The Wikimedia Foundation (WMF) section of the village pump is a community-managed page. Editors or Wikimedia Foundation staff may post and discuss information, proposals, feedback requests, or other matters of significance to both the community and the Foundation. It is intended to aid communication, understanding, and coordination between the community and the foundation, though Wikimedia Foundation currently does not consider this page to be a communication venue.
Discussions of proposals which do not require significant foundation attention or involvement belong at Village pump (proposals)
This page is not a place to appeal decisions about article content, which the WMF does not control (except in very rare cases); see Dispute resolution for that.
Issues that do not require project-wide attention should often be handled through Wikipedia:Contact us instead of here.
This board is not the place to report emergencies; go to Wikipedia:Emergency for that.
Threads may be automatically archived after 14 days of inactivity.
Behaviour on this page: This page is for engaging with and discussing the Wikimedia Foundation. Editors commenting here are required to act with appropriate decorum. While grievances, complaints, or criticism of the foundation are frequently posted here, you are expected to present them without being rude or hostile. Comments that are uncivil may be removed without warning. Personal attacks against other users, including employees of the Wikimedia Foundation, will be met with sanctions.
The Wikimedia Hackathon, the yearly software engineering conference, may stop being a global conference
I can't say I'm very happy about this proposal: mw:Wikimedia Hackathon/Community Proposal 2026. This is a conference for Wikimedia software developers that I attend yearly, and that I consider to have an extremely successful format. The above proposal appears to want to take the money spent on this conference and basically kill it off to spend it on smaller regional conferences, except for every third year.
I'm not sure this is the correct place to post this or if it will resonate with people here. But this is something in the pipeline that is frustrating me. I wish WMF would leave stuff that's working alone and not make big changes to it.
I really hope this doesn't become a wishlist situation, where we have to put up with a poor format for a couple years, then finally there is a big backlash to get it changed back. I'd rather they just not make the questionable changes in the first place.
Thanks for listening. –Novem Linguae (talk) 06:09, 18 October 2026 (UTC)
"Our goal is to better understand whether concentrating our resources on one global event held in the same region each year creates barriers to participation or limits our ability to develop technical contributor communities in other parts of the world." There seem much easier ways to trial this than an immediate three-year experiment. For example, there are already regional conferences, trial a hackathon at one or two of those. Much less money spent for the same experiment without immediately affecting the current process. CMD (talk) 10:36, 18 October 2026 (UTC)
@Novem Linguae - Thank you for the feedback and I appreciate you making sure that the entire community has an opportunity to comment. To be clear, the goal isn't to mess with something that is working, but to provide an opportunity to discuss something that others in the community have raised questions about.
So far, there's been lots of really valuable feedback on what works and doesn't with hackathons, as well as alternatives to the one provided in the proposal (like @CMD's proposal above). To ensure that others have an opportunity to see your comments and engage with them, I'd encourage you to comment on the proposal's Talk Page. ~2026-57722-47 (talk) 19:04, 29 October 2026 (UTC)
My apologies! Just realized I wasn't logged in when I replied, this is from me (@BPiovesan-WMF) BPiovesan-WMF (talk) 19:05, 29 October 2026 (UTC)
I've said for a while that Wikimania should be around 1/3 as expensive to run as it currently is, and that we should be learning how to optimize for self-funding events that recover their costs (including the cost of scholarships) through partnerships, sponsorships, and tickets. I think something similar holds for hackathons. A version of this proposal that I would appreciate would look something like:
* A commitment to reduce the cost per hackathon-year while increasing accessibility (in geographic diversity, in # of participants)
* An approach that keeps a central global gathering, and adds a couple new regional hackathons (either entirely new ones, or existing ones that are expanded to have more of the WMH spirit), while trying to reduce total costs in the coming year
* An invitation for proposals and suggestions from people who have run extremely efficient / participatory / generative / well-sponsored hackathons, about how to apply the best parts of that organization to the global event(s). (I will point to decetralized 24-hour Wikidata birthday celebrations as one joyful alternative;)
Our current events schedule and cost per event (and especially net cost per participant! and low viewership of the recordings) remind me of other global communities that have gone through unwanted dry spells where they stop having global events at all after a big-budget heyday. I would prefer to see us get to a sustainable version of this -- specifically one where it is obvious that the global community can continue to hold a productive event even without a large centrally funded push by WMF or WMDE -- sooner rather than later. –SJ+ 18:58, 9 November 2026 (UTC)
New features for quickly finding bad-faith content
We (the Wikimedia Foundation Product Safety and Integrity team) have been working with volunteers for the last several months on new ways to automatically detect different kinds of bad-faith edits, which are then shown to volunteers to take action on, through a new special page called AbuseReview.
We have focused on two kinds: personal information that likely needs to be suppressed, and vandalism that likely needs to be reverted.
We figured the community here would appreciate some insight into the data we’re seeing and how these features work, including how it interacts with ClueBot NG and existing patrolling efforts. We have a project page that we’ve been updating along the way.
Detecting personal information
For this feature, oversighters are shown likely disclosures of non-public personal information, which typically should be suppressed or revision-deleted according to the oversight policy.
As volunteers described here a couple months ago, it took us working together through several iterations and a lot of manual labeling to develop an accurate classifier. Fancy AI stuff can only get you so far - we would not have been able to make something good enough without hard human work and expertise.
Since deploying this feature, the amount of personal information that is getting caught and removed on English Wikipedia by volunteers has roughly doubled. Oversighters have told us the accuracy rate is consistently better than the human reports they are used to receiving.
Personal data flagged on the Special:AbuseReview page and suppressed
This is built using an AI-powered classifier that runs on top of a language model. After evaluating multiple models, we are using CoPE-B, an open-weight model that is trained specifically for trust and safety purposes, and is maintained by the team at zentropi.ai. We are self-hosting this model, so the live edit analysis is all happening inside our infrastructure.
To use the classifier, we provide the model a plain language “prompt” that describes the specific factors to include and exclude. This prompt was written to match community oversight policies, and has been refined over time based on feedback from volunteers and observed performance.
If you are interested in how CoPE-B works, the creators have published a paper and talk about it on their blog. As a rapid classifier that emits yes/no decisions with confidence scores, CoPE-B bears a strong resemblance to what have more recently been coined “decision models”, like Jev and Clef. In fact, Zentropi recently published a new variant of CoPE-B that is more directly compatible with the “decision model” framework, and which we expect to evaluate soon.
Detecting vandalism
For this feature, volunteers are shown likely vandalism that should typically be reverted. This uses the same model (CoPE-B) as personal information, with a vandalism-focused prompt.
This has been running on English Wikipedia for about a month now, but as a limited alpha test accessible only to a group of volunteers involved in patrolling who raised their hand to help us test the feature. Similar to our development of the classifier for personal information, many of these volunteers also put in the time to manually review many edits to help give us solid data and feedback to work with.
We know there’s already a lot of energy on English Wikipedia when it comes to tackling vandalism, in human and bot form. We consulted with patrollers and with the ClueBot NG team, and we’ve been clear that we don’t intend to displace existing efforts. To avoid creating disruption as we introduce this tool, we are currently defaulting to a delay of a few minutes before surfacing likely vandalism, to let ClueBot NG and others “go first”.
By doing this, we’ve seen that there is a steady stream of vandalism that is slipping through the existing net of bots and humans, and that this classifier can help catch it. We’ve seen less than 1% of flagged edits reported as false positives, and we see a measurable speed-up in how quickly vandalism is getting reverted on English Wikipedia – especially in the longer tail that otherwise might have stayed online for many hours or days.
Time to mitigate vandalism flagged on the Special:AbuseReview page
What’s next
Given the above, we’re now ready to move the vandalism feature out of testing and open it more widely. To start, we have opened Special:AbuseReview to all rollbackers and admins on English Wikipedia. If this community thinks that the feature should be gated on different rights, it isn’t meant to be set in stone, but we do think it is important that the judgments being made within Abuse Review represent experienced patrollers.
Overall, it really seems to us that this new generation of classifiers is a huge quality improvement over what we and volunteers are used to in this space. We are internally working next on a classifier focused on link spam, and early results are already promising.
Finally – while we started with English Wikipedia, we have since worked with oversighters and others on other large Wikipedias to deploy this feature there too, and it seems to naturally translate pretty well. We are open to bringing this to other wikis that are interested.
We are interested in your thoughts – and of course, we deeply appreciate all of the volunteer collaboration on this project so far. EMill-WMF (talk) 16:30, 27 October 2026 (UTC)
I have used the AbuseReview tool for private information extensively, and it works great for quickly identifying likely PII exposure for quick suppression. Often, I'm reverting the edit before anyone is likely to have seen it, so we're likely suppressing before any exposure, as opposed to the old model either finding it while patrolling or someone seeing it and reporting. I've also used the vandalism detection mode, which has also been effective at surfacing edits that may have slipped past recent changes patrollers and Cluebot. Count this as a strong endorsement of both of these tools from me. ScottishFinnishRadish (talk) 17:54, 27 October 2026 (UTC)
I also want to say that I hope that we'll see a way for editors to report issues for revdel and oversight directly into the AbuseReview tool so there is a private on-wiki way to make these reports, which would be a huge upgrade over the current status quo. ScottishFinnishRadish (talk) 11:35, 28 October 2026 (UTC)
+1 to a private on-wiki "button" to request revdel and oversight Netstars22(talk to me!) 18:32, 30 October 2026 (UTC)
This seems like a good idea. I've gone ahead and created phab:T443112 to move the idea along to the next step. –Novem Linguae (talk) 10:22, 31 October 2026 (UTC)
Also a very strong endorse from me. I was honestly shocked to see how much blatantly obvious vandalism gets past recent changes patrol, since my newbie experience of attempting to help out with that was that all the vandalism I found was reverted faster than I could get to it. As for the suppressible information... well, I knew it was that bad already, and I think we're surely still missing out on quite a lot. But this has been a huge help in surfacing things that should have been suppressed, but were not reported. In solidarity, asilvering (talk) 19:58, 27 October 2026 (UTC)
I don't have much to say except that I love this. Keep up the good work.
Is there any chance that we will be able to run this on historical edits/pages, so that we can chip down on long term vandalism? EatingCarBatteries(contribs|talk) 20:08, 27 October 2026 (UTC)
I had a short conversation about this with @EMill-WMF a week or so back after becoming aware of potentially libellous material that had been present in a BLP article for several months (since before AbuseReview went live). Based on that conversation, my understanding is that, at least for personal information, something like this is seen as desirable but the practical implementation has not yet been determined. We agreed that simply scanning the whole database and dumping any hits into AbuseReview would overwhelm the system, and this would be even more the case with vandalism (instances of which are more numerous). Several ideas have been suggested but the priority has been to get the live version production-ready first. Thryduulf (talk) 21:58, 27 October 2026 (UTC)
It's great to hear that's being considered. I hope that this does come to fruition, as well as for different purposes like promotional text.
If the system prompt being used is public, I'd imagine someone can repurpose it to run it on their machine on an enwiki dump. But it would take a metric ton of compute to scan on one of those all-revision dumps. EatingCarBatteries(contribs|talk) 07:04, 28 October 2026 (UTC)
I share my colleagues' enthusiasm for this. Finding PII in real time has been a consistent challenge, and the tool for that has made a measurable improvement. Our core purpose here is building an encyclopedia written by humans. Handling vandalism and private information is a necessary task to that enterprise, but it's an unpleasant janitorial task that sucks up editor time. I deeply appreciate the fact that PSI developed a means to automate some of this effort, and also that they worked with volunteers to bring in a level of quality control we are happy with. Vanamonde93 (talk) 20:13, 27 October 2026 (UTC)
It's been a few years since I've regularly done much anti-vandalism work, but from a brief look I really like this tool.
