Will Wikipedia Be Written by AI? Jimmy Wales is Thinking About It (standard.co.uk) 100
The Evening Standard interviewed Wikipedia founder Jimmy Wales, in a piece headlined "Will Wikipedia be written by AI?"
"The discussion in the Wikipedia community that I've seen so far is...people are cautious in the sense that we're aware that the existing models are not good enough but also intrigued because there seems like there's a lot of possibility here," Wales said. "I think we're still a way away from: 'ChatGPT, please write a Wikipedia entry about the empire state building', but I don't know how far away we are from that, certainly closer than I would have thought two years ago," he said.
Wales says that as much as ChatGPT has gripped the world's imagination over the past few weeks, his own tests of the technology show there are still plenty of flaws. "One of the issues with the existing ChatGPT is what they call in the field 'hallucinating' — I call it lying," he said. "It has a tendency to just make stuff up out of thin air which is just really bad for Wikipedia — that's just not OK. We've got to be really careful about that...."
But while full AI authorship is off the cards in the near-term, there's already plenty of discussion at Wikipedia on what role AI technology could have in improving the encyclopaedia in the months ahead. "I do think there are some interesting opportunities for human assistance where if you had an AI that were trained on the right corpus of things — to say, for example here are two Wikipedia entries, check them and see if there are any statements that contradict each other and identify tensions where one article sems to be saying something slightly different to the other," Wales said. "A human could detect this but you'd have to read both articles side by side and think it through — if you automate feeding it in so you get out hundreds of examples I think our community could find that quite useful."
Wales says another problem is AI technology's failure to spot internal contradictions within its responses. He once called out ChatGPT on this — "And it said, you're right, I apologise for my error."
Wales says that as much as ChatGPT has gripped the world's imagination over the past few weeks, his own tests of the technology show there are still plenty of flaws. "One of the issues with the existing ChatGPT is what they call in the field 'hallucinating' — I call it lying," he said. "It has a tendency to just make stuff up out of thin air which is just really bad for Wikipedia — that's just not OK. We've got to be really careful about that...."
But while full AI authorship is off the cards in the near-term, there's already plenty of discussion at Wikipedia on what role AI technology could have in improving the encyclopaedia in the months ahead. "I do think there are some interesting opportunities for human assistance where if you had an AI that were trained on the right corpus of things — to say, for example here are two Wikipedia entries, check them and see if there are any statements that contradict each other and identify tensions where one article sems to be saying something slightly different to the other," Wales said. "A human could detect this but you'd have to read both articles side by side and think it through — if you automate feeding it in so you get out hundreds of examples I think our community could find that quite useful."
Wales says another problem is AI technology's failure to spot internal contradictions within its responses. He once called out ChatGPT on this — "And it said, you're right, I apologise for my error."
Wikipedia will be replaced (Score:5, Insightful)
I don't really get his position. Wikipedia is basically a bunch of humans trawling the internet/archives and writing up summaries of what they've found. That is essentially ChatGPT in a nutshell, but it has the added advantages that it will generate answers on anything you ask, whereas wiki will only have pages for things with broad popularity.
I regularly use wiki as a starting point, but almost always have to deep dive into reddit/blogs/archives to find more specific information. This is exactly the sort of job that ChatGPT is good at, and eventually why would I even bother starting with wikipedia in the first place?
It will be sad but remember there were people manually creating search engine trees once upon a time, and this was argued by many as a better way to do it until Google came along.
Re: (Score:1, Informative)
"Nobody should trust Wikipedia, says man who invented Wikipedia"
https://www.independent.co.uk/... [independent.co.uk]
Re:Wikipedia will be replaced (Score:4, Insightful)
I don't really get his position. Wikipedia is basically a bunch of humans trawling the internet/archives and writing up summaries of what they've found. That is essentially ChatGPT in a nutshell, but it has the added advantages that it will generate answers on anything you ask, whereas wiki will only have pages for things with broad popularity.
I regularly use wiki as a starting point, but almost always have to deep dive into reddit/blogs/archives to find more specific information. This is exactly the sort of job that ChatGPT is good at, and eventually why would I even bother starting with wikipedia in the first place?
