Raf Guns
Shafik YaghmourThis is a damning article from the Wikipedia editors on GenAI articles written for Wikipedia: https://wikiedu.org/blog/2026/01/29/generative-ai-and-wikipedia-editing-what-we-learned-in-2025/
Ok this has left my network.
I am better known for my C++ quizzes:
https://hachyderm.io/tags/Cpppolls
and my cursed code:
https://hachyderm.io/search?q=from%3Ashafik+cursed+code&type=statuses
Also my dad jokes but I don't have handy reference for those, so you will have to just follow if you like dad jokes.
"AI Translations Are Adding ‘Hallucinations’ to Wikipedia Articles": https://www.404media.co/ai-translations-are-adding-hallucinations-to-wikipedia-articles/
Arthfach 🐻@shafik This also implies to me that Wikipedia is less and less trustworthy as time goes on and AI slop replaces researched material.
Martin Vermeer 86 47
Alex ☕🇨🇦@shafik this aligns with my experience, where an AI "summary" will turn up a useful reference, but the reference does not corroborate the summary.
That is how I use the summaries, to obtain vocabulary and references and I tend to ignore the content as once I know the vocabulary and good source.
I can just read those and no have to verify the summary, which you always will need to do.
Martin Rundkvist
Cybermatron@shafik Welcome to the life of every University lecturer marking essays right now.
I am sorry
@shafik Honestly, we all are 😒 It is very exhausting and dispiriting to have to read so much mediocrity and blatant falsehoods. Makes marking essays take twice as long and feel like wading through treacle.
The verbosity of LLMs is a big drawback. In the open source community we are struggling w/ LLM submissions for many reasons but their sometimes extraordinary length compounds the problem. It just takes so much longer to get through them.
Stephen Bannasch (316 ppm)@shafik
Interesting!
“Our early interventions with participants who were flagged as using generative AI for exercises that would not enter mainspace seemed to head off their future use of generative AI. We supported 6,357 new editors in fall 2025, and only 217 of them (or 3%) had multiple AI alerts. Only 5% of the participants we supported had mainspace AI alerts. That means thousands of participants successfully edited Wikipedia without using generative AI to draft their content.”
Interesting!
“Our early interventions with participants who were flagged as using generative AI for exercises that would not enter mainspace seemed to head off their future use of generative AI. We supported 6,357 new editors in fall 2025, and only 217 of them (or 3%) had multiple AI alerts. Only 5% of the participants we supported had mainspace AI alerts. That means thousands of participants successfully edited Wikipedia without using generative AI to draft their content.”
Yes, education definitely helps. There have been several reports that if people using LLMs in a way that makes their flaws obvious they learn and use them in more appropriate ways.
If they can keep up, it should not be a big problem. Sounds like they have a hold of it.
Looking for explanations…It’s not all bad. They do refer to where AI can be useful, but clearly copying & pasting from it is not recommended. A good read for teachers.
For sure, tools are useful when you use them appropriately.
Martin Escardo@shafik Yes. Indeed.
Today I opened Claude to try to find a reference for something I know is true, but is not original with me, to cite in a paper I am writing.
The first answer was a proof, which (in this particular case) was correct.
But then I told it that I didn't want a proof, only a reference to cite. I had told it in advance that I already knew it is true.
So it gave me a reference. When I looked at it, there was nothing in there stating or proving what I wanted.
So I complained and I got an "apology" (I am not sure machines can or are even entitled to apologize - at best, they should apologize on behalf of their creators).
Then it tried again, and it again gave me a reference that didn't have what I wanted.
The third time I tried, it said it gave up, that what I wanted is nowhere to be found in the literature. But this is wrong. I've seen it before, I know it is true because I can prove it (and Claude itself can prove it (correctly this time), but course not out of nothing).
Don't ever trust a reference given by genAI unless you check it yourself. The references I got after explicitly asking for a reference, and nothing else, didn't have what I asked for.
The machine just makes things up in a probabilistic way. When it starts "apologizing" then you can know for sure that it is rather unlikely that you will get anything useful from it.
Even more concerning is if it doesn't apologize. You may suppose that the answer is right and use it for whatever purpose you had in mind. Good luck with that.
The Turtle@MartinEscardo @shafik google Gemini can be browbeaten into an apology and then will go right back and do the same shit over again exactly.
Dr Jess Birch@shafik This certainly mirrors my personal experience (and why I'll keep calling them lie-bots) where all these tools added "citations" and "links" (in Google, Bing, etc "AI summary") but it's not "real" because it's just looking for a plausible real link not one that actually says what it has imagined (as a summary).
If you want specific lists of things, you can get all sorts of links attached (the lists are wrong) or just the weather (link will usually contradict the claim).
If you want specific lists of things, you can get all sorts of links attached (the lists are wrong) or just the weather (link will usually contradict the claim).
Luc@shafik the few times I've looked up references on Wikipedia, the experience was similar and I don't think this was related to LLMs having written the thing I wanted to learn more about but turned out to not be mentioned at all in the alleged source
A comparative study might be better than a blanket "it's frequently wrong". Of course I expect LLMs to perform worse than humans but the context would be helpful to put the numbers into perspective
I have never had this experience and use Wikipedia extensively for digging into science, math, history, economics etc
Outside of history these are mostly subject I have studied extensively and was looking into areas I was not familiar with but knowledgeable enough to know problems if I saw them.
So please provide examples and please provide these to the Wikipedians as well so they have something to sink their teeth into.
@shafik fair enough, the number of times I look at sources on Wikipedia in my life is probably countable on one's fingers and I definitely don't remember what page I was on, say, 1–3 years ago when this last occurred. (Also can't see it from edit history since, if the source doesn't say and I'm there for information, evidently I'm not qualified to write on the topic)
Will bookmark this post and try to remember to reply when I next check a reference and find that it checks out or not!
A-nom-nom-nom-aly BSC SSCThat's probably because most genai is referencing his Nazinesses grokipedia, which has been proven to be full of bullshit and white supremacy rhetoric.
antimu0n 🤘 metal lynx ♿ 🏳️🌈🏳️⚧️@shafik
It's becoming clearer every day that every AI datacenter on the planet must be destroyed. Wether by military saturation bombing, legal action, individual sabotage, or (preferably) a mass movement of millions marching onto premises with hammers and blowtorches, it needs to happen, and soon.
It's becoming clearer every day that every AI datacenter on the planet must be destroyed. Wether by military saturation bombing, legal action, individual sabotage, or (preferably) a mass movement of millions marching onto premises with hammers and blowtorches, it needs to happen, and soon.
gkrnours
Daniel Demmel@shafik I think this is an excellent article without the usual hype or hate, clearly articulating appropriate use cases and setting the right expectations. Wikipedia might end up being not just a great encyclopedia, but an organisation that figures out (under duress) how to incorporate LLMs into writing about facts that is helping instead of destroying the process.
It is very well written and I think their recommendations are super helpful for educators trying to figure out how to reset folks use of these tools.
Fredrik@shafik yes, an LLM doesnt known if its knowledge is true or false, its just information that looks the same to it. Training an LLM is already expensive, and having it check information on sources every time it build an answer would be more expensive still and I don't think the companies care about that.