Making sure AI serves people and knowledge stays human
diff.wikimedia.org119 points by benbreen a day ago
119 points by benbreen a day ago
I wonder if the timing of this announcement with regard to Musk's announcement of Grokipedia is coincidental. https://gizmodo.com/elon-musks-wikipedia-competitor-is-going...
Yes, I am interested to see if Grokpedia can live up to the initial hype. As much as I dislike wikipedia it is a difficult thing to get right. In particular where lots of political and business interests (and others, etc) apply pressure to manicure pages. With everyone banning everyone they don't like it may just turn into another boring partisan internet split.
Pages about companies and contemporary politics make up a relatively small percentage of Wikipedia. And what's the realistic alternative? These pages are always going to represent opinions. Traditional encyclopedias usually represent views closer to the establishment mainstream, but that's just a function of who gets a job at Britannica versus who has the time to edit Wikipedia all day.
Will Elonpedia be better in that respect? Its owner is not exactly known from having a healthy distance to internet culture wars.
More importantly, even if all these pages disappear overnight, Wikipedia is still extremely valuable and beneficial.
Yes but with AI, in principle, there is more wriggle room to present a spectrum of opinions on a single topic. It is all about what is baked into the models. Like can it discuss x topic from multiple competing angles and capture the complexity. That is important for history, politics, medicine, science and so on. This does not currently happen on many wikipedia pages in these domains imo.
Narrative engineering with AI is a pretty scary prospect so I can see how the wikipedia model with genuinely random human editors might have a major advantage there. If it is simply a vanity project, and not an uncensored Grok model, then obviously it would be garbage and hilariously biased.
> Yes but with AI, in principle, there is more wriggle room to present a spectrum of opinions on a single topic.
I mean, this sounds impartial but isn't. If you have an article on the planet Earth and present a spectrum of opinions on whether it's round or flat, you're not being balanced; you're implicitly supporting flat-Earthers and their trollish beliefs by taking them seriously.
This has countless parallels in political discourse. Trade flat-Earthers for (actual) neo-Nazis; it's probably not a "spectrum of opinions" you want to broadcast without passing any judgment. It's not the role of an encyclopedia to be a free speech platform.
I'm not really saying this to defend political articles on Wikipedia. It's just that I don't think you fix it by taking an LLM trained on Wikipedia, and telling it to regenerate the articles "without bias".
Wikipedia actually has multiple articles on flat Earth beliefs https://en.wikipedia.org/wiki/Flat_Earth Presenting a spectrum of opinions doesn't mean you have to avoid giving the impression of passing judgment. If an opinion conflicts with available evidence, you can just acknowledge that the opinion exists and give the evidence against it.
I hadn't heard about Grokpedia, but Conservapedia already exists, and seems to be what they want?
Edit: yes, it's mentioned in the article: "Conservapedia was launched in 2006 and is widely regarded as a joke by anyone who tries to wade through its ridiculous articles."
Edit: It seems that the supposedly liberal-based pedia remains far more popular then the explicitly conservative-biased pedia. But both are out there, so take your pick.
> As much as I dislike wikipedia
What specifically do you dislike about Wikipedia?
It is a great example of "objectivism" where on the surface it seems neutral and rational but actually is the subject of intense motivations that corrupt the content. This is particularly true for anything related to politics, history, nation states, celebrity tech bros, and so on. An average article on "the history of bicycles" is not a problem and not what I am referring to ofc.
For a different reason it is host to a wide range of superficial treatments of scientific and medical information. Which is not ideal. See scholarpedia for a better alternative for this kind of information (although its not well populated).
Do you have specific examples where you think the biases show?
For example, the page on Adult neurogenesis had an overreliance on early and limited evidence in 2010s with lots of editing wars occurring by academics. Then it got shifted to "controversial" which is better but the process for dealing with new scientific results is not ideal.
Thanks for responding with an example! I’m not very familiar with that topic.
What were the biases that affected its coverage on Wikipedia?
Academics with diverging views just started editing wikipedia to suit their views and the whole thing was very preliminary and not accurate for a long time. Then it was bailed into a "controversial" section. I intentionally picked a mild example lol.
Thanks, that’s helpful! What would your ideal outcome be in a situation like that where there is a scientific controversy?
We should have access to a Borges/Thomas Pynchon style dictionary of many possible paths to information. That kind of thinking should be encouraged. It can be confusing and controversial.
- I would avoid giving the impression that a simplified narrative is established (unless it won a nobel or is really beyond any reasonable doubt).
- End the use of "controversy" sections where it is used to minimise the impact of perfectly acceptable science. This is clearly used as a tactic in certain instances to minimise results.
