Wikipedia talk:Wikipedia Signpost/2025-11-10/Opinion

Sorry to say this, but substantial parts of this analysis are, frankly, bogus. Sure, the author is very knowledgeable about the Wikipedia side of things, and the article offers many thoughtful observations in that regard. I also appreciate the essay's historical perspective, reminding us that for an encyclopedia, Wikipedia's lack of central control by a commercial owner is the exception rather than the rule.

• Good read. Minor correction, "BritishPetroloeumpedia" should be spelled "BritishPetroleumpedia". ―Howard • 🌽³³ 13:46, 10 November 2025 (UTC)
  • Fixed, thanks. Half-expected an ENGVAR correction coming (BritishPetroleumpaedia). :) — Rhododendrites ^talk \\ 14:42, 10 November 2025 (UTC)
• "Biographical Entries // High-priority sources: Primary self-published facts (e.g., verified X posts..." Letting people write their own realities via X posts. Incredible! What could go wrong! PhotogenicScientist (talk) 14:31, 10 November 2025 (UTC)

However, this article goes awry in its "analysis" of the sourcing used by Grokipedia, which betrays the author's ignorance of how such AI systems actually work: [...] I asked [Grok] to explain the way it prioritizes sources for different kinds of content, and it provided a table that's worth including here.

Such an approach is a serious mistake, as e.g. explained back in 2023 already (when ChatGPT et al. were still new and such errors were maybe more excusable) by Simon Willison ("Don’t trust AI to talk accurately about itself"):

Asking questions like this feels like a natural thing to do: these bots use “I” pronouns (I really wish they wouldn’t) and will very happily answer questions about themselves—what they can do, how they work, even their own opinions [...]. These questions are likely to produce realistic but misleading answers. [...]

By definition, the model’s training data must have existed before the model itself was trained. Most models have a documented cut-off date on their training data [...].

If it was trained on content written prior to its creation, it clearly can’t understand details about its own specific “self”.

(While Rhododendrites unfortunately didn't document the exact version of Grok he asked, and with which prompt, XAI's documentation currently states: The knowledge cut-off date of Grok 3 and Grok 4 are November, 2024, i.e. about 10 months before Musk even decided to build Grokipedia. Now, newer chatbots use RAG-like methods to access information from after their cut-off date and get over other limitations of their parametric knowledge (i.e. the information stored in their weights during training); these involve retrieving external sources - in particular web pages and, in case of Grok, tweets. However, this will be public information, i.e. also available elsewhere, in more reliable form. Also, Grok and other chatbots like ChatGPT are generally indicating these retrieved sources explicitly in their answers, and if Grok's answer as reproduced in this essay came with such citations, the author failed to mention them.)

To illustrate this problem, I just tried to replicate Rhododendrites' query as closely as possible (with the default "Fast" version of Grok that one gets as a non-paying X user at https://x.com/i/grok ). As expected, Grok's answer differs a lot from that reproduced in the essay. For starters, it indeed listed various external sources used in its answer (including https://en.wikipedia.org/wiki/Grokipedia and https://en.wikipedia.org/wiki/Grok_(chatbot) ;-). And while it again produced a plausible-looking table, its content was very different - e.g. for me, it hallucinated the existence of a numerical source weighting system (with contradictions like "arXiv/peer-reviewed journals (e.g., Nature, Phys.org) (1.5–2.0×)" vs. "Unverified preprints (0.7×)").

The author seems to have had some slight awareness how dubious the research method underlying his far-reaching conclusions is, so he preceded this by a feeble disclaimer: It's not clear how reliably Grok will explain its own internal processes, but it should at least communicate the way its developers want Grokipedia to be seen. But besides the fact that, as explained above, it is actually very clear that this explanation will not be reliable at all (even compared how reliable Grok's answers may be on average), this also betrays other misconceptions:

• Rhododendrites' assumption that a LLM chatbot answers' about a particular topic will "communicate the way its developers want" that topic to be seen shows that he is not at all familiar with basic facts about LLMs, in particular the difficulty of aligning them with such preferences of their makers. (This is a well-known problem which has attracted a lot of research, but is nowhere near fully solved. Chatbot makers try to address with techniques like RLHF that require a lot of effort but can still fail. In fact, Rhododendrites could have learned about a prominent such failure involving Grok itself simply by reading our article Grok (chatbot).)
• What's more, as folks who have read research publications about previous efforts to use LLMs for generating Wikipedia-like articles would know, such systems don't simply leave source selection and weighing to the underlying LLM(s) or chatbots used. Rather, they tend to devote significant effort to try to steer this with e.g. prompts (which, as is well known, can make a great difference for the output of LLMs). See for example Stanford's STORM system for writing Wikipedia-like articles (which, as mentioned in our review at the time, even filtered sources according to Wikipedia's own RSP list - i.e. was very far from simply leaving this to the underlying LLMs from OpenAI that were used). In fact, it's very likely that XAI's engineers had to devote extensive efforts to this too, given that after they had already worked on Grokipedia for weeks, Musk had to delay the launch by another week because "We need to do more work to purge out the propaganda".

