Page MenuHomePhabricator

[SPIKE] What would it take to enable Recent Changes patrolling on all Wikipedias?
Open, Needs TriagePublic

Description

As we think about building new patrolling experiences for Wikipedia editors, it has become apparent that it would be useful to know if a patroller has already looked at an edit and deemed it acceptable. As we're aiming to surface edits needing patrolling, in general we want to avoid showing users known good edits.

Wikis can use the $wgUseRCPatrol configuration (Patrolled edits) to add per-edit patrolling capabilities, or use the FlaggedRevs extension. On wikis not using RCPatrol or FlaggedRevisions there is generally no easily accessible signal that a given edit has been reviewed already. It's not desirable to roll out FlaggedRevs to more wikis, but RCPatrol is a core MediaWiki feature which would be very easy and lightweight to turn on for more wikis.

There has been more than one discussion on at least English Wikipedia, through which editors have expressed that they do not want per-edit patrolling. We want to review these discussions and see if there are code or configuration changes we could make that would make rolling this feature out more acceptable, and assess prioritising those changes.

How RCPatrol works

  • On Special:RecentChanges, Special:NewPages, and Special:NewFiles, users with the patrol user right see a red exclamation mark next to edits which have not been marked as patrolled.
  • Users with the patrol user right can mark edits as patrolled.
  • Users with the autopatrol user right have their edits marked as patrolled automatically.
  • French Wikipedia automatically grants autopatrol to users who have made more than 500 edits and created their account 90+ days ago.
  • By default, bots and sysops have the autopatrol user right, and sysops have the patrol user right.
  • Wikis can configure patrol of individual edits, patrol of new pages, and patrol of new files, separately. English Wikipedia, for example, uses the new page patrol function.
  • Edit patrol status is only tracked for 30 days.

RCPatrol is enabled on 32 Wikipedias - the biggest Wikipedias using it are French, Italian, Portuguese, Persian, and Dutch. It is also deployed to Wikidata, Wikifunctions, and Commons. It was deployed to English Wikipedia (and others?) in 2005, but a discussion at the time was overwhelmingly negative about the feature, and it was disabled a few days later.

Rationale/benefits

  • Moderators save time by only needing to review edits which are likely to still require human review.
  • When software invites newer editors to review edits, we can be more confident that their time will not be wasted reviewing known-good edits.
  • New editors may find it motivational to learn that their edit has been reviewed by an experienced editor (T410882).
  • Existing tools could use a central definition of 'patrolled' rather than inventing and storing their own data on which edits have already been reviewed.

Prior discussions

Potential issues to address

  • On English Wikipedia, some users highlighted that patrol and autopatrol rights have been granted in the context of pages, not edits, and it wouldn't be desirable to have all new page patrollers have all their edits autopatrolled without further review.
    • We could split the patrol user right for new pages from new edits (and new files) - T308153
  • General concerns that this feature does not add any benefits, because existing patrolling workflows already catch bad edits, and could lead to obscuring bad edits if they're erroneously marked as patrolled.
    • One data point - Huggle, which a lot of patrolling on enwiki is done through, has a 'Good edit' button, which is analogous to patrolling an edit, except that it adds to a score for the user, and their edits are not displayed to users after so many 'Good edit' flags, not immediately from one.
    • On English Wikipedia 39% of vandalism patrolling is done with the assistance of tools, many of which are already doing this kind of filtering (T348862)
    • We could add optional configuration to hide all user facing elements of this system and only use it for any new patrolling tools we build. In this way all existing processes continue to work as they currently do, but we get the benefit of an existing patrol status system.
  • Editors should double check each other's work - one user marking an edit as patrolled shouldn't hide it from other editors.
    • Nothing is hidden by default, users can optionally filter out patrolled edits in RecentChanges. That said, we could imagine an optional config that requires X patrols before an edit is considered 'patrolled'.
    • More broadly, perhaps patrolling shouldn't be a binary 'patrolled' / 'not patrolled' classification. We could be storing more granular data so tools and users can decide what level of patrolling they want to filter out. e.g. only edits patrolled by an admin?
  • Exclamation point to flag unpatrolled changes is poorly designed.
  • RCPatrol generates a new backlog, and editors will feel stressed that there are a large number of edits which have not been reviewed.
    • We could investigate adding features to Automoderator to mark good edits as patrolled, not just to revert bad edits - T392825

Event Timeline

Restricted Application added a subscriber: Aklapper. · View Herald Transcript
Restricted Application added a subscriber: Huji. · View Herald TranscriptNov 4 2025, 1:06 PM
Samwalton9-WMF updated the task description. (Show Details)
Samwalton9-WMF added a subscriber: Ladsgroup.

