Wikipedia appealing to conservatives.
Not to mention this exact question is being discussed in the article.
Can't wikipedia ask who and it what way writes content on their pages?
Correct information written by AI vs incorrect, biased information written by humans?
But now when AI does it, it's deemed bad? If it's factual and at all useful to someone, why not leave it? I feel like we're waging the same battles when Wikipedia became an open encyclopedia. Surely you need to control who can edit right? We all know how that ended.
https://en.wikipedia.org/wiki/Steven_Pruitt
There are over 7m articles in English Wikipedia.
AI doesn't even have a motive, it's just dumb - so you can't even track the mistakes to specific motivated users or networks of users. There's no way to clean it up.
Although I have to agree with you (probably, because I'm going to put words in your mouth) that this is a basic problem with wikipedia and how it is structured and governed, rather than a problem with AI. Anybody could always just put nonsense anywhere at any time; AI is just a chaos monkey. Wikipedia can't rely on groupthink to maintain itself, it has not worked and will not work, chaos monkeys are needed.