- Allow for mixed expert, AI, conventional editing. There is nothing wrong or elitist with including edited articles by academic experts as cut outs of a particular topic. Almost like if scholarpedia and wikipedia were merged.
- Try to combat omission biases. There are some pretty wild ones on wikipedia.
- New results should be handled properly in terms of editing and language.
- Reproduce multiple versions of the same text to show how it would be different if certain hypothesis/experimental results were valid. This is easy to do if using LLMs.
You really think Elon Musk is going to make a non biased wikipedia??
Definitely not! It will be better on certain topics and worse on others. It is all down to whether you think Grok is more thorough with controversial topics, scientific and medical information, etc compared with the current human edit-a-thon on wikipedia.
A wikipedia AI could be the most balanced ofc with suitable changes and actually taking major criticisms on board. One is omission bias where very important information is just left out of articles. Another is lack of comparison of conflicting narratives (history, politics, science, etc).
https://nitter.net/grok/status/1965863232478077127
> Charlie Kirk takes the roast in stride with a laugh— he's faced tougher crowds. Yes, he survives this one easily
> It's a meme video with edited effects to look like a dramatic "shot"—not a real event. Charlie Kirk is fine; he handles roasts like a pro.
This thing as the basis for an encyclopedia will be worse than useless. Comparing it to Wikipedia is several layers removed from reality.
What's ur opinion on adult neurogenesis? Do you think the progression of that article suggests that wikipedia is the model to follow for scientific information?
What about important and controversial historical events? Do you think that wikipedia omitting information is acceptable?
Do you think that other models would insert "controversy" before making statements as part of a dark pattern to potentially encourage investigation fatigue?
All of the above are also possible in an Elonpedia style situation or Grok. I never said it will be absolutely BETTER.
> It will be better on certain topics and worse on others.
This sounds like a "both sides" kind of statement and I don't think it's fitting literally immediately after acknowledging that you don't think Musk is going to be unbiased.
What topics, specifically, will Grokpedia be better on? Which race has the lowest IQs? Whether trans people are mentally ill and should be committed? Whether a CEO should be able to run 5 different companies while on ketamine? Was Hitler really all that bad?
This is old hat. There are many Wikipedia alternatives which are based on the idea that Wikipedia is not conservative/hilarious/libertarian enough:
Let's take a spin through what the various Wikis have written about Musk:
https://rationalwiki.org/wiki/Elon_Musk
https://www.conservapedia.com/Elon_Musk
https://en.uncyclopedia.co/wiki/Elon_Musk
http://www.scholarpedia.org/ - no Musk article
https://citizendium.org/wiki/Elon_Musk - nothing here (yet)
... and so on.
Honorable Mention to Encyclopedia Dramatica which had some funny parody content.
they all were not part of hyped roadmap with XXB of investments and XXXB evaluation.
Throwing tons of money can often make things happen.
scholarpedia is the best because it has almost zero biography pages/knuckleheaded political slants. However, it seems to have just have stopped in its tracks c. 2008
The best thing wikimedia can do is make sure Wikipedia is unbiased as possible in sensitive topics - not just really on random online editors to notice things but to actually actively check for biases and omissions.
I'm not sure if moving from the biases of wikipedia editors to the biases of wikipedia staff would necessarily be a positive
It depends. Often staff are a lot more even handed than volunteer editors because you have to be a bit weird and love moderating to do it voluntarily, but it's pretty normal to do it as a job.
StackOverflow has this problem (or had, before it died) - the mods were hugely invested in closing questions for basically any reason, so normal users ended up hating it and the company couldn't make any changes to improve things because whenever they tried the mods revolted.
It's not as much of an issue with Wikipedia because most Wikipedia users aren't actually editing articles and running into any moderation issues.
Then take a really sensitive topic and check it:
https://en.wikipedia.org/wiki/Gaza_genocide
Ask yourself how neutral this is, and it of course isn't neutral at all by Wikipedia standards (obviously both parties don't agree, and one party's viewpoint is not represented ... at all, which is normally a hard no on wikipedia) and then look at the discussion page, to see it become 100x worse, and to see openly racist viewpoints that I'm surprised they allow even on discussion pages.
What's especially troublesome is that one side (the one who's viewpoint is the only represented on the page) is totally opposed to even discussing the existence of the other side.
If CISA is reaching out to Facebook and Twitter to censor content, you can bet there is heavy government involvement in Wikipedia as well. Haven't seen that mentioned in the comments yet. And hopefully that is not a controversial statement.
[flagged]
Wikipedia is very much in maintenance mode now. The vast majority of editors are interested in tweaking things like external links and categories. Some people (many!) are overly invested in the admin's noticeboard.
If you are a content creator, good luck. You aren't really valued.
You might not like what I write, but that is very much the situation on the ground.