Here is my 2c. It's more social/product than technical. Over the past 25 years many wikis have made diverging moderation workflows. It'd be nice and useful to consolidate them as much as possible and extend proven ones (such as RC patrolling) to more wikis. The reason I'm reiterating this is that I think in this regard, we have basically four different groups of wikis that need different works:

  • Wikis that have rc patrolling enabled.
  • Wikis that don't have rc patrolling enabled but they have pending changes enabled on all pages (dewiki, ruwiki, etc.)
  • Wikis that have neither of those
  • enwiki.

English Wikipedia is a dedicated category because they have built a bespoke system for moderation and integrating that cathedral with rc patrolling will require a non-trivial amount of both technical and social work both from WMF and the volunteers.

My suggestion is to enable it first on wikis that don't have FR on all pages nor rc patrolling so they can have at least a new functionality to use and rely on. Then remove FR on wikis where it's clearly is not working (very long and growing backlog of edits to be reviewed) and replace that with RC patrolling. Then in after certain time when we are more comfortable with rc patrolling usage, push for more adoption of it as either replacement of FR or in case of English Wikipedia, as the backbone of their anti-vandalism infrastructure.

Unrelated to the above: One complaint I heard when rc patrolling was being rolled out was that the red "!" was not user friendly (It wasn't clear to the users what it means, etc.) Maybe it could be looked at before further roll out.

RCPatrol generates a new backlog, and editors will feel stressed that there are a large number of edits which have not been reviewed.

One note about this is that since RC patrolling is bound to recentchanges table, the backlog falls of after 30 days since the edit so unlike FlaggedRevs the review backlog won't grow indefinitely.

New editors may find it motivational to learn that their edit has been reviewed by an experienced editor (T410882).

Also a lot of new editors feel motivated to become "autopatrolled" seeing it as the next thing to "unlock" and wear that as a badge on their user page.

This would be pretty hard for enwiki:

wgUseFilePatrol: true
wgUseNPPatrol: true
wgUseRCPatrol: false
  • user rights would need to be created and user groups changed around. this would require more code modifications to PageTriage, which is hardcoded to use the patrol and autopatrol user rights, but these same user rights are hardcoded into the patrolled edits system. and who has access to these user rights would need to be different. for example antivandalism folks (who need to be trusted to do antivandalism) do not need nearly as much qualifications as new page patrollers (who need to be experts at notability). would recommend using the existing rollbacker user group rather than creating another antivandalism user group.
  • FlaggedRevs has like 4 modes it can operate in: override, protection, neither, and not installed. Sounds like wgUseRCPatrol would definitely conflict with override, but may also conflict with the other modes as well. This would need investigation. (enwiki uses protection. other wikis could use any of these)

Recommendations:

  • skip enwiki and start with "easy win" wikis where it is technically easy to implement this. These would be wikis without FlaggedRevs installed, without PageTriage installed, and without $wgUseFilePatrol $wgUseNPPatrol $wgUseRCPatrol turned on.
  • hold a wider community consultation that includes small wiki monitoring team folks, small wiki admins, etc. and make sure that you're not missing any downsides to forcing this system to be turned on for many wikis.

Overall, my impression is that this would benefit small wikis more than large wikis. On large wikis, there's the risk of creating an infinite queue/backlog that could not be kept up with. Editor time is valuable, and implicitly directing people to spend their time reviewing every edit in a shiny new queue could take them away from other more important queues.

