User talk:DErenrich-WMF/Add A Fact Experiment

Please feel free to leave feedback here!

Thanks DErenrich-WMF (talk) 15:50, 9 August 2024 (UTC)[reply]

AI, really?

This seems like a terrible idea, AI spits out nothing but garbled false statements. Kirby54 — Preceding undated comment added 13:35, 27 September 2024 (UTC)[reply]

Hey Kirby54, see my reply to MRSC below noting that this AI implementation is just checking whether the claim is documented on the relevant Wikipedia pages. It is not attempting to propose a specific insertion of text into those articles. Thus, an error-prone LLM would not be capable of adding falsehoods to Wikipedia pages in this experiment BluePenguin18 🐧 ( 💬 ) 17:23, 27 September 2024 (UTC)[reply]

Thanks for chiming in here & elsewhere @BluePenguin18 on the question of how AI is being used here. You are correct that we are using AI in this tool only to search Wikipedia and retrieve back some contextual information to the user of the extension (i.e., whether or not there is a relevant article on this topic on Wikipedia – though the user can also just search manually; and whether or not the claim that the user has found/selected is already contained in the article). AI is not involved in writing any text that is posted to the talk page. I'll elaborate on this in the FAQ so hopefully it's less confusing to others! Maryana Pinchuk (WMF) (talk) 21:03, 27 September 2024 (UTC)[reply]

'Spam' concerns?

My initial gut reaction to this is that it seems great, but I also see how it could, especially on breaking stories, pollute talk pages with impetuously created topics from users acting in good faith but without much consideration. Mittzy (talk) 20:54, 26 September 2024 (UTC)[reply]

Oh, and I do hope that there are not plans to create a similar app that modifies articles directly. While automated edits are useful for housekeeping, I think that major (non-minor per WP:MINOR) changes to a page should be authored with care. Mittzy (talk) 20:57, 26 September 2024 (UTC)[reply]

Agreed on the above, @Mittzy! Tod Robbins (talk) 21:07, 26 September 2024 (UTC)[reply]

AFAIU, currently this requires an autoconfirmed wikipedia user, so the spam potential is reduced.

It could perhaps also be nice if the extension could check if talk page has a mention of this, then offering to add it as a reply to it (if the source is not the same). Aveaoz (talk) 21:47, 26 September 2024 (UTC)[reply]

I think we've seen from newcomer tasks that any structure that encourages specific editing behavior will produce a significant volume of non-constructive edits. Volume alone doesn't tell the whole story here, as some edits are less constructive/more destructive than others. With that in mind, I think both that and this have the potential to be a net positive for the site, while not unduly burdensome for those trying to maintain the existing quality of articles. Remsense ‥ 论 05:02, 27 September 2024 (UTC)[reply]

Thanks everyone for the feedback! This is all really helpful. Yeah polluting/spamming talk pages is definitely something we're thinking about and we're thinking that the talk page might not be the right long term place for this kind of information. If we decide to move forward with this concept we'd definitely want to check that duplicative facts are not being added (either just not letting you add the source or adding it as a reply). Currently the edit rate is low enough that we skipped that for the purposes of getting this in people's hands quickly. DErenrich-WMF (talk) 05:29, 27 September 2024 (UTC)[reply]

One other concern I hold is that Wikipedia can be a highly technical / in-depth resource... The potential for largely automated inclusion of short 'fact' or 'fun-fact' type segments in the middle of an article does not, I feel, fit too well with Wikipedia's goal. As I said last night, I think that such automated / 'AI powered' tools poses quite a risk to the integrity of Wikipedia if it is ever allowed to be used to directly edit articles. Mittzy (talk) 15:12, 27 September 2024 (UTC)[reply]

Thanks for the sharing your thoughts @Mittzy! Did you get a chance to try out the extension and play around with fact/claim selection? It's interesting that you mention "short 'fact' or 'fun-fact' type segments" because our intention was to facilitate adding important new information on topics that may not have yet been updated, not fun facts or trivia. There are many articles (especially on less-popular topics) that haven't been updated to reflect new information/sources because no Wikipedian is closely watching/working on them, but people might be encountering new information on these topics outside of Wikipedia.

But I can see how adding random fun facts could be something a user of this could do. Tangentially related to this experimental browser extension (which we're just demoing/getting feedback on, not planning to make available to non-editors), but I am curious if you think there is any place anywhere on Wikipedia (maybe not in the article directly) for non-Wikipedians being able to submit well-sourced "fun facts/trivia"? I.e. something like a reader version of DYK? Maryana Pinchuk (WMF) (talk) 20:49, 27 September 2024 (UTC)[reply]

Does not function on Opera GX

I immediately downloaded this extension onto my Opera GX browser after being intrigued by the demo video and tested it out. I searched for an external website I'd been looking at for a while, highlighted a sentence and clicked the extension icon, but nothing happened. I tried toggling all sorts of settings to no avail. Am I doing something wrong or is the extension just not compatible with Opera GX? SleepDeprivedGinger (talk) 21:17, 26 September 2024 (UTC)[reply]