One comment: can we get a button to update the list, rather than having to refresh the page? (Or am I missing something?) LittlePuppers (talk) 21:00, 27 October 2026 (UTC)
You're not missing something, the product is.:) We are in fact working to deploy a "live refresh" mode soon, hopefully next week. EMill-WMF (talk) 21:07, 27 October 2026 (UTC)
I've been using the personal information side of AbuseReview and I can't endorse it highly enough. It's catching the additional of personal information before even new page patrollers can, so we (Oversighters) can suppress it before even they see it, reducing the exposure to sometimes very sensitive information. Thryduulf (talk) 22:01, 27 October 2026 (UTC)
Out of curiosity, what kind of personal information are people trying to add? Is this a matter of people trying to dox another person, or something a newer editor may think is innocent? EatingCarBatteries(contribs|talk) 06:29, 28 October 2026 (UTC)
@EatingCarBatteries Speaking from my personal experience (both as a former child on the internet, and the stuff I've sent into OS): children like posting stuff like which elementary school they go to, what street they live on, their phone number/email address, their full names, ect. This is obviously, not a good idea, but children do not always have a good sense of internet safety. GreenLipstickLesbian💌🧸 07:20, 28 October 2026 (UTC)
@EatingCarBatteries all of the above. People posting their own information is the most common, especially children as noted by GLL, but also apparent adults posting their contact details (email, phone number, sometimes street address) and occasionally other personally identifiable information. Posting the details of others also happens, most often this is in good faith, e.g. children including details of their family members alongside themselves or people adding too much information to biographies, but doxxing does happen. There are also cases that might be self-promotion/spam, might be innocent oversharing or might be doxxing but we have insufficient knowledge to tell (e.g. a pseudonymous editor sharing a phone number for Joe Bloggs might be Joe themselves or might be someone attempting to doxx Joe). Thryduulf (talk) 11:28, 28 October 2026 (UTC)
And to add to both of those, there are also people who think that the talk pages are for contacting an article subject. Those people tend to leave contact information. On pages dedicated to government agencies, they tend to leave... much more than just contact information. In solidarity, asilvering (talk) 17:16, 28 October 2026 (UTC)
Holy cow, I didn't know it got that bad. This has been insightful, thanks you guys. EatingCarBatteries(contribs|talk) 02:29, 3 November 2026 (UTC)
+1 to above comments - 9 times out of 10 the PII is fairly "typical" (in that in the era of data breaches and data brokers it's probably already out there if you look) so names, email addresses, postal addresses and so on. Once or twice I've seen ID numbers (which in that case is clearly being done with malicious intent). Netstars22(talk to me!) 18:48, 30 October 2026 (UTC)
Hello! I'm not an administrator, but I do enjoy reverting vandalism. Is there some way with which I can apply to access the vandalism detections without running for admin? I'm not interested in seeing any of the personal info stuff, obviously. GrinningIodize (you just lost the game) (talk) 17:21, 28 October 2026 (UTC)
Ooh, this looks interesting and promising - nice work y'all. One interface request I do have: in addition to live refresh, would it be possible to adjust the "rollback" button to do it in a new tab or window, rather than in the same window? Netstars22(talk to me!) 18:31, 30 October 2026 (UTC)
Netstars22, I do this habitually, but depending on your device clicking with the middle mouse button or three fingers on a trackpad will usually do that. (That said, you suggestion is definitely a reasonable default.) LittlePuppers (talk) 21:40, 2 November 2026 (UTC)
We could potentially look at it. The thing is, as @LittlePuppers describes, the middle-click (or can be done by holding Ctrl/Cmd and a regular left-click) lets users choose whether to do it or not. There's not a way to do it the other way around, though - if we say that they always open a new tab, there isn't a shortcut to easily disable that behavior in the same way. EMill-WMF (talk) 22:21, 2 November 2026 (UTC)
After using AbuseReview for a few days now, I can say that I really like the idea and I think that it's a great prototype; however, I still have a couple of questions:
What's the purpose of keeping this tool's vandalism detection features restricted to rollbackers but leaving the (arguably worse) vandalism detectors in Recent Changes open to everyone? Are there any plans to expand access to normal users, and if not, why?
Would it be possible to add a button that lets you skip an item? I find a lot of diffs where I can't easily tell the intent and I'm too lazy to look into it further, but I don't want them to clutter my UI, and I also don't want to feed the tool bad classifications.
Are there plans to feed the classification data obtained from this tool into ClueBot NG for training purposes?
What's the purpose of keeping this tool's vandalism detection features restricted to rollbackers
I'd guess it's probably a combination of this tool being in testing and the flag buttons being used for training data. I think you can view flagged items in recent changes via queries with the "Likely Vandalism" tag, like this. MetalBreaksAndBends (One for all) 14:37, 3 November 2026 (UTC)
Glad it seems to be working well for you. Happy to answer what I can:
It's restricted to rollbackers for quality control - we want the judgments inside the tool (the flag buttons) to represent experienced patrollers. As @MBAB surmised, those judgments are used to improve the tool over time, and how we have confidence that it's working well.
We could consider a skip button. The main thing I would be hesitant about there is whether that would encourage enough people to hide the tougher ones that it ends up weakening the overall inbox-zero "culture" that exists now with Abuse Review, and which we're hoping to maintain.
That would be a better question for the ClueBot team.:) But I will pass that question on, we are staying in touch with them on this project.
@MBAB The tag itself is just as rights-restricted as the AbuseReview queue it powers, so only rollbackers can run that query on recent changes by that tag. EMill-WMF (talk) 15:32, 3 November 2026 (UTC)
In theory, I agree that it would be best to discourage users from attempting to skip trickier stuff, but in practice, I think that this will simply scare people away and crowd out more clear-cut issues. To be clear, I'm not asking for a button that will hide the item for everyone, just for that user. GrinningIodize (you just lost the game) (talk) 16:24, 3 November 2026 (UTC)
This appears to be an industrialized doxxing threat from the Anti Defamation League. Is there a way to formally ask the WMF to request clarification on what exactly they are doing?
What exactly is the concern here? They are allowed to do this. There's nothing illegal, People do the same thing manually when gathering evidence for SPI or ArbCom. The tools that let people track editors are built into mediawiki. It's the user contributions and page history buttons. Packaging that information up into something that highlights shared pages, or related contributions, or similar content of edits is not objectionable and in fact would be a license violation to restrict. What part of CC-BY-SA or GFDL would let us restrict a reuser from using a users contributions to pull out patterns of purported bad actors? --~2026-51667-44 (talk) 07:51, 29 October 2026 (UTC)
If I remember correctly, you have to sign a disclaimer when creating an account consenting to it(It's been a while since I created my account so I could be wrong here), but the ADL aren't requiring that people consent, which is illegal at least in Australia. This is almost certainly a matter for the lawyers because they are breaking Australian law at the very least, and many Wikipedia editors are from Australia. ғѳmɑɭʱɑυե(Any pronouns)ᵀᵃˡᵏ 08:06, 29 October 2026 (UTC)
Why would Australian law be a factor? Its an American non-profit doing the analysis, of a website owned by another American non-profit whose primary hosting is in America. There's no nexus to Australia. Having an Australian user isn't enough. The EU tries to claim having European users is enough to apply EU law as well, and despite that claim Wikipedia has no Data Protection officer, and will not turn over all data relating to a person (withholds arbcom deliberations and discussions) as required by the GDPR. Because it operates under US law.
The group has a First Amendment right to do what they do. Other groups have a right to loudly criticize them. They have the right to not work with them. Wikipedia's license does not allow us to stop them doing what they're doing. Nothing in Australian law has any bearing on that. --~2026-51667-44 (talk) 21:02, 29 October 2026 (UTC)
Because Australian law applies to Australian users. ғѳmɑɭʱɑυե(Any pronouns)ᵀᵃˡᵏ 21:03, 29 October 2026 (UTC)
As the TA says, the Anti-Defamation League are Americans using an American website that operates on American servers. It's still not clear what Australia has to do with this specific case; Australian law only applies to parties who are Australian. (The mere act of analyzing an Australian user on an American site doesn't mean that the analyzer falls under Australian law.) Since neither party is Australian, the presence of Australian law is not relevant here. Accessedgrant (Epicgenius mobile alt) (talk) 03:26, 30 October 2026 (UTC)
Something not being illegal doesn't mean it's not wrong or immoral. It's pretty clear that what they're doing, or planning to do, is POV-pushing via intimidation. sapphaline (talk) 08:58, 29 October 2026 (UTC)
Also, per my post above, it is illegal in at least one jurisdiction. ғѳmɑɭʱɑυե(Any pronouns)ᵀᵃˡᵏ 08:59, 29 October 2026 (UTC)
It is not illegal to analyse someone's wikipedia contributions. In solidarity, asilvering (talk) 09:02, 29 October 2026 (UTC)
Seemed to me like they were implying that they were stalking Wikipedia editors. ғѳmɑɭʱɑυե(Any pronouns)ᵀᵃˡᵏ 09:03, 29 October 2026 (UTC)
@Sapphaline There is no disclaimer when you create a Wikipedia account. Every time you publish an edit you agree to the Wikipedia terms of use, which references the privacy policy, which states that your contribution history is a permanent public record, and goes on to say You should be aware that specific data made public by you or aggregated data that is made public by us can be used by anyone for analysis and to infer further information, such as which country a user is from, political affiliation and gender. There is no requirement for your to sign a disclaimer with the ADL or anyone else for them to do analysis of your contribution history. --Ahecht (TALK PAGE) 20:46, 29 October 2026 (UTC)
I may open a discussion at RSN about this per the precedent established with the Heritage Foundation. ғѳmɑɭʱɑυե(Any pronouns)ᵀᵃˡᵏ 08:09, 29 October 2026 (UTC)
It sounds to me less like they are making a doxxing threat and more like they are announcing that they have been cyberstalking us. ғѳmɑɭʱɑυե(Any pronouns)ᵀᵃˡᵏ 08:16, 29 October 2026 (UTC)
It contains the usernames, not the names. --Ahecht (TALK PAGE) 20:51, 29 October 2026 (UTC)
Although it shows some poor understanding of Wikipedia and some terrible political positions, it does not give evidence of cyberstalking, of tracking users off-wiki or of attempting to doxx. It is an analysis of the publicly available edit histories. Our editing should be up to such scrutiny, and our public edit history is designed to make it so. BobFromBrockley (talk) 12:38, 30 October 2026 (UTC)
By names, do you mean the anonymous usernames we all electively publish under? ꧁Mangroves꧂ 22:50, 2 November 2026 (UTC)
The concern here seems to be that there is an organization with a huge amount of money and political power that is trying to influence and modify what's written on Wikipedia to align with its own world view. This in itself is quite problematic for the same reason that we ban canvassing. Ita140188 (talk) 09:00, 30 October 2026 (UTC)
Just going to link my comment from RSN here - Disturbing if true, but considering the source provided is seemingly an open white nationalist I'm not so sure this should be taken at face value. TheKip(contribs) 11:04, 29 October 2026 (UTC)
Also going to point out here that it's highly suspicious that they're in synagogue attire in a synagogue for a press release and that may be a sign of an antisemitic nutjob using AI to back up their own claims. ғѳmɑɭʱɑυե(Any pronouns)ᵀᵃˡᵏ 11:08, 29 October 2026 (UTC)
No more than when a Christian wears a cross. Slatersteven (talk) 17:04, 29 October 2026 (UTC)
Eh, no it is in fact weird; at least I find it weird as a Jewish person who has attended synagogues belonging to variety of different Jewish denominations over time. A tallit is only worn in the context of specific prayer services, most of which you wouldn’t normally be filming at. It’s not exactly forbidden to wear in other contexts, but it’s not normal to wear outside of specific rituals. Wearing it for a press release about Wikipedia is weird. If we’re analogizing to Christian practice, I’d compare this to donning a nun’s habit or a priest’s frock as a layperson. signed, Rosguilltalk 17:14, 29 October 2026 (UTC)
The supposed source of the clip is a more general talk, rather than a press event. But there are exactly 0 reliable in this whole matter, so I would lean towards this being some kind of missinfo.MetalBreaksAndBends (One for all) 17:26, 29 October 2026 (UTC)
The timing of the event right after Saturday services makes the wearing of tallesim more normal, although it's a weird time for a Conservative congregation to be having a public, filmed event (but maybe they've changed practice around recordings a bit since COVID) signed, Rosguilltalk 17:30, 29 October 2026 (UTC)
Just irreparably ruined my ad targeting, turns out the clip was from some other event (unreliable sources have accuracy issues, who would've guessed!). But, I think he also mentioned monitoring wikipedia at the Sinai Temple event too? Here's the twitter clip, this one seems a bit more verifiable. I'm pretty sure the other guy on the poster was there for it, as there's another clip where the camera isnt only on Greenblatt. MetalBreaksAndBends (One for all) 17:55, 29 October 2026 (UTC)
This source (infolibnews - Chris Menahan) is the white nationalist conspiracy theorist referred to by The Kip above. We should not be linking to their content. BobFromBrockley (talk) 12:41, 30 October 2026 (UTC)
Only rabbis wear them? Slatersteven (talk) 10:04, 30 October 2026 (UTC)
Editors who promote real-life violence are legitimate targets for FBI surveillance. I'm not the one accusing them of crimes, just they should not be surprised if the FBI comes looking for them.tgeorgescu (talk) 07:04, 30 October 2026 (UTC)
And the relevance of this vague accusation (oh sorry, "I'm not" accusation) to this thread is...? Fram (talk) 09:26, 30 October 2026 (UTC)
Or, the ADL is not the FBI (leave to the law that which is the law business) and promoting killing is against the law and should be reported, not lead just to user bans. Ignoring, what the hell do you mean "promote real-life violence" (see wp:agf). Slatersteven (talk) 10:01, 30 October 2026 (UTC)
If the editors were genuinely "pro-Hamas" they would be promoting real-life violence. The issue is that what ADL counts as "pro-Hamas" is not what most informed humans would. BobFromBrockley (talk) 12:42, 30 October 2026 (UTC)
No, you see, we are defending Hamas... somehow... LightNightLights (talk • contribs) 13:11, 30 October 2026 (UTC)
Note that the Israeli mainstream position is that every member of and supporter of Hamas should be killed. Technology exists to kill anyone deemed pro-Hamas based on the location of their online devices. Refaat Alareer and his family were killed as a result of his social media content. There are many other examples.