In Discord, someone posted some low usage statistics for existing wikis with $wgUseRCPatrol turned on. That should be investigated a bit more to make sure the Quarry query was correct. And if it was, then that is useful data and could be used in making a decision about this.

This would be pretty hard for enwiki:

wgUseFilePatrol: true
wgUseNPPatrol: true
wgUseRCPatrol: false
  • user rights would need to be created and user groups changed around. this would require more code modifications to PageTriage, which is hardcoded to use the patrol and autopatrol user rights, but these same user rights are hardcoded into the patrolled edits system. and who has access to these user rights would need to be different. for example antivandalism folks (who need to be trusted to do antivandalism) do not need nearly as much qualifications as new page patrollers (who need to be experts at notability). would recommend using the existing rollbacker user group rather than creating another antivandalism user group.
  • FlaggedRevs has like 4 modes it can operate in: override, protection, neither, and not installed. Sounds like wgUseRCPatrol would definitely conflict with override, but may also conflict with the other modes as well. This would need investigation. (enwiki uses protection. other wikis could use any of these)

Recommendations:

  • skip enwiki and start with "easy win" wikis where it is technically easy to implement this. These would be wikis without FlaggedRevs installed, without PageTriage installed, and without $wgUseFilePatrol $wgUseNPPatrol $wgUseRCPatrol turned on.
  • hold a wider community consultation that includes small wiki monitoring team folks, small wiki admins, etc. and make sure that you're not missing any downsides to forcing this system to be turned on for many wikis.

Overall, my impression is that this would benefit small wikis more than large wikis. On large wikis, there's the risk of creating an infinite queue/backlog that could not be kept up with. Editor time is valuable, and implicitly directing people to spend their time reviewing every edit in a shiny new queue could take them away from other more important queues.

There are a lot of misunderstandings on how RC Patrolling works. For example, "infinite queue" is not correct. The data for rc patrolling is in rc table and will fall off after 30 days. Meaning it can't possibly grow beyond a certain bound (unlike FlaggedRevs)

I'd argue since editor time is valuable, we should enable it so one editor doesn't need to review the same reviewed edit.

In Discord, someone posted some low usage statistics for existing wikis with $wgUseRCPatrol turned on. That should be investigated a bit more to make sure the Quarry query was correct. And if it was, then that is useful data and could be used in making a decision about this.

I don't know the discord discussion but it doesn't seem too low to me:
https://quarry.wmcloud.org/query/100544
https://quarry.wmcloud.org/query/100545
https://quarry.wmcloud.org/query/100546

Basically, in many large wikis a lot of edits get automatically auto patrolled, so it doesn't get reviewed (rc_patrolled = 2) then you have zero which is being reviewed and then 1 which is manually reviewed. Ratio of 0 to 1 isn't that bad.

Good food for thought. But I still worry about it being an infinite queue (never reaches zero backlog) even with the 1 month dropoff.

Would be good to calculate the # of non-autopatrolled revisions for enwiki per month to see how many (tens of thousands? hundreds of thousands? millions?) of extra revisions we'd need to patrol per month if we were to turn this on.

Good food for thought. But I still worry about it being an infinite queue (never reaches zero backlog) even with the 1 month dropoff.

I do not think it is necessary for this number to reach 0. The main objective here is to avoid two people reviewing the same edit, not to ensure that everything is patrolled.

I'd argue since editor time is valuable, we should enable it so one editor doesn't need to review the same reviewed edit.

As one of the folks involved in the discussion on Discord, I disagree with this take. The POV of "one editor doesn't need to review the same reviewed edit" only works if you work back of the queue/middle of the queue, and you want to look at (say) tougher edits. MediaWiki, in its current state, does not expose the back or the middle of the 30-day queue in a meaningful way (unlike FlaggedRevs). If you are working front of the queue, you don't really care, you continually advance looking through the front of the queue in terms of edits and collisions/double-reviews are not unexpected (this is kinda/sorta true even in like NPP where we do have a pretty robust patrolling system which is why policies like WP:NPPHOUR are in place to prevent NPP folks from jumping on the same article and potentially biting newbies).