Hey, thanks for the feedback. We did not test this on Opera GX and I do not have a lot of experience with that browser so it's not surprising that it doesn't work. That said I don't think we did anything that would make it explicitly not work. DErenrich-WMF (talk) 05:32, 27 September 2024 (UTC)[reply]

That's alright, do you at least plan on adding support for this browser in future? SleepDeprivedGinger (talk) 11:59, 27 September 2024 (UTC)[reply]

Safari support

I use Safari and would try this extension out if it had a Safari version. I understand that this would require quite a lot of extra work, so it might make sense to hold off on Safari support until and unless Add A Fact exits the experiment stage. Once that happens, reaching as many editors as possible should be a priority, and I hope that the existing Wikipedia app and Apple developer account can streamline the process of getting the extension onto Safari. Ilovemesomenachos (talk) 02:44, 27 September 2024 (UTC)[reply]

Yeah we targeted Firefox and Chrome first because of their wide use in the community. Wider support would be on the table if this exits the experimental stage. DErenrich-WMF (talk) 05:34, 27 September 2024 (UTC)[reply]

Oppose on principle

Just logging that I opposed any use of "AI" on Wikipedia on principle, for reasons that should be obvious. MRSC (talk) 04:03, 27 September 2024 (UTC)[reply]

I remain similarly opposed to AI LLMs directly adding content to Wikipedia, especially because of model collapse as AI bots could begin training themselves on potentially incorrect portions of Wikipedia that other AI bots wrote. However, I think you should give this proposal a second look, since this AI implementation only checks whether a statement is already represented on the articles that seem most closely related to the claim. If the statement does not seem to be represented, then human users receive the option to use a non-AI script for proposing the fact to talk pages. BluePenguin18 🐧 ( 💬 ) 07:16, 27 September 2024 (UTC)[reply]

I'm also inclined in principle against AI in Wikipedia. But I'll try to keep an open mind, and will at least give the extension a try. Mike Marchmont (talk) 12:08, 27 September 2024 (UTC)[reply]

Hey, thanks for the feedback even when it's negative. I would actually appreciate more detail on your reasons for opposing applications of AI in Wikipedia. AI is actually already in use in various ways e.g. vandalism detection and is planned for more applications (see [1] or [2]). It'd be useful to know why the community opposes it (e.g. is it energy use, copyright. reliability, etc). DErenrich-WMF (talk) 15:46, 27 September 2024 (UTC)[reply]

What LLM is being used?

Curious about potential bias. (Haven't been able to try it yet.) RememberOrwell (talk) 06:30, 27 September 2024 (UTC)[reply]

The extension description states that it uses Llama3-70b Bertaz (talk) 07:44, 27 September 2024 (UTC)[reply]

Keep suggestions on the talk page

Articles like Alcaligenes faecalis are not written as a loose collection of every time their subject is mentioned, instead requiring humans to weave reliable sources to comprehensively describe their topic. Using LLMs to assess whether certain statements are represented on the associated articles is an innovative approach, but I caution against posting these suggestions anywhere besides the talk page. Whereas the talk page can be easily archived to resolve spam during breaking news (though I like Aveaoz's idea to have the LLM check against existing talk page proposals), it would be difficult to cleanup if we allow these proposed facts to be posted into the article itself, such as through invisible comments. I think cautioning against unreliable and deprecated sources is sufficient because there are productive ways to use every source. BluePenguin18 🐧 ( 💬 ) 07:08, 27 September 2024 (UTC)[reply]

Thanks, @BluePenguin18! One piece of feedback we've heard is that talk pages aren't the ideal place for article suggestions, because many of them are dormant/unwatched. I agree that letting people add these suggestions directly to the article is risky, but I'm wondering if there could be a more actionable "holding space" than a talk page that doesn't become yet another moderation queue/backlog. What do you think about sending these to the relevant WikiProject's talk page? (Though the same concern as article talk holds, in the sense that many WikiProjects are also pretty dormant, I wonder if this might make it more likely for someone to look at/do something with a suggestion?) What do you think? Do you have other ideas like that, to put these in front of Wikipedians who'd be inclined and excited to do something with good suggestions? Maryana Pinchuk (WMF) (talk) 20:55, 27 September 2024 (UTC)[reply]

Please enable this for draft space

I tested this against three articles in draft space - Draft:Marshall L. Stephenson, Draft:Murder in New Jersey law, and Draft:William J. Harbison, each using sources and text selections that were clearly on point, and in each case, the tool 1) failed to detect that a relevant draft exists (in the first and third instance, it found nothing; in the second, it found various articles secondarily related to murder in New Jersey); and 2) when prompted with the title of a draft article, found the draft but then spun its wheels endlessly without getting to the step of offering to add a note to the draft talk page. This is frankly a perfect tool for the way I write drafts, it just needs to be able to find them. BD2412 T 12:17, 27 September 2024 (UTC)[reply]

this is really useful feedback. thank you. currently the way it is coded it isn't easy to let it auto-find drafts but adding support for manually tagging drafts it something we could consider doing in the short term. but improving the draft situation overall is a good idea. DErenrich-WMF (talk) 15:42, 27 September 2024 (UTC)[reply]

Well... where to begin?