The ADL calling editors pro-Hamas, irrespective of it being baseless, should be considered a very serious threat.
The best read I have is: tgeorgescu is defending the ADL and thus is attacking people who are pro-Hamas, as "[e]ditors who promote real-life violence".
Conveniently, pro-Israel violence was not brought up. LightNightLights (talk • contribs) 13:27, 30 October 2026 (UTC)
It is their POV, not mine. tgeorgescu (talk) 13:41, 30 October 2026 (UTC)
This is why we have AGF and should not be accusing any editors of promoting violence (we do not have parliamentary privilege here; you do not repeat random accusations). Slatersteven (talk) 13:52, 30 October 2026 (UTC)
White supremacist editors get quickly shunned when admins find out, but they exist.
@Tgeorgescu: can you please explain what you meant? It sounded as if you were making excuses for the ADL by starting about some unmentioned, unknown "editors who promote real-life violence" out of the blue, but perhaps you meant something else? Fram (talk) 13:37, 30 October 2026 (UTC)
That's what the ADL stated: they want to catch editors who advocate for violence. tgeorgescu (talk) 13:39, 30 October 2026 (UTC)
That's not what the ADL stated though.
"We've had to build software to track, to try to identify where editors are coordinating on edits to again poison the articles on Zionism, to poison the articles on Israel, to poison the articles on the Jewish community, our faith". Fram (talk) 13:52, 30 October 2026 (UTC)
I watched the video: "white supremacists... Christian nationalists..." tgeorgescu (talk) 13:58, 30 October 2026 (UTC)
Which video? You are not talking abou the one-minute one posted on X which started all this, but something else? Fram (talk) 14:06, 30 October 2026 (UTC)
The two minutes one. tgeorgescu (talk) 14:10, 30 October 2026 (UTC)
Do you have a link? Fram (talk) 14:17, 30 October 2026 (UTC)
My point: ADL also does the dirty work for liberals. tgeorgescu (talk) 14:19, 30 October 2026 (UTC)
No idea what you mean with that last sentence, but you seem to be doing the dirty work for the ADL, equating all people who edit against the preferences of the ADL with "people who promote real-life violence". Fram (talk) 14:32, 30 October 2026 (UTC)
Going after extremists who promote violence is not a fault. Of course, that should not get overstepped through crying wolf. tgeorgescu (talk) 14:40, 30 October 2026 (UTC)
No but going after people who do not follow the ADL's POV is (crying wolf). Slatersteven (talk) 14:51, 30 October 2026 (UTC)
"they're monitoring the DSA, anti-war activists, pro-Palestine groups, "right-wing extremists," political Islamists, Christian nationalists, and others." Equating this with "going after extremists" (like you said now) or "people who promote real-life violence" who "should not be surprised if the FBI comes looking for them" is an extremely one-sided POV and basically just amplifying their dubious message. Don't know if this is yoour usual method, but it's a very bad look. Fram (talk) 14:55, 30 October 2026 (UTC)
I think we all agree that the ADL shouldn't cry wolf. tgeorgescu (talk) 14:56, 30 October 2026 (UTC)
I don't think we are reading the same discussion. I don't agree with your stance, and am not interested in finding some tiny sliver where we don't necessarily disagree to make it seem as if we agree here. You made very dubious claims to defend the ADL and attack the editors they try to dox or intimidate, while hiding behind some sophistry.
You stated, in full: "That's what the ADL stated: they want to catch editors who advocate for violence." This is not what they stated at all. I have no interest in continuing debate, never mind "agreeing", with someone who makes such false claims to defend an organisation going after Wikipedia editors. Fram (talk) 15:07, 30 October 2026 (UTC)
But thier own statement makes it clear that is what they are doing, going after anyone critical of Israle. Slatersteven (talk) 15:12, 30 October 2026 (UTC)
tgeorgescu, Wikilawyering that "I'm not the one accusing them..." doesn't play. This is an extemely concerning statement and I highly recommend you strike it as it absolutely does read as you stating that editors are promoting real-life violence and the FBI needs to surveil them, which can at the very least be taken as intended to cause a chilling effect even if it doesn't meet the strict letter of WP:NLT. And that's even completely leaving aside that this is about the ADL, not the FBI, and the ADL, not being the FBI, has not the authority or legal power to surveil a ham sandwich. - The BushrangerOne ping only 22:53, 30 October 2026 (UTC)
Well no more than any other user. Slatersteven (talk) 15:03, 31 October 2026 (UTC)
"Software to track Wikipedians" is likely less complicated/evil than we're imagining. Nor does it mean that its a doxxing tool. Sure, it could be very devious and evil. But its more likely to be some vibe-coded contribution checker. CaptainEekEdits Ho Cap'n!⚓ 20:27, 30 October 2026 (UTC)
Proposal: Wikiproject User Safety
Didn't realize there was a new ADL report out with usernames seemed buried in the discussion above. Regardless of whatever the ADL is doing (or honestly failing to do, their analysis is shoddy and lazy, there may be need for users to crowdsource tips together and discuss these. Would a Wikiproject User Safety be a useful idea? User:Bluethricecreamman(Talk·Contribs) 02:56, 1 November 2026 (UTC)
Usernames can be checked just by looking at a page edit history; it's public information. All they are doing is collating it. They are not (as far as I know) publishing real names or address links to those usernames (indeed, I think anonymity is one of ther whines). So (as I said above) this is no more than I could do, the only difference is they can be arsed. The (ADL) needs watching over this, but no more action is needed right now, with a clear statement: "THIS FAR, AND NO MORE". Slatersteven (talk) 11:04, 1 November 2026 (UTC)
Yeah, i think the edit contribution analysis anyone can do. Interesting to hear how they make it sound high tech analysis, i suppose User:Bluethricecreamman(Talk·Contribs) 17:41, 1 November 2026 (UTC)
Source Verification: Internal Review Findings
Experiment source verification suggestion live in Suggestion Mode
Hi y'all – we wanted to follow up about the source verification project, where we’ve just completed an internal review of the suggestions it produces and have a proposal about making it accessible for testing with experienced editors.
With your help, we defined criteria to select the 1,000 articles we would generate an initial dataset of source verification suggestions for.
We (staff) have created this dataset and completed an internal review of it.
What follows is a summary of what we found, what we think these findings mean for this work going forward, and the open questions we're left with.
What do you think about the conclusions we've reached and next steps we're proposing?
Conclusions
We reviewed 134 source verification suggestions from the dataset and discussed them together as a team. What follows are the conclusions we're reaching:
Article claim and source URL extraction worked well. Generally, the model succeeded at extracting claims and source URLs from Wikipedia articles. This enabled us to focus on the quality of the verdicts the model reached and the rationale it provided for each.
The model outputs sufficiently reliable verdicts. In 71% of the suggestions we reviewed, internal reviewers agreed the cited source did not fully support the claim, including 49% where it did not support the claim at all. Several of these were concrete discrepancies between article and source (see the Lucille Ball and Alamgir examples below).
Taken together, we think the source verification suggestion is reliable enough to warrant a wider volunteer-led review. Please see "Proposed next steps" for how we see this playing out.
Proposed next steps
With the "Conclusions" above in mind, we see the next steps as the following:
On Tuesday, 3 November 2026, we (staff) make the initial dataset of source verification suggestions available to you all in two forms:
Note: These suggestions would only be available to people who have both 1) published ≥100 cumulative edits and 2) enabled experimental edit suggestions within Preferences
As a spreadsheet so that you can more easily see patterns across suggestions and locate articles where you can experience the suggestions in-context via Suggestion Mode
Note: we do not think reviewing suggestions in the spreadsheet alone is sufficient for forming an opinion about the usefulness of the suggestion.
Note: we see the purpose of y'alls review as making sure we wouldn't be wasting unsuspecting experienced volunteers' time with a large number of erroneous suggestions by enabling source verification suggestions as default-on within Suggestion Mode.
We (staff + volunteers) will then look at data from both reviews to decide: Do we make the source verification suggestion available by default within Suggestion Mode to experienced volunteers so that we can evaluate how helpful they find it in real world use?
Note: to evaluate "how helpful [volunteers] find it in real world use" we would be analyzing engagement with the suggestion, the quality of edits made with it, the edit completion rates of edits where it was activated, and qualitative feedback from experienced volunteers who tried it.
Open questions
What additional filters might be useful to place on the model and/or what articles it runs on to create a more useful user experience? E.g. Setting a cap on how many suggestions are shown per article, preventing suggestions from appearing on list pages, etc.
What information, shown alongside a suggestion, might help you decide more quickly whether it’s right? E.g. deep-linking to the relevant passage within the source, linked to the diff where the claim was added, etc.
Currently if the claim is not supported, we expect that the editor will fix it or add a {{failed verification}}. Is that reasonable or are there other actions we should try to directly enable?
How should the suggestion behave when it can't read a source (e.g. paywalled, offline, etc.)?
Detailed findings
Internal review findings
Process
We generated a total of 16,585 suggestions for a sample of articles that met the criteria listed here
We compiled those 16,585 suggestions into a spreadsheet, split up by category (e.g. blp, history, ctop, {{dubious}}, {{Failed_verification}}, {{ai generated}}, etc.)
6 staff members reviewed a total of 134 of the 16,585 suggestion
For each of the 134 suggestions, staff:
Reviewed the claim text the model extracted from the article and evaluated whether they thought the claim was complete
Reviewed the verdict the model reached (Not supported or Partially supported), the rationale the model gave for why it reached said verdict, determined to what (if any) extent they thought the model prediction was correct, and offered an explanation as to why they reached the conclusion that they did
We then looked over the suggestions our colleagues reviewed and documented general reflections about them
Lastly, we calculated the rates at which staff verdicts did and did not align with the verdicts the model reached (see "Quantitative findings" below)
Quantitative Findings
In 49% of cases, reviewers agreed with model in concluding source does not support claim at all
Flagged claim
Notes
AdventHealth opened a 200,000 square feet (19,000 m2) emergency room and primary care facility on Walt Disney World's campus on May 17, 2023.
"AdventHealth facility is only 19 thousand sqft, not 200 thousand" (page: AdventHealth, check id 3c1bac7e1c82eaae)
[Lucille Ball’s] acting forays were stilled at an early stage when she became ill with rheumatic fever and was unable to work for two years.
"Claim states she had rheumatic fever but source states it was rheumatoid arthritis." (page: Lucille Ball, check id: 3f99f519c8f527fc)
[Mary of Teck] is buried beside her husband in the nave of St George's Chapel, Windsor Castle.
"Queen Mary was buried in the North Nave Aisle, not the nave itself." (page: Mary of Teck, check id 012418710cd2e07e)
"Though Alamgir was unhurt, three vehicles (including the one in which he was riding) were vandalized, and ten people were injured." Link to claim
The claim included within the Wikipedia article states that the car the subject was riding in was damaged. Although, the source cited says the car he was riding in was NOT damaged. These two statements contradict one another. (page: Mirza Fakhrul Islam Alamgir, check id: 01836695cfc49863)
In 22% of cases, reviewers agreed with model in concluding source does not support claim fully
Flagged claim
Notes
On 17 June 2026, Keir Starmer announced that the Children's Wellbeing and Schools Act 2026 has passed, which puts a ban on children under 16 having accounts and accessing social media apps
The act was passed on a different date and doesn’t ban anything by itself (page: Online age verification laws by country, check id: cbeeec4e88959d7a)
[Amis Harvest Festival] takes place every July to September.
Article claims that the Amis Harvest Festival takes place every July to September. The source only mentions summer time, not fixed dates, and that the dates can vary across villages. (page: Amis people, check id: 740542a585970b96)
A statue of Earhart by Ernest Shelton was erected circa 1971 in Los Angeles, California.
Source talks about the statue and when it was erected, but doesn’t mention the sculptor’s name as listed in the claim. (page: Amelia Earhart, check id: 0366ad517930495e)
In 28% of cases, reviewers disagreed with model in concluding source does not fully support claim
Flagged claim
Notes
“Marie Tussaud was employed to make a death mask of her head.” -
The model thought that “her” refers to Marie Tussaud herself (page: Marie Antoinette, check id: 770d5c8b71c195f2)
It would also be played by Weir's RatDog after the dissolution of the Grateful Dead.
Source does say that Ratdog played "Victim or the Crime", and the dates mentioned in the article are after the dissolution of the Grateful Dead. (page: Victim or the Crime, check id: 2d912ad52fc7efe4)
Following the release of Citizen Vigilante in late June 2026 and its ban in Germany, the film was promoted by Elon Musk on X, greatly boosting its reach.
A second source verifies the part of the claim that the model says the source doesn’t verify (page: Uwe Boll, check id: 0e4b8066c9135210)
Qualitative findings
Perceived significance of discrepancies: There are some claims that the model flagged as not being supported by their source(s) that, upon closer inspection, involved relatively small discrepancies between the information included within Wikipedia and the source. Editor opinions may vary on how consequential they consider these discrepancies to be.
E.g. a film reaching some sort of revenue milestone within 3 vs. 4 days after launch.