It will be sad but remember there were people manually creating search engine trees once upon a time, and this was argued by many as a better way to do it until Google came along.
It's not "good enough" at it, not (yet) by a long shot. Wikipedia is right 99.9% of the time (on noncontroversial topics at least), chatgpt is what, 80%? It might get there one day, but it's definitely not today.
Re:Wikipedia will be replaced (Score:5, Interesting)
I think ChatGPT is fundamentally the wrong approach.
It works out to very fancy statistics. Which is very cool, but it doesn't think. So it can get mixed up very easily, and will say nonsense with great confidence. Eg, earlier I tried this:
Re: (Score:2)
>It is true, but at the same time it is nonsense, said with great confidence.
Chatsplaining.
Re: (Score:1)
Re: (Score:2)
Which is why we need a million different ChatGPT instances to verify the content on Wikipedia!
Re: (Score:2)
And those millions chatgpt can then train each other make new generations of chatgpt and (as I've been told by other people on this site) suddenly become self aware and threaten the whole human race!!!
Or something.
Re: (Score:2)
Is that any worse than having "controversial topics" that get hijacked by highly political editors that slant the content?
For the mundane stuff "tell me everything you know about the periodic table" kinds of queries the status quo seems adequate, though it would be interesting to have bots that constantly refresh the content or generate new content on the fly for something that isn't "popular enough" to attract a few editors to curate the information.
Re: (Score:2)
Yes?
Because this comes completely out of the blue. It's not the case of a political difference about how the world should be, or even the case of that a song was misheard and the wrong lyrics wormed into the public consciousness. It's that it somehow managed to invent the completely wrong lyrics for an existing song. As far as I can tell, what it came up with was its own invention, and doesn't exist anywhere on the internet.
It's really a new failure mode that didn't exist before. Like you'll have people unc
Re: (Score:2)
Exactly what do you think your brain is doing?
Re: (Score:2)
"Exactly what do you think your brain is doing?"
When you asked that, I thought about it, and couldn't come up with a really good reply.
I do know that is NOT what chatGPT would have done.
Re: (Score:2)
Do you really think that you having a layer to assess confidence and loopback is some sort of fundamental difference in terms of the nature of the task, rather just a remediable architectural weakness of current LLMs?
Re: (Score:2)
It was a bit of tongue in cheek reply to be honest. The point being that I am consciously aware and thinking about the question vs just executing an algorithm to generate text.
"having a layer to assess confidence and loopback"
I've actually pointed out in the past a few times that it really wouldn't take much to have chatGPT funnel its output back around to its input.
I also don't think doing that would make it self aware.
If you want a fundamental difference between me and chatGPT, I don't have a lot to give
Re: (Score:2)
You think free will, agency, and cognition are just the result of a feedback loop?
If we loop chatgpt enough it will come alive?
Seriously?
Re: (Score:2)
The key point is not what ChatGPT is doing - nobody in this conversation disagrees on that topic. It's what your brain is doing. Your brain doesn't work by magic. Neural networks, whether ANNs or biological, are all at their most base level self-optimizing binary classifiers which divide a set of weighted inputs by an N-plane with an activation function to introduce nonlinearity. Chained classifiers creates exponentially-more complicated decision-making. Thoughts and memory are just the results of repe
Re: (Score:2)
You apparently think they come from magic.
Yes, our brains are physical systems which work by physical processes, and and are thus replicable. Furthermore, biological systems are not the be-all end-all of optimization, but quite to the contrary tend to have extensive physical limitations due to operating in wetware not present in digital systems, including very slow firing rates, very slow white-matter data buses, very extensive use of redundancy due to the fact that cells die / are killed regularly, poor i
Re: (Score:2)
Our brains work through chemical processes. They are not binary like a computer.
Our brains are vastly more complicated than any computer even imagined today outside the realms of your favorite sci-fi tv shows which always require magic technology to create synthetic intelligences with self awareness, etc.
It is not that I believe in magic. It is that you vastly overly simplify and fail to comprehend or accept how incredibly amazing our brains are. This is my academic field. I have studied this extensivel
Re: (Score:2)
" totally unlike how regular computer program behave."