I agree completely with BluePenguin18 above about the genesis and development of articles like Alcaligenes faecalis: it's a really good and concise summary of how WP should and can work. But thousands of less experienced editors really don't get it. As a professor of Mediaeval Literature pointed out to me last year, Wikipedia can be seen as the interface between academia (often dealing with highly technical subjects) and the general public: and such snippets of 'breaking news' detract from the long-term validity of similar articles.
1. BTW, Alcaligenes f. has already been edited on this very subject (apparently deliberately ignoring the example in the video) by someone whose edit history and talk page do not inspire confidence, possibly just to get there first. I'm not qualified to judge whether it's a useful edit.
This "Add a fact" tool works in exactly the opposite manner to the way I (and I suspect many others) actually edit on WP. If I do happen to come across something useful relating to an article I am familiar with, I will open it in the editor, add the relevant info in the article text and make a full cite there and then. Many of the articles I edit depend on scholarly articles, usually via pdfs, which the tool is not capable of addressing.
The very idea of "Facts" is barely consistent with the fundamental concept of Wikipedia's approach to articles. It could easily be said that there are no facts, only opinions, and WP attempts to balance them using the best of reliable sources.
Who is actually going to be doing the work once this factoid (cue breathless "Did you know...") has been posted on the talk page? The result, as shown in the video, is incapable of being transferred into the article without further effort from an experienced editor: if it is not simply ignored, someone will have to use their time and effort to give reasons why it's not suitable for inclusion. Only few of these posts could easily lead to exasperation.
Use of this tool could easily encourage lazy spamming of talk pages by people eager to increase their edit count.
I find that the use of |quote= makes it relatively hard to distinguish the quote from the cite.
I suggest you re-record the video, speaking at approximately half the speed. I know you are very familiar with the tool, having developed it yourself, but as an intelligible introduction to a noob I'm afraid it fails. Your speech comes across as almost garbled. Please take your time.
The foisting of AI on Wikipedia and Wikimedia has not yet found wide approval. Although there may be positive reasons for using the hallucinatory, imitative mumblings of a village idiot, I for one remain utterly unconvinced. MinorProphet (talk) 14:01, 27 September 2024 (UTC)[reply]
Agree with all statements regarding Wikipedia. Mittzy (talk) 15:08, 27 September 2024 (UTC)[reply]

The "Add a Fact" workflow of proposing statements is obviously very different from how experienced editors approach articles, but I think DErenrich-WMF is reasonably claiming that the act of proposing these facts would serve as a gateway to mainspace editing for new editors. I do not think that this tool would be abused for edit count purposes, simply because there are much easier and more fulfilling ways to boost one's count as a gnome. Regarding AI hallucinations, note that this implementation only uses the LLM to check whether the claim is represented on associated articles. It does not propose a particular text insertion to the talk page. BluePenguin18 🐧 ( 💬 ) 18:22, 27 September 2024 (UTC)[reply]

Does it, or does it not, work with Firefox?

The description of the tool states, "Add A Fact is available for the Google Chrome and Mozilla Firefox browsers." And there is a comment (above) which says, "Yeah we targeted Firefox and Chrome first because of their wide use in the community".

So in installed it in Firefox. It seemed to install OK, but I couldn't get it to work. So I delved into the docs, and I found this statement in the FAQ: "For technical reasons the extension will not just work on Firefox due to some manifest 3 API's that Firefox doesn't support".

If it is not supported in Firefox, I would understand that. I am aware of the issues of browser compatibility. And I congratulate you on what you have achieved so far. But can you perhaps amend the information to clarify this point. (And if it is not supported by Firefox, perhaps it should be removed from the FF add-ons page? Mike Marchmont (talk) 16:35, 27 September 2024 (UTC)[reply]

I'm running Firefox 130.0.0 and it's working for me. Which version do you have installed? Tod Robbins (talk) 16:45, 27 September 2024 (UTC)[reply]

Please explain the required permissions

So I said to myself, why not give it a try, and followed the link to the Firefox Add-Ons store. Lo and behold, the extension wants to be able to read all the data on all websites. Which, of course, includes anything entered on a form on any website, including usernames, passwords, bank card numbers, CVV2 numbers, and so on. Of course, I politely declined the installation of the extension. The required permissions and why they are needed would be most welcome, as would be an official statement from the Wikimedia Foundation regarding the collection and use of data. Imerologul Valah (talk) 21:17, 27 September 2024 (UTC)[reply]