Selection bias: The quantitative findings should be taken with a grain of salt considering we did not review suggestions randomly. As such, there may be some selection bias at play.
URL Extraction: as part of the initial internal review process, each person was asked to verify whether the model had accurately extracted the source URL from Wikipedia. In 100% of cases, the model was successful in this operation. As a result, we removed this URL checking from the review process.
Claim Density: the number of claims the model flagged as needing review varied article-by-article. Some of this variance can be explained by genuine differences in content quality, others, as we came to learn, can be explained by the model handling claims that start with pronouns (e.g. "he/she") less well. There are a series of improvements we think could improve the model that we'd value y'all input on. See "Open questions" below. For context:
A typical article in our sample has over 150 citations, and the model flagged about 17 of them as not fully supporting their claims.
There are outliers: 2% of articles have >100. Lists and some biographies (e.g. 2026 deaths in the United States - see screenshot). This is due to a combination of the higher background level of sloppiness, citation density and tool deficiencies, for instance the pronoun problem.
Number of Source Verification Suggestions per article (initial en.wiki dataset).Experimental Source Verification Suggestion live in Suggestion Mode.
Suggestion invalidation
Suggestions appear if and only if the claim and source have not changed since the article was initially analyzed and the source is not followed by a qualifying template[i]. That is, if the suggestion has been addressed in any capacity - a user has updated the source or the claim, whether prompted by suggestion mode or not, or has added a qualifying template - then the suggestion will no longer show to any user.
If a user dismisses a suggestion by clicking the ‘dismiss button’, they won’t see that suggestion for the remainder of the edit session. However, they - and other users - will continue to come across it in future edit sessions until it is actually addressed.
The Editing Team recognizes the importance of preventing what experienced volunteers agree are invalid suggestions from appearing for themselves or other people in the future. We are tracking this work in T427236 and T433147. While we have not yet defined a timeline for when this functionality will be in place, you can see the technical thought we've put into the architecture that would help enable it.
i. Wikis have the option to locally configure a ‘templateToInsert’ for Source Verification in the on-wiki config. This is the template that (1) will be inserted after the claim when the user dismisses the suggestion with the ‘suggestion is valid’ response, and (2) will be checked for to determine if the suggestion is still valid and should be surfaced to other users. For example, the recommended ‘templateToInsert’ for enwiki would likely be {{Failed-verification}}.
Hi, does the WMF Tool already have access to the wikipedia library? I testet the script from Alaexis however in all the testet articles, 90% of the sources were not available, even open access sources. If its like that, the tool is useless. While i can certainly copy the text into Alexis's tool and let it be checked, but then I might as well just ask the AI in my browser. Or myself. In my opinion, the tool is only useful if it saves me the trouble of looking up the claims in the sources. Then all I have to do is verify the AI's output. The Other Karma (talk) 07:27, 31 October 2026 (UTC)
We're working on requesting access for the source verifier, through the same route that Earwig's Copyvio Detector already uses. CMedelius-WMF (talk) 00:18, 4 November 2026 (UTC)
I am not sure 71% is sufficiently reliable - that's calling someone out to look specifically at a source to double-check something for no reason more than 1/4 of the time. And since we can't actually dismiss the suggestions, that... will be a lot of incorrect suggestions constantly coming back up.
Re: whether a source agrees fully or not, I'm not really seeing a difference between "agree not fully" #3 (the Earhart one) and "disagree not fully" #3 (Uwe Boll). In both cases, as I read these, there is some information that was not in the source, and the model is equally correct in both instances. It's just that in the second there is also another citation. This will surely affect quite a lot of articles - can the model not be trained to look at all of the footnotes given, not just the first?
Re: "north nave aisle", it's quite normal to describe that part of the cathedral as "the nave". So the claim is actually fully correct, if not as specific as it perhaps ought to be. In solidarity, asilvering (talk) 19:51, 29 October 2026 (UTC)
I am not sure 71% is sufficiently reliable…
You may end up being right in thinking this suggestion is not reliable enough, in its current form, to warrant a deployment as a default-on suggestion within Suggestion Mode.
Although to clarify, we were not seeking, nor equipped to, make a decision about a wider deployment with this initial internal review. This internal review was narrowly scoped to answering, "Do we think these suggestions are reliable enough to warrant experienced volunteers allocating valuable time to reviewing them?"
And based on the review we conducted, we do think the suggestion is reliable enough to invite interested experienced volunteers to review them.
It's only after we (staff and volunteers) have reviewed these suggestions would we together decide "Do we make the source verification suggestion available by default within Suggestion Mode to experienced volunteers so that we can evaluate how helpful they find it in real world use?" (See Step 3 above)
Does this get at the point you're making? Might there be something I'm missing/misunderstanding?
...that's calling someone out to look specifically at a source to double-check something for no reason more than 1/4 of the time.
On the time spent piece, you may well be right and I think it's only through real world use will y'all be able to make a judgement call on this.
As I see it, what matters is whether the time spent checking false positives outweighs the time saved by catching real errors. To what extent (if any) does this thinking align with how you're viewing this tradeoff?
And since we can't actually dismiss the suggestions, that... will be a lot of incorrect suggestions constantly coming back up…
We agree with you here: without the ability to dismiss suggestions, multiple volunteers could see the same unhelpful/incorrect suggestion which would in turn increase how precise the model would need to be before we’d (staff + volunteers) be comfortable putting it in front of people. Potentially, to an unachievable point even.
So it's not ambiguous: we see work on dismissals as a question of when not if and we're open to the possibility that this functionality would need to be in place before source verification suggestions are enabled by default.
Thank you for thinking through this critically. Please let me know if any facet of the above leaves questions/concerns open in your mind or causes new ones to emerge. PPelberg (WMF) (talk) 22:33, 29 October 2026 (UTC)
@PPelberg (WMF), we must be misunderstanding each other, because I read "Do we think these suggestions are reliable enough to warrant experienced volunteers allocating valuable time to reviewing them?" as basically "is this reliable enough to be deployed?" with extra words. What's the difference, for you? In solidarity, asilvering (talk) 02:05, 30 October 2026 (UTC)
@Asilvering, the distinction is between enabling them as experimental suggestions, visible only to editors who have opted in to experimental suggestions, and making them regular suggestions, visible to all editors who meet the edit-count criteria. Alaexis¿question? 07:34, 30 October 2026 (UTC)
Thank you for adding this context, @Alaexis. @Asilvering: does what Alaexis shared above address the confusion you spotted? PPelberg (WMF) (talk) 22:30, 30 October 2026 (UTC)
No, I think we may have a fundamental disagreement. But I think that disagreement could be resolved by reframing how the edit suggestion is presented to editors. Right now, the wording of the suggestion begins from the presumption that the model is intended to highlight problematic sentences; the purpose of this edit suggestion is to find and fix incorrect sentences on Wikipedia. It is understandable that the suggestion is worded in this way, because that was indeed the intent of this project. However, with a 71% hit rate, we cannot truthfully say that the model is highlighting problematic sentences. So in fact what we have is an edit suggestion that fails 29% of the time even if the editor correctly applies WP:V when responding to the check, which we can't be certain about either.
If you cannot raise the hit rate over 71%, the suggestion should be reframed as a spot check. We do sourcing spot checks all the time (for example, at WP:GAN), and they do not start from the presumption that the sentence is incorrect. (At GAN, you're instead starting from the presumption that the sentence is correct, if anything.) So instead of "a model has identified that this claim might not be supported by its source", you end up with something neutral like "a model has selected this sentence for source-checking". Under those terms, you have an edit suggestion that works 100% of the time, and ideally results in improvements to the article 71% of the time or more often. Does you follow?
In any case, we definitely need the ability to dismiss suggestions as incorrect (ie, to mark it as "fact checked, it's fine") before this leaves "experimental mode". In solidarity, asilvering (talk) 00:03, 31 October 2026 (UTC)
I like the idea of carefully framing the suggestion as a prompt to spot-check a citation. As a relatively neutral framing, it could reduce automation bias and increase the chance that a person actually goes and looks at the source. It also opens up the opportunity to prompt spot-checking of sensitive/high-risk statements where the automated verifier can't access or parse the citation. The research for the reference need model (paper) might suggest ways to help find those statements. Dreamyshade (talk) 00:57, 31 October 2026 (UTC)
+1, I assumed this was already going to be the case Kowal2701 (talk, contribs) 01:03, 31 October 2026 (UTC)
Agreed re the dismissal and I think that it should be recorded on-wiki. This way the community will be able to measure the effectiveness of suggestions (edits vs dismissals). Having it recorded will also reduce abuse as the editor will know that the dismissal is recorded and attributed.
Not sure about the framing as "spot checks". Spot checks ("a quick examination of a few members of a group instead of the whole group" ) make it look like a random sample, while in reality it's more like a screening (medicine), where a positive result means that a more thorough check is needed.
I'm sure the language can be adjusted, I think one widely used tool that has the same probabilistic nature is Earwig's Copyvio Detector. It describes likely copyvios as "Violation suspected" (example - here the top one is actually not a copyvio but a wiki mirror). Alaexis¿question? 17:20, 31 October 2026 (UTC)
Re: whether a source agrees fully or not, I'm not really seeing a difference between "agree not fully" #3 (the Earhart one) and "disagree not fully" #3 (Uwe Boll). In both cases, as I read these, there is some information that was not in the source, and the model is equally correct in both instances. It's just that in the second there is also another citation. This will surely affect quite a lot of articles - can the model not be trained to look at all of the footnotes given, not just the first?
Good catch. The model is trained to evaluate all footnotes following a claim, not just a subset of them. If it happens to return a verdict that doesn't cover the complete set, then we do a second check on the suggestion mode side to ensure this suggestion isn't shown to users. So this example is one that was not caught by the model but would still not have been shown in suggestion mode; including it as an example here was our mistake, as it doesn't reflect the actual user experience. CMedelius-WMF (talk) 22:47, 29 October 2026 (UTC)
With regards to #1, I think it'd make sense to avoid checking articles being rapidly edited, new articles, and claims that cite paywalled or offline sources as referred to in #4. On #2, making the model's reasoning available could help the community figure out where strengths and weaknesses are. MetalBreaksAndBends (One for all) 19:57, 29 October 2026 (UTC)
On #2, making the model's reasoning available could help the community figure out where strengths and weaknesses are.
@MBAB, could you please say a bit more here about the sort of reasoning you'd find helpful?
I ask this because currently, the model outputs a rationale for each suggestion it outputs. E.g. “Claim states she had rheumatic fever but source states it was rheumatoid arthritis”. This leads me to think you might be thinking of something else/in addition...
I think it'd make sense to avoid checking articles being rapidly edited, new articles...
You and @CMDeach raised this. Could you please share a bit more about where this recommendation is coming from?
For what it's worth, we'd been assuming (perhaps inaccurately!) that it'd be useful to check articles that are new and being rapidly edited because they'd be more likely to contain inaccuracies. Then again, maybe you're thinking adding another actor into these situations (in this case a machine) could complicate things? PPelberg (WMF) (talk) 23:46, 29 October 2026 (UTC)
I was talking about reasoning tokens, as reading them can be helpful for understanding a model's thinking in further depth than a 50 word summary can provide. Obviously, this shouldn't be exposed to editors by default (especially if your model of choice has a GPT-6-astra-esque affinity for not using the spacebar).
I suggested avoiding rapidly edited and new articles for 2 slightly different reasons. New articles because they're going to be in the NPP queue, which means they're going to be checked soon (although ai verification built for the NPP workflow could be promising. e.g. not doing the check until the page is being reviewed and factoring in other steps like a source's value for notability). For rapidly edited articles, it's because there's a higher chance that any given checked text will be modified requiring rechecking. My concerns on both of these is less that the system might get something wrong and more that compute might be misallocated. MetalBreaksAndBends (One for all) 00:26, 30 October 2026 (UTC)
I was talking about reasoning tokens, as reading them can be helpful for understanding a model's thinking in further depth than a 50 word summary can provide. Obviously, this shouldn't be exposed to editors by default...
Ah, got it. I agree with you in thinking that exposing this level of detail is best suited for somewhere deeper in the experience so folks who are especially curious can find it. Thank you for adding this context.
In case it's interesting to know: I think the checks we ran used an un-reasoning version of Qwen 3.8. Although, @SSalgaonkar-WMF will be able to say for certain. PPelberg (WMF) (talk) 22:33, 30 October 2026 (UTC)
I suggested avoiding rapidly edited and new articles for 2 slightly different reasons. New articles because they're going to be in the NPP queue, which means they're going to be checked soon (although ai verification built for the NPP workflow could be promising. e.g. not doing the check until the page is being reviewed and factoring in other steps like a source's value for notability).
Understood. To confirm the NPP use case, you're thinking that it would be worthwhile to wait to run the suggestion on new articles until after an experienced volunteer has validated the article meets notability requirements? Specifics aside, we're with you in thinking there is/are NPP-specific use cases here that feel promising.
For rapidly edited articles, it's because there's a higher chance that any given checked text will be modified requiring rechecking. My concerns on both of these is less that the system might get something wrong and more that compute might be misallocated.