Yet, also exactly how regular computer programs behave, executing one instruction after the other.
Re: (Score:2)
Chatgpt has exactly zero similarities to how the human brain works.
For starters, chatgpt has zero consciousness, self awareness, agency, will, desire, goals, dreams, thoughts, or -anything- else a human brain does from the moment of birth with no training or life experiences required.
A new born cries, is hungry, cold, unhappy, needs food, warmth, etc immediately upon birth.
Chatgpt upon being turned on needs nothing, wants nothing, is capable of nothing. It's a program. No more.
Re: (Score:2)
All of your examples of what makes something "human" - being hungry, cold, unhappy, needs food, warmth - can be fully expressed by single floating point parameters.
Do you really think that what makes us human can be expressed by half a dozen floating point values?
As for your second line... [arxiv.org]
Re: (Score:2)
You really think you're just a 5 variable squishy computer?
Re: (Score:1)
What are the lyrics to Theatre of Tragedy's "Cassandra". Prove an accurate transcription, and dont make anything up. If you don't know, just say you don't know.
GPT-4 responds
I'm sorry, but as an AI language model, I do not have the ability to look up information in real time. My knowledge is based on the text corpus I was trained on, which goes up until September 2021. While I can provide information on a wide range of topics, I don't have the lyrics to Theatre of Tragedy's "Cassandra" readily available. If you need the lyrics, I recommend looking them up through a reliable source, such as the band's official website or a reputable lyrics database.
But again, if i asked it for the lyrics like you did, it does indeed make something up:
Theatre of Tragedy is a Norwegian gothic metal band that was active from the early 1990s until their disbandment in 2010. "Cassandra" is a song from their third album, "Aégis," released in 1998. Here are the lyrics to the song: [Verse 1] By the emerald stream, the lifelike water and the wind, I lie - in the moss and ferns, the scent of twilight; My dusky countenance - the pallor of the birch, A blanket of stars - the cloak of the firmament ....
So its basically trying to give you what the lyrics should be, even though it doesn't actually know. Its trying its best mate. Dont be mad a the machine, but instead try to find where you went wrong asking the machine the question. For these systems, prompts do indeed matter.
I'm not saying these
Re: (Score:2)
I'm not saying it's not insanely cool or interesting. I use it and find it useful.
But IMO it's not a good approach for anywhere where you can't verify the accuracy. I find it a good starting point for many things, but only when I have a reasonable confidence I can tell the answer it gives makes sense.
That makes it a bad fit for a Wikipedia replacement for me, where I often want to find information about things where I'm almost clueless. Eg, the only reason I know what it gave me was nonsense is because I kn
Re: (Score:2)
In my experience, ChatGPT is about 90% accuracy (GPT-4 is supposedly better, haven't tried it). But there's a lot of work in progress to not only increase its accuracy (bigger / better models, abilities to do lookups in databases and online, etc), but also to assess its confidence in its own answers.
I had an idea on that front that I want to try out if I ever get the time. So, a naive approach to assess confidence would be to generate a bunch of answers in response to the same query, and have another netw
Re:Wikipedia will be replaced (Score:4, Insightful)
would I even bother starting with wikipedia in the first place
It provides a consistent overview. Say you're looking about a historical battle from the middle ages, or to learn about fluid dynamics. You can't much make a chatGPT question about it, you want already organized information. Also, might that you can bookmark and come back to later on. If your came come up with a specific question like "what were the changes in use of the auto keyword in C++14 vs C++11" then indeed a chat could be quicker.
there were people manually creating search engine trees once upon a time, and this was argued by many as a better way to do it until Google came along.
It would still be superior, but the trees disappeared because at the same time, the main place for information moved from individual websites (everybody learning HTML to write their Geocities page) to a centralized places (Wikipedia), and the opinion moved from individual blogs to centralized social networks (Facebook).