Based on my (admittedly anecdotal memory, while using the IRC cvn-bot framework (or Wikiloop Battlefield back back in the day on enwiki), folks tend to heavily depend on external indicators, like ORES/algorithms offered by external tools, or stuff like "person was blocked on other wikis", "lots of text removed" etc to sort through and find offending edits.

In Discord, someone posted some low usage statistics for existing wikis with $wgUseRCPatrol turned on. That should be investigated a bit more to make sure the Quarry query was correct. And if it was, then that is useful data and could be used in making a decision about this.

I don't know the discord discussion but it doesn't seem too low to me:
https://quarry.wmcloud.org/query/100544
https://quarry.wmcloud.org/query/100545
https://quarry.wmcloud.org/query/100546

Basically, in many large wikis a lot of edits get automatically auto patrolled, so it doesn't get reviewed (rc_patrolled = 2) then, you have zero which is being reviewed and then 1 which is manually reviewed. Ratio of 0 to 1 isn't that bad.

Those are abysmal/not good numbers imo, though; 13% of edits in the last 30-day period were patrolled for frwiki, 14% of edits in the last 30-day period for nlwiki, and 19% of edits in the last 30-day period for fawiki (during what is effectively a holiday season). If anything, this proves that patrolling is a niche workflow and is not the one majorly used by editors. Given that antivandalism is fairly robust on large wikis, if patrolling were used in large amounts, I would expect a significantly higher coverage in terms of edits (at least a number approaching ~50%?) . A patrol rate of 13% implies that we are missing/not looking at/not patrolling a whopping 87% of edits.

Regarding the numbers I shared on Discord, my queries pin the number of patrol actions to revisions in 2025 across months to be pretty low overall, at 3-4% for frwiki mainspace (where I assume the most amount of patrolling occurs). I'm aware that because of the way I am looking at autopatrolled edits, there is a chance for a discrepancy, but a 10% discrepancy doesn't seem right, maybe there is some kind of extension that is automatically marking pages without log entries?

As a contributor from frwiki myself, I can say that I rarely mark edits as patrolled. This action does not feel sufficiently meaningful currently, as marking an edit as patrolled does not appear to prevent other patrollers from reviewing it again. In addition, I mark edits as patrolled only when I am confident that they are correct, which is relatively infrequent. Confidently assessing the quality of an edit in a subject area in which I lack sufficient expertise is more difficult than identifying clear vandalism on the same subject. That said, I do relatively little patrolling.

The total of actually patrolled edits should be the total amount of edits which have been either patrolled or reverted, since edits that are cancelled are rarely marked as patrolled.

The total of actually patrolled edits should be the total amount of edits which have been either patrolled or reverted, since edits that are cancelled are rarely marked as patrolled.

Okay, so I threw together a few python scripts and got:

Patrolled or reverted %:                         19.93%
All patrolled %:                                 15.66%
- Patrolled but not reverted %:                  8.71%
- -  Patrolled but not reverted with logs %:     8.67%
- -  Patrolled but not reverted without logs %:  0.04%
- Patrolled but reverted %:                      6.95%
- -  Patrolled but reverted with logs %:         4.68%
- -  Patrolled but reverted without logs %:      2.27%
- All patrols with patrol logs %:                13.34%
Reverted %:                                      11.22%
- Reverted and patrolled %:                      6.95%
- Reverted but not patrolled %:                  4.27%
- Total count (non-AP + mainspace):              114859 edits
No logs, not reverted, but patrolled:            8 edits

based on frwiki's RC table on Toolforge today. Approx 20% of edits (~19.9%) saw some kind of review action, with 11.22% seeing a revert. While 15.66% saw patrol actions, a lot of this, approx 7% (~6.95%) can be attributed to patrol actions associated with reverting edits (for example, the rollback action?). Only 8.67% of the edits that could be patrolled saw a patrol without an associated revert (patrolling). Additionally, around 2% of edits were reverted without patrol actions.

Based on this, I really don't see the value of RCPatrolling as much of it is partially duplicating the signal already offered by the mw-reverted tag and the usage to mark good edits/edits already looked at appears to be very low (in the context of frwiki atleast).