Oh, okay. I see. Understood. Thank you for explaining that. PPelberg (WMF) (talk) 22:37, 30 October 2026 (UTC)
...and claims that cite paywalled or offline sources as referred to in #4.
Right now, the suggestion ignores offline and paywalled sources. Although, this leads me to wonder: can you see it being useful for the UI to indicate when the model tried to evaluate a claim and was unable to do so because a source was offline or paywalled? Do you think it ought to instead show nothing silently? Something else? PPelberg (WMF) (talk) 23:49, 29 October 2026 (UTC)
This is also mostly based in compute allocation, although there is also a volunteer allocation angle too. If a paragraph has multiple citations at the end with one or more being inaccessible, it can't be known if any given claim not supported by the accessible sources is actually unsupported or just supported in the inaccessible ones. So, for those paragraphs any given suggestion is less likely to be helpful, due to it only being trustworthy where the content explicitly contradicts a source. A flag for inaccessible sources would help avoid errors where correct content is removed because inaccessible sources were skipped. But unless a system were devised to direct verifiers with access to inaccessible content to suggestions, I think bailing at the first sight of problematic sources would offer the best rate of "tokens per abstract unit of improvement". MetalBreaksAndBends (One for all) 00:57, 30 October 2026 (UTC)
In light of WP:PAYWALL, I think claims that cite paywalled or offline sources should be up for review as well. But we'd need additional options for feedback, like "Paywalled/offline, checked", "Paywalled/offline, couldn't check". Jo-Jo Eumerus (talk) 10:22, 6 November 2026 (UTC)
I've been working with a Master students evaluating the script, and they found a couple of categories of mistakes it often makes. I will ask them to share. Perhaps that can raise the 71% to a higher level. I agree that 71% is not quite the accuracy we would like to see. Source verification is incredibly time-consuming as an activity, and newish (100-1000 edits) editors are also not great at it. Will this result in properly cited text being removed?
In my two GAs I've run it on today, I disagreed with all 5 suggestion that things were wrong (2 were subjective), but maybe the script uses a different model than the one you have access to? In solidarity, —Femke (talk) 🐦 21:19, 29 October 2026 (UTC)
...I disagreed with all 5 suggestion that things were wrong (2 were subjective), but maybe the script uses a different model than the one you have access to?
@Femke: might you be able to share the version of the script and/or model you were using so that we can check if they're the same? Also: might you be able to share the following so that we can see the suggestions you encountered? It'd be helpful to know:
The permanent link to the version of the article you were editing
The text of the claims the model generated suggestions for
The verdicts and rationale the model gave
I've been working with a Master students evaluating the script, and they found a couple of categories of mistakes it often makes. I will ask them to share.
Seeing what they've found would be a big help. Thank you for offering!
I agree that 71% is not quite the accuracy we would like to see. Source verification is incredibly time-consuming as an activity, and newish (100-1000 edits) editors are also not great at it. Will this result in properly cited text being removed?
I'm going to hold off on responding to the "71% accuracy" part for a moment because I'm going to come back to that in a response to the point Asilvering raised about this above. Now, with regard to who the proper audience might be for this suggestion, we think y'all are best equipped to decide and implement what you think is best via Special:EditChecks. PPelberg (WMF) (talk) 21:41, 29 October 2026 (UTC)
I would love 100+ edits to be the audience, as that's the type of editors we want to teach to fact check everything. But.. I think putting the weapon in their hands would be a bit too dangerous as is. I suspect many people will trust the tool, and where the tool is wrong, it's usually in more complicated facts I also think editors will get wrong.
I'm off to bed soon, but I ran it on post-exertional malaise (4 claims flagged as incorrect) and Ursula Bethell (1 claim flagged). Two of the errors seems to be, if I understand it correctly, from sentences supported by two sources, where 1 of the sources didn't completely cover all the content. One said that 'irreversible decline' was not the same as 'permanent decrease', which is probably a bit more subjective as permanent may be a slightly stronger word, but I'd say that's the same. Don't remember the other two, would have to rerun. Will rerun tomorrow with the requested info.
I used Aleaxis' script with the default settings, so Lift Wing. In solidarity, —Femke (talk) 🐦 21:50, 29 October 2026 (UTC)
For reporting issues that you notice from particular suggestions, it may be easier to use the tool's built-in "comment" button, which sends reports to User talk:Alaexis/AI Source Verification with metadata about the suggestion itself. Sdkb‑WMFtalk 18:39, 30 October 2026 (UTC)
The edit suggestions have a different UI which doesn't have a comment button (see the screenshot above). @CMedelius-WMF, maybe you can explain what options the editor has and how they work? Alaexis¿question? 19:17, 30 October 2026 (UTC)
Yes, the edit suggestion appears with two options: "Revise", which brings them into a focused edit state, and "Dismiss", which opens up a survey to provide an answer on why.
It also has the normal dropdown that is shared by all edit suggestions, which contains an option to provide feedback that is then posted to mw:Talk:VisualEditor/Suggestion_Mode/Feedback. We've internally been discussing a change to make the feedback destination configurable; and in this case, it may be helpful to direct feedback to User talk:Alaexis/AI Source Verification. CMedelius-WMF (talk) 19:48, 30 October 2026 (UTC)
I would love 100+ edits to be the audience, as that's the type of editors we want to teach to fact check everything. But.. I think putting the weapon in their hands would be a bit too dangerous as is. I suspect many people will trust the tool, and where the tool is wrong, it's usually in more complicated facts I also think editors will get wrong.
1) It's important newer editors be equipped with the know-how and tools to help out with verifying claims are corroborated by their accompanying sources
2) I see potential in this suggestion helping newer editors onboard into and complete this work
3) Before "2." can happen, we need to be confident that the user experience is – let's call it – resilient to people assuming the suggestion is always right.
Assuming the three points above accurate reflect the points you were making @Femke, I think we're aligned here. PPelberg (WMF) (talk) 22:45, 30 October 2026 (UTC)
Great to see this work continuing and progress being made.
For my part, idk if ~70% is too low or not. Obviously, higher is ideally better, but I think most people aren't right 70% of the time, so seems to me the model is doing pretty good:-)
The concern I have is that ~150 out of 16k is a tiny percentage, and as they were apparently not randomly selected, I'm not sure if the results are representative at all of anything. Of course, the solution to that is what you're already planning to do: have more people do more testing, which seems like the right next step regardless. Levivich (talk) 22:01, 29 October 2026 (UTC)
yeah, ~70% is well worth it, way better than any other way we have of finding WP:V failures (which is nothing). Issue is whether people will put absolute trust in it or be lazy/go with the balance of probability and not actually read the source before removing stuff Kowal2701 (talk, contribs) 00:50, 30 October 2026 (UTC)
@Kowal2701, that's true for, say, Alaexis's userscript. I'd be quite happy to use a userscript with a 71% hit rate, because I'd be using that mostly informationally - ie, my use case is that I'm already trying to fact-check the article, and the script would be helping me focus on the parts that seem most likely to be wrong. That's great. But in Suggestion Mode, what it's doing is surfacing an error to an editor, possibly a relative newbie, and directly asking them: hey, does this source verify, because I don't think it does? And that's a very different kind of mode that ought to have a higher hit rate than 71%, in my opinion. In solidarity, asilvering (talk) 02:03, 30 October 2026 (UTC)
@Asilvering, agreed that default-on suggestions is a different mode from a userscript. However right now these suggestions will be experimental, visible only to opted-in users, and the whole point of the experimental stage is to check whether these suggestions work before anyone decides to make them regular ones. It's quite possible that based on the findings of the experimental stage we'd fine-tune the pipeline and/or enwiki could raise the edit count threshold. Alaexis¿question? 07:43, 30 October 2026 (UTC)
I think people should easily be able to opt-out of it. Tbh what I don't like about Alaexis' script is reading the LLM's reasoning, I'd rather it just listed the claims with their corresponding sources, possibly on the talk page, and people can just {{tick}} or {{cross}} and sign or whatever. I expect its false positive rate is much higher on good quality articles, we could exclude FAs and GAs but idk what else Kowal2701 (talk, contribs) 15:20, 30 October 2026 (UTC)
...hat I don't like about Alaexis' script is reading the LLM's reasoning, I'd rather it just listed the claims with their corresponding sources, possibly on the talk page, and people can just Y or N and sign or whatever...
@Kowal2701: the simplicity of the workflow you're describing resonates with me and I'm curious: what don't you like about seeing the LLM's reasoning? E.g. might you not trust it and therefore find its presence non-additive?
...I expect its false positive rate is much higher on good quality articles, we could exclude FAs and GAs but idk what else...
The thinking there being those articles have already been closely reviewed by knowledgable volunteers? PPelberg (WMF) (talk) 17:14, 30 October 2026 (UTC)
what don't you like about seeing the LLM's reasoning? Yeah, partly because I wouldn't trust it, partly because I'm sick to death of reading AI-generated text lol. I don't think there's really any benefit to it since in all cases the human will have to check the source for themselves anyway, though ofc having the reasoning is useful for testing the feature re: how accurate it is.
The thinking there being those articles have already been closely reviewed by knowledgable volunteers? Yes, and content written by experienced editors is usually of a similar calibre. Maybe we could use the mw:WikiWho API to exclude content written by editors who have GAs/FAs, or if we wanted it to be more restrictive exclude content authored by users w autopatrolled? Kowal2701 (talk, contribs) 18:24, 30 October 2026 (UTC)
One of the findings from my student this summer was that the script detected loads of issues at a FA that they believed were true errors. I'm curious to see whether I agree. In solidarity, —Femke (talk) 🐦 18:29, 30 October 2026 (UTC)
yeah, I'm sure it would identify true WP:V issues in FAs and content by users w autopatrolled, but given such issues generally occur less often in such content I assume there'd be a higher rate of false positives? Kowal2701 (talk, contribs) 18:32, 30 October 2026 (UTC)
Yeah, you're right that there will be more false positives in such articles (the false discovery rate would be higher). On the other hand, it's also more embarrassing to have such errors in GA/FA, so fixing them may be worth extra time spent on checking duds, and Femke’s student found real errors at an FA. I think it's worth tracking this while the suggestions are in the experimental phase: errors fixed vs discarded suggestions in GA/FA. Alaexis¿question? 18:58, 30 October 2026 (UTC)
it's probably worth incorporating it into GAR/FAC instead of random spotchecks Kowal2701 (talk, contribs) 22:37, 30 October 2026 (UTC)
@Kowal2701, I don't think that's a good idea, because the script only works on a subset of sources, and that subset of sources is already the set that are easier to check. Because reviewers who make random spotchecks are often quite determined to maintain that randomness, they will request access to harder-to-find sources from the nominator and go digging where otherwise they might not, and this is an important part of that process. Alaexis's script can be perfectly useful at GAN/FAC, especially if you're pretty sure the whole article is dubious and you're looking for some low-hanging fruit so you can fail the review and send it back for more editing. But that random spot-check has real benefits and it would be a mistake to incorporate this script instead. In solidarity, asilvering (talk) 00:13, 31 October 2026 (UTC)
Yeah, partly because I wouldn't trust it, partly because I'm sick to death of reading AI-generated text lol. I don't think there's really any benefit to it since in all cases the human will have to check the source for themselves anyway, though ofc having the reasoning is useful for testing the feature re: how accurate it is.
I do get that, I'm tired of seeing AI-generated text everywhere too. Though in terms of benefit, I'd argue that it does save time for the user; it tells you exactly what to check in the source, so it's easy to spot when it gets something wrong. Without that reasoning, you'd have to work out from scratch why the claim/source was flagged.
You can see examples of the reasoning in the spreadsheet that we'll share. I'm curious to know your thoughts after reading through some of the rationales. CMedelius-WMF (talk) 00:16, 4 November 2026 (UTC)
Another thing, we might want to exclude articles with {{AI-generated}} tags, since it's not really fair to have someone hone/polish such content when it'll have to be removed or rewritten from scratch Kowal2701 (talk, contribs) 18:30, 30 October 2026 (UTC)
Oh, the notes are the AI's reasoning. Here I thought it was the reasoning of the staff who had agreed or disagreed with the model. That makes me much less grumbly. Though I still think 71% hit rate is far too low for a useful edit check. In solidarity, asilvering (talk) 20:26, 30 October 2026 (UTC)
The concern I have is that ~150 out of 16k is a tiny percentage, and as they were apparently not randomly selected, I'm not sure if the results are representative at all of anything. Of course, the solution to that is what you're already planning to do: have more people do more testing, which seems like the right next step regardless.