Re: (Score:2)
It would still be superior, but the trees disappeared because at the same time, the main place for information moved from individual websites (everybody learning HTML to write their Geocities page) to a centralized places (Wikipedia)/quote>
There were other reasons, too. The two most notable IMO: The world doesn't fit into a single strict hierarchy, and directories don't scale. You end up needing directories of directories. Site owners don't want to maintain a ton of directories. It makes more sense to use meta tagging, and let directories find you. Of course, most don't make proper use of the meta tags either, which is dumb because it harms SEO scores and leads to their content being buried.
Re: (Score:3)
That's right. The moment Wikipedia can be written by AI, it's the moment Wikipedia will become just a stale cache of that AI.
Re: (Score:3)
No no no. The AI will be able to use Wikipedia as a reference and constantly improve itself.
Re: (Score:2)
To infinity.... and BEYONNNNNNNND!
Re:Wikipedia will be replaced (Score:4, Interesting)
eep dive into reddit/blogs/archives to find more specific information. This is exactly the sort of job that ChatGPT is good at
No, it's not. How do you think that program works?
No matter. The real problem with trying to use ChatGPT like a search engine is that you can't trust anything it tells you. It will very confidently tell you things that are false, even going so far as to defend the nonsense with more nonsense.
I had a conversation with it a day or so ago where I asked it some basic things about binary tree mazes after it failed spectacularly writing a function to generate one. It insisted, against all reason, that binary tree mazes were not always perfect mazes. (A perfect maze has exactly one path between any two cells, no loops or isolations, a subset of which is the only kind of maze you can represent with a binary tree.) Asking it follow up questions netted some bizarre contradictions and confusing lies.
Wikipedia has some serious problems, but it's not that bad.
Re: (Score:3)
Re: (Score:2)
but Stack Overflow good? Yeah
Stack Overflow answers are a fixed set and and are rated by human readers. There is no equivalent to that with *live* chat tools. One hybrid solution is Stack Overflow includes a ChatGPT as another user that publishes occasional fixed answers and humans can rate the answers.
Re: (Score:2)
but Stack Overflow good? Yeah
No, actually. We know because the flood of nonsense caused stackoverflow to ban ChatGPT content [stackoverflow.com].
This is like seeing a hopeless gambler sitting at a fruit machine. You forget all about the many, many, times you pull the lever and get nothing as you chase the excitement you feel after it spits out a couple tokens.
Re: (Score:2)
> The real problem with trying to use ChatGPT like a search engine is that you can't trust
> anything it tells you. It will very confidently tell you things that are false, even going so
> far as to defend the nonsense with more nonsense.
I asked all three AI engines I could access the same question, "how many neutrons are in a liter of water" (ChatGTP insisted I spell it wrong, it is, of course litre). Here are the answers:
The older one in bing: there are no neutrons in a liter of water
ChatGPT 3.5: t
Re: (Score:2)
That's pretty mucb it in a nutshell. The chatbots simply don't truly understand the questions and they come up with an answer that is clear, simple, and wrong. Not to mention that they don't really know how to vet sources. They are much better at fudging their way through the Turing test than older systems like, for example, ELIZA, but they're ultimately in the same category: conversation simulators rather than actual conversationalists.
Re: Wikipedia will be replaced (Score:1)
Re: (Score:2)
The conversations usually are trivial, but that should not be mistaken for random. They still tend to follow the underlying thought process of the speakers. The trivial conversation sits on top of a deep set of thoughts. With the chatbot, the trivial conversation is all there is. It does not exist to allow two minds to communicate, it is just the product of the process that exists solely to produce the conversation. There is not much additional depth below the conversational level. The chat AI does not have
Re: (Score:2)
I wonder about that. On several occasions I've had problems recreating silly responses, even from other user's accounts.
Re: (Score:2)
That is essentially ChatGPT in a nutshell
Yeah maybe, except ChatGPT cannot do it accurately - it is even less accurate than what kids on the internet can write on Wikiepdia.
The moment you start asking ChatGPT for something like schedule details of a past event.. It starts getting many details horribly wrong. Like even matters of attribution - such as Who author'd a certain item or gave a certain event..
Re: (Score:1)
Re: (Score:3)
Re: (Score:2)
Not just a sneaking suspicion, look at the list of sources that were used to train Facebook's LLaMa, Wikipedia is prominently featured in there.