I appreciate that this research must have been done very quickly, judging from the fact that it was first pitched to the Signpost on October 29, i.e. a mere two days after Grokipedia's launch. However, one would have still expected better from a Senior Research Fellow at a major US research university. One is also curious about the publication process of Tech Policy Press - it's not quite clear from https://www.techpolicy.press/contributor-guidelines/ , but it seems they publish submissions without any fact-checking or review by subject matter experts. Fortunately other academic researchers appear to be working on more systematic, less hallucination-infested comparisons of Grokipedia and Wikipedia right now, so we should be able to highlight better analyses in the Signpost soon.

Regards, HaeB (talk) 08:20, 11 November 2025 (UTC)

Whoa, HaeB. First, you're treating a clearly marked piece of commentary like it's intended as scholarship. Second, your objections and all of the text above is apparently only to... asking Grok? None of this, in my view, remotely justifies the surprisingly harsh final paragraph.

Asking Grok begins with a clear qualification and functions primarily to highlight what anyone can plainly see for themselves: that Grokipedia relies heavily on primary sources. Call asking Grok a gimmick or a stunt if you like -- that's largely how I see it, too. These kinds of "using the master's tools against him" exercises can provide some fun and some insight, but of course I'm not going to put "I asked Grok and here's what it said" into a research study (although you would be correct to assume I'm not an AI/LLM researcher).

If I were reading this with a "critical about the credulous way people talk about LLMs" hat on, I might also get hung up on communicate the way its developers want. However, it's there not as a technical statement but based on the entirely non-technical observation that Musk/xAI are frequently on the record about trying to intervene to change how Grok answers questions about things, including about themselves/itself. How that happens technically, how hard it is, how well it works, how foolproof it is, etc. would be important in many contexts, but mostly beside the point here (although the response to your query does seem to emphasize primary sources in several places, too, despite being presented as counter-evidence). It would probably be better worded ~"might, based on the number of public statements Musk and xAI have made about their interventions, communicate the way its developers want Grokipedia to be seen". I think it was something like that in an earlier, longer version. Still, though, we're not talking about any of the central arguments.

this research Again, it's not research, it's commentary.

How about this: can we pretend instead of my "I asked Grok" bit, I just said something like "from looking through a few dozen Grokipedia articles, one thing that jumps out its reliance on primary sources (much more so than Wikipedia)". The rest of the essay can just remain as-is, and we can proceed as normal with the substance of this opinion piece rather than get into the technical weeds about the limitations of LLMs. — Rhododendrites ^talk \\ 14:47, 11 November 2025 (UTC)

• As an additional data point on the above discussion, I don't think the "references" Grokipedia offers are good for very much at all, not necessarily even insight into how Grokipedia created the article. I happen to have written an article with a somewhat unusual title, and Grokipedia of course has its own copy, which is extremely similar and clearly just taking "my" article... but... it's thrown away all the boring ol' book references Wikipedia has and replaced them with random weblink "references". And when I say "random" I mean it - there are links that have nothing to do with anything in there, and don't include any content that Grokipedia has even taken. Basically I wouldn't assume very much from the Grokipedia-created refs, either. I think it's some script that was told "find relevant webpages for this passage and insert them as references" that spat out ref-guesses after Grokipedia had already written the article. A secondary post-hoc justification, in other words, not insight into the creation process. SnowFire (talk) 17:30, 11 November 2025 (UTC)

  Do you find that Grokipedia's citations are much different with regard to reliability than those generated by other LLMs? All of the models, AFAIK, are pretty infamously unreliable (e.g.). But looking at the content of the articles, it's clear many of them integrate a lot of material from primary sources and sources connected to the subject. Musk said they took "Grok and said ok cycle through the million most popular articles in Wikipedia and add, modify, and delete. So that means research the rest of the internet -- what is publicly available -- and correct the Wikipedia articles and fix mistakes but also add a lot more context...". So the "what is publicly available" is the probably the big overarching sourcing difference, but the primary/connected/self-published portion of "what is publicly available", which is so visible in the articles, is to me the salient difference with Wikipedia. — Rhododendrites ^talk \\ 18:28, 11 November 2025 (UTC)