Very fair - the results of our internal review shouldn't be taken as a measure of overall accuracy, for the reasons you've stated. However, the purpose of the internal review was more so to perform a sanity check that the suggestions were good enough to be worth additional reviewers’ time. We see volunteer review as the next step because we think you all, as people experienced with verifiability work, bring the nuance needed to differentiate between suggestions that identify genuine issues and those that you'd consider noise. CMedelius-WMF (talk) 08:23, 30 October 2026 (UTC)
For question 1, I agree with MBAB, an automatic tool should avoid frequently edited articles (if someone wants to run it manually, they know what they might be getting into). I don't see a problem with running these on all lists, if the question is whether the tool can understand table sourcing then have it avoid tables on all articles. For question 2, the linked diff is very helpful, especially as it will pick up WP:TSI issues where sources and text get separated by later edits. For 3, please include a nudge for an editor to use the |reason= parameter in fv. For 4, shouldn't the tool simply avoid those situations? More broadly, there seem to be different categories of errors that could be picked up, and it might be useful for the model to distinguish. Of the four "does not support claim at all", the first seems a case of somehow putting in the wrong units. I've found these sorts of numeral mismatches something that models are good at picking up with relatively few errors, if the model can figure out if something counts as a numeral mismatch that might provide a useful subset. The rheumatic fever is another sort of mismatch, worth flagging and seems correct in this case but it seems vulnerable to an llm not understanding wider context. The nave example asilvering has mentioned above, our nave article describes "nave" as having sensu strico and sensu lato meanings, this is perhaps an llm not understanding context and being overly prescriptive. The fourth example is a mix of numbers and prose, harder to classify. However, I don't see how it is distinguished from "does not support claim fully", as the source does support other parts of the text, such as the individual being unhurt. CMD (talk) 23:11, 29 October 2026 (UTC)
In general, I like these next steps. I think that they're reasonable and would allow a much better analysis of reliability. I would guesstimate that in my own experimentation with Claude, about 75% of its suggestions on source-text matches are accurate, so your findings so far roughly track my own experience. In terms of answers to the questions, I think a cap on suggested edits per-page, per-editing-session, is reasonable and that a link to a relevant passage would be helpful. I think if the model/suggestion cannot access a source, it should not take any action to avoid a situation where a new editor also cannot access the source and improperly removes a valid claim. ThadeusOfNazereth(he/him)Talk to Me! 23:39, 29 October 2026 (UTC)
Adding support for dismissing suggestions [1], potentially reframing the suggestion as a neutral spot-check [2], deciding what (if anything) to communicate within the UI when the model is unable to access a source [3], enabling volunteers to configure what wiki page the feedback people share via the suggestion's built-in "Report feedback" function gets published on [4], revising what articles the suggestion is not shown on [5][6[7]...to me, these all sound like important topics to revisit after we've made the initial dataset available to you all and before we consider a default-on deployment of any kind.
In line with the above, you can expect the following...
Tomorrow:
We will share a spreadsheet containing ~14,000 source verification suggestions
Continuing to respond to points you've raised and that we've not yet responded to
Please let me know if anything here brings questions/concerns to mind. PPelberg (WMF) (talk) 00:21, 4 November 2026 (UTC)
Updates
1. The experimental source verification suggestions are now enabled for experienced editors who have opted into both A) Suggestion Mode and B) experimental edit suggestions. To see the suggestion in action, please do the following:
Visit Beta Features and enable the Suggestion Mode beta feature
Visit Editing preferences and make sure the Enable experimental edit checks and suggestions setting is enabled
2. The 14,000 sample source verification suggestions can be viewed in this spreadsheet
Asks
For any issues you notice with the suggestions, please use the feature's built-in Report a problem function (pictured) and How to report issues with a suggestion please post any patterns you're noticing on User talk:Alaexis/AI Source Verification/Suggestions. PPelberg (WMF) (talk) 00:33, 5 November 2026 (UTC)
I just did a handful to check that this is working, and it worked as expected. These two were interesting because they helped me fix minor issues in the article text: List of fictional ships (fixed a copy-paste mistake), Beef clod (removed a questionable claim). Dreamyshade (talk) 09:08, 5 November 2026 (UTC)
Great to hear that, thanks for the feedback! Alaexis¿question? 09:57, 5 November 2026 (UTC)
Not a problem with the tool per se, but should we change the prompt for the lead. A frequent mistake is that the lead summarises the body, and someone places the wrong citation for a claim supported later. We want folks to check that first before removing the claim. In solidarity, —Femke (talk) 🐦 17:41, 5 November 2026 (UTC)
@Femke to confirm: would it be accurate for me to think the following?
1. You think it makes sense for the suggestion to ignore (read: not evaluate) article lead sections
2. You think this because, based on what you've observed, people place citations within the lead that do not support their associated claims even tho said claims are properly cited/source elsewhere in the article PPelberg (WMF) (talk) 23:44, 5 November 2026 (UTC)
It is a question of what the "associated claims" are defined as. A citation in the lead may specifically only cover say the word immediately before it, rather than the whole sentence. (This is not the way citations are expected to work in the body.) CMD (talk) 23:47, 5 November 2026 (UTC)
@CMD, mmm. Okay. Got it. Thank you for explaining that. So you're aware: if you all were to decide the suggestion ought to ignore lead sections, you could make that happen by setting the on-wiki config for the source verification suggestion to ignoreLeadSection = true. PPelberg (WMF) (talk) 00:06, 6 November 2026 (UTC)
That is very useful, I personally would prefer tasks were shifted towards the body rather than the often given more attention leads, but moving off-topic from this particular use case with a specifically technical reason to avoid leads. CMD (talk) 00:09, 6 November 2026 (UTC)
The special rules for citations in lead sections are at MOS:LEADCITE. Most claims in a lead don't need to be supported by a citation if the claim is supported in the body. So if a claim in a lead is only partially supported by a citation (for example, Femke may have run into row 4014 in the spreadsheet), that may not be a real problem. But if a sentence or clause in a lead is contradicted by the citation, or if the citation doesn't support anything in the sentence or clause, that should be checked, but the editor should also look at the material and citations in the body while checking it, instead of just removing it. Summary: it's complicated, and I agree with skipping leads at least for the first iteration. Dreamyshade (talk) 00:16, 6 November 2026 (UTC)
Some more complications here. Eventually, we do want an adjusted lead check, I imagine, as errors in the lead are the most consequential.
If there is a quote, it needs to be supported for instance. But what needs support is messy in a way that experienced editors something make mistakes in, and would likely differ significantly by language edition. So turning it off temporarily makes sense. —Femke (talk) 🐦 08:31, 6 November 2026 (UTC)
@Dreamyshade,@Sadads, @Femke, totally agree with your points, the leads demand special treatment. I've tried to implement it in the user script first and it was surprisingly non-trivial. There are many possible combinations since leads also include inline citations, so the claims should be checked against the body and against cited sources. The challenge is not just to come up with an algorithm but also not to confuse the user of the script. Hopefully I'll release a new version of Source Verifier (not edit suggestions) and I'd be grateful for testing and feedback. Alaexis¿question? 11:30, 6 November 2026 (UTC)
If it's helpful to have an example of the tool running into a LEADCITE issue, you can see this one I've reported. Sdkbtalk 16:35, 9 November 2026 (UTC)
@Dreamyshade, @Sdkb, @Femke, @Sadads, please try the latest version of the userscript which checks lead against body and lists unsupported lead claims separately. The UI will probably change soon, it's now too crowded for my liking, but the logic should be fine. Let me know how it works. Alaexis¿question? 22:19, 9 November 2026 (UTC)
I had to fiddle with the relationship to references a lot in the design of: User:Sadads/Lead_Balancer -- one of the steps actually does a "lead to section comparison and checks if there is corresponding content elsewhere in the article. According to my Claude (which built most of it): "Lead-to-section mapping (no AI). coverage.js:112-147 scores each lead sentence against each body section with tf-idf cosine similarity. Words found in every section, and the subject’s own words, are ignored. A sentence only maps to a section if its shared words add up to real evidence, such as a heading word, a word unique to that section, or two specific words. Otherwise it stays unmapped (sectionIndex: null). This is a similarity heuristic, not a check for citation support." @PPelberg (WMF) -- you could do a word similarity hueristic for checking if the lead sentence without a citation, is well covered by cited material elsewhere in the article, even if its not a 1:1 match. Sadads (talk) 00:12, 6 November 2026 (UTC)
@PPelberg (WMF) I will send you the recommendation from Claude on the hueristic for lead section citations, which could be implemented as a slightly different flag for uncited lead sentences, Sadads (talk) 00:18, 6 November 2026 (UTC)
With permission, sharing the results of the types of error found over summer in a sample of climate change related articles. The overall error rate was the same at about 70%. Some of these are discussed above already.
Error Type
Description
Frequency
Partial support error
The tool is not correctly describing labelling claims where partial support is present
38
Website not found
URL does not link to a currently active website
30
Book
The source is a book and so is not readable for this experiment
23
Omission
The AI claims information is not in source, when it is
19
Graphical
AI tool failing to read information that is in a graphical format
14
Implication
The AI does not infer a logical implication that a human could reasonably understand
12
Context
The AI is not properly reading relevant context to an article
11
Structural Error
The AI tool is being confused by non-standard citation/article structure
7
Paywall
Source is paywalled
6
Contradiction
The source contains information that contradicts the AI's interpretation
5
Source error
Source is not found or otherwise unreadable
4
Landing page
The citation links to a landing page rather than the source itself, thus confusing the tool
2
I've not checked, but I imagine that when the source is unavailable, the SuggestionMode already ignores it. Some errors may be fixable by the prompt.
I think it was context where the AI often doesn't read the date the article was published, so that sentences like "Yesterday, the prime minister said X" translated to "On day X, the prime minister said X" would be flagged as incorrect.
The structural error, e.g. citations around tables or in front of bullet points, are another one where a prompt can help. At least, in front of bullet points (where we put the citation before the text in general), may be easy to include in the prompt or exclude from the SuggestionMode selection.
We can see captions, and can perhaps prompt that statements that seem to be present in the graphs should not be flagged as incorrect. Not all captions start nicely with "Figure 1" however.
Fixing the 'implication' (e.g. simple maths) and contradiction may simply require us to use a better model.
The biggest error is interesting, as the tool, in its text, correctly says that something is partially supported, but the verdict differs from the explanation. I imagine such an almost-hit again would simply require a slightly smarter model. —Femke (talk) 🐦 09:15, 15 November 2026 (UTC)
Thanks, that's very helpful. Speaking of smarter models, have you been using the default model all the time? Alaexis¿question? 10:21, 15 November 2026 (UTC)
I have yes. The others require subscriptions, don't they? —Femke (talk) 🐦 10:25, 15 November 2026 (UTC)
Mostly yes. There are two free options now, Qwen hosted on Liftwing (default) and GPT-OSS hosted on HuggingFace (used to be the default until not long ago), so just wanted to make sure you didn't change the model in the settings. Alaexis¿question? 10:30, 15 November 2026 (UTC)
@Alaexis, do you have a sense of which models are the most useful for this at present? In solidarity, asilvering (talk) 19:30, 15 November 2026 (UTC)
@Asilvering, last time I checked, Gemini 3.7 was the best but the two free options (Qwen 3.6 and GPT-OSS-20B) weren't that far behind. The most useful metric is "Supported-vs-rest (strict)" - this is the% of cases in our benchmark dataset for which the model's binary verdict (SUPPORTED VS EVERYTHING ELSE) is correct, you can see the numbers here. I checked Qwen 3.8 when it became available on LW and found it to be slightly better than 3.6. Note that the benchmarking dataset is not representative (on purpose), so these numbers are only for comparing different models. Alaexis¿question? 20:48, 15 November 2026 (UTC)
How easy is it to retest the accuracy upon a changed prompt? I wonder if we can change the following (suggestions are where claude, chatgpt and gemini sort of agreed)
Make the output json give the source_quote first, then the number, then the verdict. That way, we might prevent the number 1 error of the verdict being wrong but the reasoning right. It may force the model to reason before deciding. The description of the source_quote would change to "Quote the passage that comes closest to supporting or contradicting the claim"
Give more examples of things being right, e.g. very simple implications
I'm less sure about this, but Claude said the prompt for unavailable sources might be overly broad. No text to quote might be interpreted as no relevant text to quote (e.g. an unrelated source), instead of no text available.
Good ideas, definitely worth checking. It's not too hard, I'm now testing your first suggestion. Feel free to fork the repo and try your other ideas yourself (the benchmark dataset is here and the readme is , and the LiftWing and Hugging Face models are free, so no API key is needed). Alaexis¿question? 16:09, 16 November 2026 (UTC)
How about a few confusion matrix displays? Ideally a top matrix for summary with 3 or 4 below it. I suggest avoiding complex representations so a quick view can be possible. Yesterday, all my dreams... (talk) 09:39, 16 November 2026 (UTC)
WMF's amicus brief in In re: OpenAI, Inc. Copyright Infringement Litigation
The Wikimedia Foundation filed a motion to submit an amicus brief (Commons link) in the consolidated In re: OpenAI, Inc. Copyright Infringement Litigation case (including e.g. The New York Times v. Microsoft and OpenAI). There is also a summary on Diff.
My read (IANAL) is that the WMF's position boils down to:
Scraping is about access and Terms of Use, and is not a "use" in the sense of "fair use".
Using data to pre-train a "general purpose" model is fair use.
Retrieval at query time is not fair use.
And now here's my take on these positions, and some implications (again, IANAL):
The claims of the publishers that tie copyright to acquisition fail, but server load is a consideration. This is already more or less the WMF's de facto policy.