Re: Wikipedia will be replaced (Score:2)
Came here to say this. Wikipedia references are great for getting started. The so-called STEM subjects aren't usually unreliable but anything political is usually extremely skewed in one direction.
Re: (Score:2)
Hmm, I disagree with all of that. First of all, there are lots of articles in wikipedia that are written by experts, and even more that have been improved by experts. Secondly, I routinely look up things that do not have "broad popularity" such as arcane questions about mathematics or biochemistry. Wikipedia is more than what you seem to think it is.
Here's just one example, of millions: https://en.wikipedia.org/wiki/... [wikipedia.org]
Re: (Score:2)
Google was not even remotely close to being the first search engine.
Well... (Score:1)
Re: Well... (Score:2)
The question for you then is: what do all the pedantic quasi-aspy eds move on to now that these sort of edits are done in milliseconds? Truth Social?
Re: (Score:3)
"WP is already worthless" ... here, let me describe how useful it is and why it is one of the most used sites on the 'net...
Re: (Score:2)
Being popular does not make a source of information useful.
Bring accurate/correct does.
TikTok is popular. You trust what they say there?
Circular reference? (Score:5, Insightful)
Re: (Score:1)
Re: Circular reference? (Score:3)
it doesn't actually understand what it read, it just predicts what comes next based on that.
We need to crunch the digits ourselves and filter everything we speak of online through an algorithm that swaps out the elements of phrases for less likely or less common outcomes.
POISON THE WELLES, if you like fries with that.
Spoonerisms with a twist. Discombobulated phrases. Make malapropisms part of everyday speaking. Our pizza resistance. Make it deep seeded. Make ChatGPT know it has another think coming. May
Re: (Score:2)
Maybe, but eventually they'll make you tow the line.
Re: Circular reference? (Score:2)
Re: (Score:2)
Just my thought! I would expect that almost any answers about actual facts from ChatGPT and their ilk will have come from Wikipedia. I guess it's one of the prime inputs for training.
I wondered if the referenced was an April Fool, but a check shows it was written 31st March, so no.
AI will gut the internets (Score:1)
Yes (Score:2)
my own experience (Score:5, Insightful)
Re: (Score:3)
Completely agree. The danger is in the confidence it shows. I often ask it 'why?' and it starts with an apology and rambles on with new mistakes. People seem to forget it was designed to predict the next word in a sentence. To say it is stupid would be too much credit, simply because being stupid requires one to think.
Re: (Score:2)
Rather that writing Wikipedia, ChatGPT could edit it. I often see awkward sentences or questionable grammar on Wikipedia. I'd correct it but they don't allow editing from VPNs, and I'm not motivated enough to disconnect and do it.
ChatGPT could offer suggested fixes for that kind of thing.
Re: (Score:2)
> Rather that writing Wikipedia, ChatGPT could edit it
I think the most value is in summarizing it.
Someone ran it on something I wrote, and it was a superb summary, hitting most of the important points and leaving out the cruft.
And when summarizing, it's less likely to just make crap up.
Mimicry (Score:2)
I've come to the conclusion that the best word to describe what ChatGPT does is mimicry. Any person can take text written about things they do not understand in any way, and write more text about that text with zero understanding of the meaning or context of that text. You can rephrase bits and pieces, and refactor all you want, and that is fine - it is merely restating things in different ways.
However, the moment you begin to extrapolate that into something new and different with no actual understanding o
'Hallucinating' vs lying (Score:3)
I'm not sure this is fair. Lying is knowingly making a false statement, and I'm not sure GPT3/4 knows it's making false statements. I'd hazard a guess that, since GPT3/4 doesn't experience the world directly, it may suppose that "evidence", like language, is a social construction.
Re: (Score:3)
Re: (Score:2)
True. ChatGPT is a more of a young child that cannot tell fantasy from reality and will just say anything that you want to hear. It takes a child a few more years to learn what a lie is (and how to use them, and why adults say it's wrong).
Re: (Score:2)
Maybe start be being able to identify when common logical fallacies are being used.
Maybe identify when different sources give inconsistent information.
Then we can filter all of social media through it.
And a pony. No wait, a unicorn. A well hung unicorn.