• Nearly every encyclopedia asserts some version of "neutrality." Wikipedia's definition is unusual: its "neutral point of view" policy aims not to pursue some Platonic ideal of balance or objectivity, but rather a faithful and proportional summary of what the best available sources say about a subject. As a Classical scholar and a Wikimedia Laureate, the arguments about both topics feel wrong: The first pillar clearly says that Wikipedia is an encyclopedia which per definitionem is a description of the world as it truly is. With Plato we can conclude that we can only strive for this truth (or ideas, perceived reality, or whatever you might call it), but we can agree that there is just one of these. Certainly, we can and should have different opinions about how to get there, what is the right/good/neutral way, etc. The second pillar, Neutral Point of Views, instead, only describes the method how we think we reach this most easily, by looking for sources, treating them as unbiased as possible, and weighting their points against reality. But that doesn't set us free from thriving for truth.
  Of course, our encyclopedia would give first the information with the highest probability to be closest to the reality and then other opinions, regardless of the percentage of people who believe one or the other. If a majority would still think that Earth is a flat and only a minority that it's a geoid, we would still give the latter information first. And there are countless of community members who fight hard for keeping the perceived reality in our articles against broadly popular but still wrong perceptions of the world. — Whether you call one or the other perspective left or right, doesn't matter at all. And if one side thinks that the other one is overly represented, maybe they just don't want to accept that the truth might be more on one side than the other. What bother me most is my perception that they try to confuse us with the unscientific and untrue approach to ask for multiple realities and shamefully equate this with multiple opinions/pluralism of opinions just to call for neutrality. I don't want a neutral encyclopedia when the truth is not neutral. It would be like asking myself to lose my values just to be neutral about things. That makes no sense. This encyclopedia has values, this encyclopedia seeks for the truth, this encyclopedia tries to be respectful/neutral to all the sources out there, but not for the sake of an abused definition of neutrality/NPOV. Best, —DerHexer (Talk) 18:58, 11 November 2025 (UTC) PS: @Jimbo Wales: Maybe some food for your book tour about “truth”. Sadly, I cannot attend the reading in Berlin because I'm traveling in the Czech Republic.

• Thanks, DerHexer. I'm not sure what you're disagreeing with, though? If a majority would still think that Earth is a flat - but you quoted a passage about summarizing what the best available sources say, not about what's popular. If this is a hypothetical -- i.e. "if the consensus among scientific sources is that the earth is flat we would still say it's round" -- then I would either disagree or reject the impossible hypothetical, as we then exist in an alternate universe where the earth is flat or one of our chief mechanisms to describe the world as it is (i.e. science) has failed.
  It's not that truth doesn't matter (although the essay "Verifiability, Not Truth", which I presume you disagree with :), is to me at least a useful polemic), but that we defer to external sources to figure out what's true, evaluating their reputation for truth/accuracy (via other sources), and then summarize them faithfully. In either case, it's not Wikipedians debating what's true in a vacuum but rather Wikipedians debating what's true according to other sources, then debating how to select sources and summarize them.
  which per definitionem is a description of the world as it truly is - At the risk of going on a tangent (a risk I'm willing to take since it's relevant to a piece of this essay), it's hard to define an encyclopedia, but while it is true that encyclopedias often aim for representing truth (or at least express a desire to do so), they manage to nonetheless incorporate a wide range of unintentional and intentional biases and inaccuracies, are have historically served a range of powerful interests. Ruwiki, Conservapedia, Metapedia, and dozens of others have, in recent history, launched in pursuit of their own version of "the world as it truly is", directly countering Wikipedia. I would agree that we should not treat these truths as equal, and tried to explicitly play against the "weight according to popularity" critique of Wikipedia in the essay. What I find most useful and interesting to talk about, however, is not their truth but how they have modified our process of pursuing neutrality in order to accommodate their version of truth. I suspect we probably agree about the process on a practical basis, even if we're using a different epistemic frame to talk about it. — Rhododendrites ^talk \\ 19:22, 11 November 2025 (UTC)

Indeed, (moral) neutrality as a goal is nonsense. Wikipedia’s goal is to describe the world as it is. When the world is not (morally) neutral, so be it. But certainly, our methods to describe the world should be as unbiased as possible. But form follows function, neutrality follows reality. — And that's where I disagree with, for example, Sanger’s theses, a lot of “you are so unneutral” weiners on the politcal edges, as well as with the over-importance of unclearly defined neutrality in this essay and elsewhere. The fight about (moral) neutrality is misleading and counter-productive to what we want to achieve.
And just a few words about verifiability: that is epistemological nonsense. We can prove things wrong (and most of the stuff from ruwiki, conservapedia, metapedia, etc. we easily can), but we can only confirm that somebody has said something about something in a logical or convincing way, unbiased, without neglecting other sources and scientific data, etc. But if it's actually true (lat. verus -> verifiablity) or not, we can never guarantee. And if we call that “verifiability” or “seeking the truth”, I don't care. Funnily, there is no (good) word for “verifiability” in my native German language. — Having said that, I would describe the task of an encyclopedist slightly different than you: After their research on a subject, Wikipedians do not debate what's true, but they describe what is the most likely truth, refering to the convincing arguments of the sources, covering other opinions and their arguments too, in the fairest and most factual possible way; but of course they can call nonsense “nonsense”. Because they intend to use factual neutrality, not moral neutrality, when they describe the most likely truth. (And coming back to Plato, we must be cautious not to mix our frameworks with his. Morality was a completely different thing in his time. Therefore neutrality, truth, etc., too.) Best, —DerHexer (Talk) 22:59, 11 November 2025 (UTC)

What matters regarding items in an article is:

• Is it intelligible? Can the target readers understand it?
• Is it relevant? Does it answer questions they might have?
• Is it true? Does it help them to relate to reality rather than misleading them?

These are the only things we should care about. JRSpriggs (talk) 19:18, 11 November 2025 (UTC)