Disagrees with the New York Times' (and others') core claim. This is the surprising position in my view, both because of the claim itself and even more so because taking a position on this matter just doesn't seem like it was necessary. I would be curious to hear more about why the WMF did so. The Diff post says it's to support smaller-scale/independent AI development. We do rely on machine learning for a variety of maintenance tools, as the brief's conclusion talks about, but there doesn't appear to be any attempt in the brief to distinguish between those small/open developers and the largest corporations that stand to profit the most. #3 claws back some of the traffic lost to AI interfaces, but #2 concedes the rest of it is ok. More importantly, this means prioritizing developers above the journalists and publishers that we depend on for our citations. It's not an easy needle to thread, but it also didn't seem like this particular needle needed threading right now.
This is the consequential position for us, because a chatbot would be able to retrieve Wikipedia content, but would need to provide attribution or a backlink. That would be a Very Good Thing, though I have no idea what the effects of ShareAlike would be within a chatbot answer. For the non-free stakeholders, this would mean retrieval needs a license, which is good for them. Unclear what it would do to retrieval practices in general. Overall, this seems worth the brief itself.
Almost certainly, the third point was the main point, but that leaves the question of what purpose the other two serve. Perhaps the combination reinforces Wikimedia Enterprise? Do the first two make it easier to take the third position? Is there also a strategic-pragmatic move here, knowing that the claim that training is fair use is almost certain to be held up in the long run (given how much of the US economy could be argued to depend on it right now), so conceding that to gain a better position for #3?
I hope none of this sounds cynical, and I hope I haven't misrepresented anything. I found #2 perplexing, but I have a lot of faith and respect in WMF Legal, who I concede may be playing chess while I'm playing checkers. Just hoping for more context. —Rhododendritestalk \\ 19:16, 30 October 2026 (UTC)
I'm also concerned about #2. As I understand it, the thinking is this: if no one can train open source models because a) models trained on copyrighted material cannot be used to generate free material and/or b) it is extremely expensive to acquire training data and small-scale developers can't pay for it, the only game in town will be the big AI companies with the ability to pay, and this would be bad for the free culture movement broadly speaking. In solidarity, asilvering (talk) 21:16, 30 October 2026 (UTC)
This is pure speculation, based on nothing but gut feeling, but one possibility about #2 is that WMF is afraid that if training an AI model on copyrighted material is viewed as copyright infringement, writing a Wikipedia article with sources that happen to be copyrighted would be infringement too. There not being a light-and-day difference between "AI training" and "writing adequately sourced Wikipedia articles". Jo-Jo Eumerus (talk) 13:19, 31 October 2026 (UTC)
The crux of the copyright complaint for the training part of the case has to do with making copies of the copyrighted content first, even if the output isn't a direct violation. Still, that's not a bad take - that the WMF could be taking principled opposition to a copyright/ownership claim that's concerningly broad, given the work we do? —Rhododendritestalk \\ 13:46, 31 October 2026 (UTC)
Makes sense, if it were established that taking information from a copyrighted source was enough to make a copyright infringement, this would seriously imperil Wikipedia's content. Although I don't get the impression that "copies of the copyrighted content first" are the issue per se, at least per The New York Times v. Microsoft and OpenAI? More when (parts of) these temporary copies end up in the reply to a prompt, although from the way the complaint about competition is worded, it may extend to replies that don't meet substantial similarity but share the same information. Jo-Jo Eumerus (talk) 14:06, 31 October 2026 (UTC)
FWIW the replies that contain obvious infringement are the smoking gun, so are very prominent in coverage of the case. They are indeed the subject of copyright claims themselves, but they're also direct support for the much broader claim that training involves making/keeping copies. —Rhododendritestalk \\ 14:26, 31 October 2026 (UTC)
As an experiment, I asked Claude to write an article for me. The prompt I used was:
Write me a Wikipedia article about the SoHo Weekly News, which was an alternative newspaper published from 1973 to 1982 in New York. The article should be about 5000 words long and comply with the Wikipedia MOS (Manual of Style).
I used SWN as a test subject because I had written on the topic myself (shameless self promotion: SoHo Weekly News is going to be on WP:TFA in about a week). I'm not sure what I expected, but the result was kind of shocking. I'll have the WP:AINB folks breathing down my neck if I post it on-wiki, so I tossed a PDF into Google Drive. It's pretty much the poster child for WP:CLOP if not straight-out plagiarism. This may be an extreme case, however. SWN is a niche topic, so there's not a lot of long-form content out there about the paper. My article probably swamps the training corpora on the subject, so I guess it's not surprising that all Claude could do is regurgitate it back.
I get what the NY Times is upset about. The product they sell is partly their investigative reporting and partly their writing. If the LLMs ingest NYT material and spit it back out in the same words (similar to what Claude did with my SWN article), that's clearly copyright infringement.
The deeper question is what if the LLMs ingest a NYT article, understand the information it contains (leaving aside for the moment that LLMs don't really "understand" things in the way we typically use the term) and create entirely new text based on their understanding. That's exactly what I did with my sources when I wrote my SWN article. So, not copyright infringement, but possibly a violation of the NYT Terms of Service which says "The contents of the Services, including the Site, are intended for your personal, non-commercial use". And later on, "Non-commercial use does not include ... the development of any software program, model, algorithm, or other generative AI tool ..." It's really hard for me to see how what the LLMs are doing is not in violation of that. I maintain a paid subscription to the NYT. It's a contract. I give the NYT money. In return, the NYT provides me with certain services, and imposes certain restrictions on how I may use those services. If I don't like the restrictions, I'm free to cancel my subscription and stop using the services.
While I'm certainly sympathetic to the WMF's claim of "fair use", I don't see how that overrides the limitations imposed by, for example, the NYT TOS. I'm also torn between wanting the WMF to be right and wanting the NYT to be right. For sure, I'm all about "information should be free", and I'm certainly preaching to the choir on that one. But I also value the information I get from the NYT (and other papers to which I maintain paid subscriptions). If the NYT is ultimately unable to maintain profitability, they will go out of business and I will lose access to a service I value. I don't want to see that happen. RoySmith(talk) 15:50, 31 October 2026 (UTC)
That sounds like an issue with ToS violations (->contract law?) and not with copyright law, though? Even if it was found that OpenAI violated the NYT TOS, that would not imply that this violation constitutes copyright infringement. This lawsuit seems to be about copyright issues not ToS. JoJo Eumerus mobile (main talk) 16:23, 31 October 2026 (UTC)
In my (non-legal advice) understanding, you are correct that fair use concerns do not override agreed-upon contract terms. As JoJo Eumerus says, the two are separate concerns. If the companies obtained their copies of the copyrighted work without entering into a contract with the publisher, though, then only the copyright concerns are applicable. isaacl (talk) 16:46, 31 October 2026 (UTC)
One additional complication is the way many corporations interpret "agreed-upon contract terms". By clicking on any link to this web page you agree to... We can change this agreement at any time without notice...
There is a refrigerator maker that is trying to enforce a "by opening this box you agree to binding arbitration" notice when the box is removed by the installers and thrown back in the truck before they wheel the fridge into your kitchen. --Guy Macon (talk) 01:57, 1 November 2026 (UTC)
I've read the brief. Here are my personal comments: I'm feeling Wikimedia Foundation is standing on the wrong side and it makes me doubt whether I should contribute to Wikipedia again.
WMF argued for a blanket fair use for AI pre-training, but they missed many technical facts. First, the current AI models are dominated by a few Big Tech (OpenAI, Anthropic, Google, Meta, etc.) because the main cost of AI training is the computing power and hardware, not licensing. Making the AI training fair use doesn't benefit smaller companies, especially those who already negotiated licenses for their training data. Second, WMF would be allowing AI trained in order to create a direct competitor to exist (there is one already, Grokipedia). I take no position on whether Grokipedia is allowed to rival Wikipedia while trained on Wikipedia content, but some other Wikipedia editors might concern the ethics of that practice. Let's see.
Third. Wikipedia content is licensed under CC-BY-SA. CC licenses implicitly allow AI training (disclaimer: IANAL) although the output might be constrained to require attribution (and, for share-alike licenses, share-alike) if it looks so much like a derivative. Fair use for AI training won't help Wikimedia's mission for a free and accessible knowledge at all. --Explorer09 (talk) 02:23, 3 November 2026 (UTC)
I think I would have a few questions there. First, do OpenAI and the other chatbots respect restrictions such as robots.txt or other metadata restrictions? If they don't, they're actually accessing a computer system when they're not authorized to, and under US federal law, that can be a felony. I don't think just reading something, whether it's read by a chatbot or a person, constitutes copyright infringement, but it could violate other laws if they're just slurping things up in violation of the TOS of sites they visit. And if they're doing close paraphrasing or directly reproducing them for their users, well, that could certainly be infringing. SeraphimbladeTalk to me 00:08, 5 November 2026 (UTC)
Just on a couple points: robots.txt is a purely voluntary convention, and there's ample evidence of crawlers ignoring it. And hiQ Labs v LinkedIn means claims based on scraping publicly accessible data need to rely on something other than CFAA. —Rhododendritestalk \\ 03:02, 5 November 2026 (UTC)
Training a model seems a textbook example of a transformative use which should be fair use; it is also a clear parallel to learning and being creatively influenced by what we read and hear and experience. Points 1 and 2 seem like the right position, legally and culturally, for a knowledge society. Point 3 seems overstated in the brief -- retrieval at query time for confirmation or sourcing, is also definitely a fair use; as are short quotes of a length we would allow in an article (i.e. the volcanic cloud was described by <SOURCE> as "most likely the highest in over a century"). Any use at query time that would qualify as fair use if I was adding a paragraph to an article should be fiar use in the output of a model.
I would have liked this brief to more carefully frame the types of reuse and derivative works that fall outside fair use, and to separately note that norms around attribution need to adapt: fair uses of source materials already need to be attributed so as not to be plagiarism, and that norm should be strengthened and extended to cover AI outputs that may draw on a panoply of sources.
Overall I'm glad to see WMF engaging with these public discussions but we need a clearer articulation of the abject failures of the copyright regime, and how we need something better to further our society's cultural and educational goals (freedom of panorama and no gatekeeping of public domain works by intermediaries, at a minimum). Without that, when we take positions that will be warmly embraced by the (c) lobby, they can be used in the name of Wikimedia to shrink and enclose standard affordances for sharing knowledge about the world: including fair use, first sale, and the public domain. –SJ+ 18:43, 9 November 2026 (UTC)
I wish WMF would not just walk into a legal battle zone to which they had not been invited. It costs money, money, money. The only winner will be the lawfirm of Dewey, Cheatem & Howe whoever they may be representing. Yesterday, all my dreams... (talk) 09:37, 11 November 2026 (UTC)
My position disagrees with SJ arguing that training is transformative use and thus is "fair". My position is closer to US Copyright Office's position on fair use: that whether the AI training is fair use depends on the ultimate purpose of that AI model, including how the end users would use it for. There's no blanket fair use for AI training, and AI training is not always unfair either. It depends on how the AI is further used for.
The closest example for Wikimedia Foundation on how an AI could be an unfair use is Grokipedia, since Grokipedia is intended to compete with Wikipedia, while trained on Wikipedia data. It's just fortunate that WMF and Wikipedia are not-for-profit, and are not financially stressed to sue Grokipedia, but trying to defend that as fair is overstepping the foundation's purpose. I really suggest WMF should stay out of this and not pretend they are smart.