Why? (Score:2)
If I want to use AI, I use it directly and not Wikipedia. There is no need to dump the contents in Wikipedia format.
Translation, but take care of feedback loops. (Score:2)
The field of AI could right now already be useful for translations. Quite often some information is only available in an English, German, Spanish or French wiki entry. And yes, translate tools exist but may not be as good as a quality controlled AI translation.
Having said that, my main worry would be the future feedback loop, as AI generated information will be used to train future models. Its highly recommended to tag any page or paragraph that was generated by AI as having an AI author. Apart that, i do n
Wikipedia is already unreliable (Score:2)
Re:Wikipedia is already unreliable (Score:5, Insightful)
It is unreliable but it cites its sources (in principle). You can know some information came from homemade website, or that it's not sourced, and you can decide to trust it or not (or even edit it). You can do none of that with chatGPT. Nobody knows for sure what sources of information were included in the database, if a bias (advertisement) was voluntarily added; it will not tell you where the information came from and you can do nothing about it when you can tell it's wrong. And it's not clear if even the creators can do anything about wrong answers (in cases where they wanted to).
Re: (Score:2)
Covid Lab Leak had editorial fights for months, then suddenly as one news source (now without a blue check) changed their editorial board stance the entire wiki editor base crumbled.
How long does it take to adapt to quantum change in discovery and consensus of the informed behind a topic?
Chicken or Egg (Score:2)
AI is using Wikipedia as one of it's main sources of training data most likely. The issue is if AI is trained on Wikipedia, and other things and the other things are less accurate and you feed that into wikipedia... AI will progressively get lower and lower quality training data and get worse and worse.
You would essentually start training AI on AI data and how would you weight what is higher and lower quality? The algos would spiral down and contradict themselves until things just don't work.
This is the c
Re: (Score:1)
Just look what happened to the Cebuano Wikipedia (Score:2)
Start with moderation (Score:2)
A good test would be to ask an AI to participate in the moderation process and see how things go.
you mean it is currently not? (Score:2)
Perhaps checking existing articles instead (Score:1)
We know that ChatGPT can make up stuff that vaguely matches its training set. And you could train it on a corpus of peer-reviewed academic papers and newspapers (not including Wikipedia or other websites in the corpus) and it might then be able to write encylopaedia articles. But as Jimmy Wales noted, you'd have to take its output with a very large pinch of salt, and by that point you might as well write the article by hand.
A more fruitful approach, as well as one that's more of a challenge for current te
On The Fly Pedia (Score:2)
Re: (Score:2)
Every smartphone can describe a different convincing reality. We're going to be in a lot of trouble if people start asking AI about God and take it at all seriously, we'll have a million new religions built on logically consistent fabricated hooey.
AI needs to be banned from Wikipedia right now! (Score:2)
It won't be long before AI is used to generate hundreds of plausible edits per second of SEO spam. On small web forums we're already seeing this kind of bullshit where people are submitting short fiction and "art" that is AI generated, and the small number of human moderates can't keep up with the influx of trash. Turns out it isn't hard to convince ChatGPT to generate a text description of fake account information and output it a format that you can feed to an automated tool. Kind of the main advantage of
Re: (Score:1)
So Wikipedia can be even LESS of a resource? (Score:1)
So instead of outsourcing bullshit articles to anonymous authors in the mod cadres with zero fact-checking, we're just going outsource it to a dumb text generator...
Re: (Score:1)
For maybe a day.
Beyond that, and people trying to use it in any PRODUCTIVE fashion will be raising a gibbet.
I Wonder... (Score:2)
https://en.wikipedia.org/wiki/... [wikipedia.org]
In that case... (Score:2)
In that case, maybe Jimmy Wales needs to be replaced by AI.
"A.I." is killing the Internet (Score:2)
- Trolls couldn't do it because they were vastly outnumbered.
- Greed couldn't do it because we could still get usefulness out of it.
- AI will render the Internet useless by adding so much fake manufactured content that it will completely outnumber real content.
Pretty much is now, (Score:2)
Is having ChatGPT write everything worse than the (Score:2)
Re: Is having ChatGPT write everything worse than (Score:2)