Wikipedia position is also impractical in another manner: When the AI can regurgitate copyrighted works, especially poems and lyrics, such an output infringement should make the AI companies liable instead of pretend it never happens or shift all the responsibility to end users. Such an output infringement is an infringment, and is an important factor on judging the "fair use" of AI training. In other words, AI can contribute to infringement through regurgitation. And to stop this contributory infringement from AI, denying fair use of AI training is the only way. Really, because any post-training AI safeguard can fail. Explorer09 (talk) 13:44, 14 November 2026 (UTC)
I am no legal expert, but I am somewhat sure that "And to stop this contributory infringement from AI, denying fair use of AI training is the only way" isn't how laws usually work. Only the AI output that is copyright infringement is actually copyright infringement. Otherwise, any person reading a book would be committing copyright infringement because of the mere possibility that they might remember enough of its text to write down a copy that violates copyright. Like, for example, by writing a WP:CLOP Wikipedia article. This is a point that the article you cite mentions: Contributory copyright infringement#Sony Betamax case and the Grokster case. As for the status of an AI that output a copyright infringement, I am not actually sure that it's settled law and jurisprudence how to handle these cases. I don't think that "the AI company is liable" is an unreasonable approach, but well, is it the approach sanctioned by laws & courts? It makes sense that the WMF would take a position here, lest the cookie crumble in a way that is bad for the encyclopedias. Jo-Jo Eumerus (talk) 08:27, 15 November 2026 (UTC)
AI regurgitation is different from the Sony Betamax case that AI companies usually cite. With the Sony Betamax case, the tech is "neutral" - the VCR only copies copyrighted expressions if the end user feeds in that expression. It's not like Cox v. Sony either as the ISPs are also "neutral". Gen AIs, however, can suggest copyrighted expressions without an explicit prompt from the user (it's demonstrated before, e.g. in this article). Such an action can be ruled by a court as a form of "inducement", thus making the AI companies liable. How can you say that a tech can be "fair use" on training while it can induce infringement for downstream uses? Many people would find that ridiculous, and I guess the judges will find that ridiculous as well. Explorer09 (talk) 13:22, 15 November 2026 (UTC)
Whether and how frequent the AI can reproduce copyrighted materials without the user's explicit prompt is a fact based analysis and I should leave that question to the court to decide case-by-case. You didn't get the difference between Cox v. Sony and this In re OpenAI Copyright Litigation case yet. It's this: ISPs don't actively suggest what copyrighted material should be presented to user when they send HTTP requests. ISPs are passive entities that only forward user's requests and forward the websites' responses, no matter whether the content is infringing. If a generative AI only does that (passively forwarding requests instead of actively generating infringing materials), then even I would consider that a fair use. Unfortunately I've read the NYT v. OpenAIcomplaint and New York Times didn't just allege that. Explorer09 (talk) 14:47, 15 November 2026 (UTC)
See, that sounds like a nitpicky difference that is unlikely to make a difference in court. In other contributory copyright infringement cases, courts inferred intent and effect from a totality of indicators, not just one point. Moreover, if you (or NYT) are citing examples of individual outputs, the "And to stop this contributory infringement from AI, denying fair use of AI training is the only way" isn't how laws usually work problem comes back. Jo-Jo Eumerus (talk) 17:02, 15 November 2026 (UTC)
You didn't present any counterargument about your "isn't how laws usually work" part. Explorer09 (talk) 02:04, 17 November 2026 (UTC)
What you describe as "not how laws usually work" is absolutely how laws usually work. It is the sort of stupid shit that lawmakers and courts do do all the time. You go on to describe how the proposed legal remedy doesn't make sense and will have unintended consequences, which are attributes found in many, many laws. See ] for a humorous list containing many examples. --Guy Macon (talk) 03:33, 17 November 2026 (UTC)
Yes, but the inherent indeterminacy of court cases is also well known. Please do a search "what the judge had for breakfast". Yesterday, all my dreams... (talk) 11:55, 18 November 2026 (UTC)
Change to Wikimedia log-in security
I was recently locked out of my account for a week because the email address associated with my account had expired.
I added an email address 14 years ago, and it was long dead.
There was no warning..."hey, add an email to your account!"...just a lock out.
The security feature for logging-in needs to be changed so there is a warning to logged-in users that a working email needs to be associated with their account. Magnolia677 (talk) 15:14, 1 November 2026 (UTC)
Thanks for raising this here, and I'm glad you were able to recover your account. We probably should consider something that asks users to reconfirm their existing email and makes clear that it could matter, especially for users like yourself whose emails were added long ago. We added a banner for unconfirmed emails earlier this year, but not for confirmed-a-long-time-ago emails.
As a side note, you can avoid email checks by enabling 2FA - if you then also add a passkey, you can do passwordless login, which is the easiest thing we offer. EMill-WMF (talk) 21:46, 1 November 2026 (UTC)
I was literally locked out because my email had long expired. I spent the week volunteering at my church, spending time with family and friends, exercising, and reading classical novels. It was a horrible week! I'm hoping a warning can be sent out so others don't have to suffer like this. Thanks! Magnolia677 (talk) 22:23, 1 November 2026 (UTC)
You can prevent this by not registering an e-mail address. That has several disadvantages but, if the powers that be stubbornly insist on providing no other way to avoid excessive security, it remains a useful tool. Certes (talk) 09:32, 2 November 2026 (UTC)
@Certes: We are "the powers that be". If we send a message that the sign-in policy needs changing, then it will change. All that's needed is the following: "you will be locked out in a week if you don't add a working email to your account". Easy peezy. Magnolia677 (talk) 11:21, 2 November 2026 (UTC)
"There is always a well-known solution to every human problem—neat, plausible, and wrong." -H. L. Mencken
The above is an obvious wrong solution. --Guy Macon (talk) 17:25, 2 November 2026 (UTC)
H. L. Mencken was a racist and antisemite. Who cares what he thinks. The sign-in policy needs to be corrected. Perhaps a message sent to all registered users talk pages, or a one-week warning before lock-out. Magnolia677 (talk) 19:13, 2 November 2026 (UTC)
You might benefit from studying our article on Genetic fallacy. That, and perhaps some actual evidence or argument other than "Something needs to be done. This is something. Therefor this needs to be done"[CitationNeeded] --Guy Macon (talk) 20:06, 2 November 2026 (UTC)
I don't care about Mencken or genetic fallacy or any other useless off-topic crap. I care about people being locked out of their account without warning, and finding a workable solution so they are about to access their account. Magnolia677 (talk) 20:59, 2 November 2026 (UTC)
if you want to play that card... your quote is not an argument, just an opinion. It's an argument from authority, and there is nothing in it to be dismissed or disputed as no argument has been presented why it would apply here. "I have this solution" - "Mencken says that there is always a wrong solution". Yeah, so what? Try posting some actual arguments based on the actual discussion, instead of lecturing people on what gems from your wisdom might benefit them. Fram (talk) 15:46, 4 November 2026 (UTC)
Out of curiosity, is this a message to all users or a message at the log in page, Magnolia677? --Super Goku V (talk) 17:11, 4 November 2026 (UTC)
@Super Goku V: It was a message at log in. I entered my username and password...all good...then got a pop-up saying I needed to check my email for a log-in code. What?? I hadn't used the email associated with my Wiki account in 10 years, and it was long dead. So I had to fill out a form and wait a week. No warning on my talk page like..."better have a working email or you'll soon be locked out". Bupkis. It could happen to anyone. Very traumatic. Magnolia677 (talk) 19:28, 4 November 2026 (UTC)
I was just referring to your suggestion. Is your suggestion to have a message sent to all users to their talk page or a message on the log-in page? --Super Goku V (talk) 19:35, 4 November 2026 (UTC)
At log-in would be most attention getting, along with a scary-looking icon. Thanks for your help with this. Magnolia677 (talk) 19:51, 4 November 2026 (UTC)
Hacked
I am seeing claims we have been hacked, is this true? Slatersteven (talk) 15:20, 4 November 2026 (UTC)
There's was some vandalism earlier, which some types of people sometimes call hacking. Is that consistent with what you're seeing? -- zzuuzz(talk) 15:34, 4 November 2026 (UTC)
Not sure I have seen claims (including screenshots) of whole pages being replaced with some hacker's message (and yes, I know that might not mean hacking, just a crapy edit). Slatersteven (talk) 16:04, 4 November 2026 (UTC)
It is/was just another vandal. -- zzuuzz(talk) 16:11, 4 November 2026 (UTC)
Editors interested in this project may also be interested in multiple discussions going on now about funding the LGBT+ User Group for the next three years. As I was not aware of this group until recently, and not very familiar with how WMF funding works either, I put together the following intro to help me, as the linked discussion page addresses a topic I was pretty unfamiliar with, and I needed some background to understand the underlying WMF structure and backstory. Maybe it will help you, too.
Wikimedia LGBT+ is a "Wikimedia user group that promotes the development of content on Wikimedia projects which is of interest to LGBT+ communities and supports LGBT+ individuals who engage in Wikimedia activities". It is a recognized Wikimedia affiliate, in this case, a Wikimedia User Group, since 2014. There are 143 user groups, overwhelmingly local groups for a given geographical area (country, region, city), but there are also a few user groups for common interests, such as Art and Feminism, Classics, or Education. The LGBT+ User Group is among the latter. (There are also affiliates that are not user groups, like chapters, but that is o/t for here.)
Among other things, WMF affiliates, including user groups, can apply for funding from the WMF to support their goals and activities. The WMF has several models of funding, such as the Rapid Fund, the General Support Fund, the Multiyear Funding program, and others. The LGBT+ User Group has received funding from the WMF in the past, and has applied for a continuing grant to support their activities for the next three years; this grant proposal is entitled, "Sustaining Wikimedia LGBT+ 2026–2029".
Now we come to the point of this message. This grant proposal has stimulated some lively discussion on the grant proposal talk page. There is a multiplicity of viewpoints, and I found it tough going, given my relative unfamiliarity with the backstory (it's an iterative process for me), but I feel it's important to understand it as it involves questions about who or what should be funded by the WMF to promote the goals of supporting LGBT+ content and community. In an effort to help me follow the conversation thus far, I asked a chat bot to summarize the main viewpoints, and this was the response; all the usual caveats about AI errors, hallucinations, and misinformation apply.
I'm hoping that someone will provide a quick summary of what grant applications are current. Of course we all want people to be nice and to write good articles, but I'm astonished that donations to Wikipedia might be used to fund such basic activity. Anyone with evidence of harassment can get the harasser permabanned by emailing either the WMF or the Arbitration Committee. Johnuniq (talk) 08:19, 6 November 2026 (UTC)
Security update: Many Wikimedia tools were taken offline because of a security vulnerability in PHP. This affected many websites worldwide. Toolforge and Cloud VPS have been back up since October 22. Many important tools have been gradually restored. Some tools need the respective volunteer maintainers to restart them. Follow the work on Wikitech.
Wikipedia Year in Review 2026: This year's Wikipedia Apps Year in Review is in progress and you can help make it available in more languages to readers around the world by helping with translations. Here's how:
Around the Globe in ESEAP: Connecting 32 geographic areas and 64 language editions of Wikipedia across East Asia, Southeast Asia, and the Pacific. See how the region is creating spaces for active editors to build relationships.
Notifications Experiment: In the coming weeks, the first of two notifications experiments on Android will be launched. The experiment will test whether personalized Discover reading list notifications will encourage them to return to the app. The three-week experiment will include English Wikipedia readers who have opted in to notifications. Learn more and share your feedback on the project page.
Reading recommendations: An ongoing experiment is testing article recommendations for new account holders who also identify as readers. This aims to encourage new account holders to read and return to the website regularly, with the goal of increasing internal referrals by 1%. See all live, upcoming, and completed experiments in Product & Technology.
New search experiment: The first of four experiments in Phase 2 of Semantic Search is running in the Wikipedia Android and iOS apps for Arabic, French, and Japanese Wikipedias. Participating readers can search inside articles and open quoted passages directly in their source articles, no answers or summaries are generated. Feedback is welcome on the project’s discussion page.
Share Cards: The Future Audiences team is planning to bring Share Cards to Wikipedia on mobile. This feature gives users the option to share a visually appealing card with friends, based on the article they are reading. The card contains the article’s title, image, opening text, and related attribution, and links back to the article for full reading.
Tech News: The latest highlights from Tech News weeks 44 and 45 include the default renderer for math formulas has changed. See also the 59 community-submitted tasks that were resolved over the last two weeks.
New Toolforge features: As part of Tools Platform's work to improve the reliability of Toolforge-hosted tools, two new platform features are launched: (1) Toolforge alerts let tool maintainers receive email notifications when any of their tools is encountering a common error condition; (2) Toolforge extended logging capabilities let tool maintainers retrieve logs for events like job creation, deletion or deployment failure, with a new CLI toolforge logs, at the same time that they retrieve their job logs.
Work also continues to make sure Citoid does not get blocked by publishers who are starting to block bot traffic to manage the impact of scraping. In September, we worked with Reuters and the Library of Congress to allowlist Citoid. You can track our progress with other publishers on Phabricator.
For information about the Bulletin and to read previous editions, see the project page on Meta-Wiki. If you have feedback or suggestions about the bulletin, let us know at foundationbulletin@wikimedia.org. For questions about the Wikimedia Foundation's work, contact us!
Awareness campaign about the importance of Wikipedia in the age of generative AI
The number of human visits to Wikipedia is decreasing because of the use of generative AI to look for information. While Wikipedia is not a business and its main goal is not to maximize the number of visits, it's important that the public doesn't see generative AI as a replacement for Wikipedia (it may be a replacement for searching very basic information about simple facts, but it's never a replacement for Wikipedia as a whole). This is not only for the benefit of the Wikimedia movement and WMF, but also for the good of the wider public. I think a public awareness campaign by WMF would be very helpful here. While the exact slogans will be better left to experts in communication, I think that sentences such as The largest work ever written and By humans, for humans would be good ones for use in the campaign. The campaign should be able to communicate ideas such as the ones expressed in this essay, and, probably, additional ones. MGeog2022 (talk) 13:47, 18 November 2026 (UTC)
WMF has already started running with a "Knowledge built by humans" slogan in some campaigns, see eg. Facebook, Diff. CMD (talk) 14:41, 18 November 2026 (UTC)
An interesting question is what message the WMF sends in its most effective method for reaching the general public; annoying obtrusive banner ads that they hide from logged in users (see eating your own dog food).
They could use the banner ads (or some percentage of them) to attract new editors, or they could make it all about donating more money. The first addresses a serious problem the Wikipedia has, while the second supports them being able to fund pet projects and grateful grant recipients. --Guy Macon (talk) 15:20, 18 November 2026 (UTC)