Research:Patrolling on Wikipedia/Literature

Documents reviewed

Peer-reviewed research

I pulled some of these from my own reference manager. I also queried Google Scholar for combinations of "Wikipedia" and terms related to patrolling and vandalism to identify candidates that I wasn't previously aware of.

Geiger, R. S., & Ribes, D. (2010). The work of sustaining order in wikipedia: The banning of a vandal. https://doi.org/10.1145/1718918.1718941
Müller-Birn, C., Dobusch, L., & Herbsleb, J. D. (2013). Work-to-rule: the emergence of algorithmic governance in Wikipedia. https://doi.org/10.1145/2482991.2482999
Gilbert, M., & Zachry, M. (2015). Tool-mediated Coordination of Virtual Teams in Complex Systems. https://doi.org/10.1145/2788993.2789843
Halfaker, A., & Riedl, J. (2012). Bots and Cyborgs: Wikipedia’s Immune System. https://doi.org/10.1109/MC.2012.82
Kumar, S., West, R., & Leskovec, J. (2016). Disinformation on the Web: Impact, Characteristics, and Detection of Wikipedia Hoaxes. https://doi.org/10.1145/2872427.2883085
Yang, D., Halfaker, A., Kraut, R., & Hovy, E. (2016). Who Did What: Editor Role Identification in Wikipedia.
Arazy, O., Ortega, F., Nov, O., Yeo, L., & Balila, A. (2015). Functional Roles and Career Paths in Wikipedia. https://doi.org/10.1145/2675133.2675257
Geiger, R. S., & Halfaker, A. (2013). When the levee breaks: Without bots, what happens to wikipedia’s quality control processes? https://doi.org/10.1145/2491055.2491061

Research reports

I reviewed these reports, as well as supporting documents when I had access to them. I discovered them from memory, through consultation with other current WMF research staff, and through keyword searches of the Research namespace on Meta.

Research:The_Rise_and_Decline
Community_process_workflow_interviews_(June_2015)
Research:Autoconfirmed_article_creation_trial
Research:New_Page_Patrol_survey/WMF_report
Research:Building_automated_vandalism_detection_tool_for_Wikidata
Research:Curation_workflows_on_Wikimedia_Commons
Research:Vandal_fighter_work_load
Research:Patroller_work_load
Research:Understanding_hoax_articles_on_English_Wikipedia
Research:New_editor_support_strategies
2018 Steward Spambot lock workflow report (internal WMF report)

Summary of themes

Patrolling as a collaborative activity

Although users of tools like Huggle and bots like ClueBot_NG seem to work on their own, in reality a lot of patrolling and vandal-fighting involves multiple-steps performed by different actors, some of which are human and some of which are not. For example, ClueBot may automatically revert an obviously vandalistic edit (like unexplained page-blanking) and leave a warning; then someone watching RecentChanges may revert a subsequent edit by that user and leave a more stringent warning; when the suspected vandal makes another edit that looks suspicious, a third patroller using Huggle will see that the editor already has two talkpage warnings, and report them to AN/I to request a block; then an admin reviews the alleged vandal's history and blocks them; then another bot places a notification on the vandal's talkpage that they were blocked. Patrollers can experience their process as collaborative, and they appreciate the team spirit that creates.

Patrolling as filtration

There's a kind of filter model to patrolling work, at least in the aggregate (and on the biggest wikis): Bots using ML or deterministic algorithms revert obvious stuff; humans focused solely on patrolling and using automated interfaces with built in triage feeds (e.g. AutoWikiBrowser, Huggle, and PageCuration) take care of stuff that's not automatically deletable (either because the vandalism was undetected by the algorithm, or because local policy requires human review in some/all cases); editors monitoring MediaWiki's built-in feeds like RecentChanges or personal watchlists identify subtler vandalism that went undetected by the first two stages, and/or more complex cases that require background research or judgement calls.

Some kinds of vandalism persist unnoticed for much longer than others; certain kinds of hoax articles can persist for months or even years without discovery.

Redundancy

There is redundancy in the patrolling system. This redundancy can lead to wasted work--e.g. multiple Huggle users 'racing' to revert the same edit. The redundancy is due primarily to lack of mutual awareness (most tools operate at revision-mediated timescales, not in real-time) and the high cost of explicit coordination among distributed actors performing similar tasks simultaneously.

There is also resilience in this redundancy. When a stage in this multi-stage filter break down or can't deal with a sudden influx, actors further down can step up to take up the slack, at least for a while--e.g. patrollers increasing their activity in response to an extended downtime for ClueBot_NG. It's not clear to what degree this response to Cluebot going down was emergent, stygmergic or explicitly coordinated.

The need for speed

Specialized vandal-fighting tools like Huggle, AWB, PageCuration, userscripts and bots are optimized for speed. The goal is to get inappropriate content off the wikis as quickly as possible, and to block bad actors before they can do more damage. Editors and some bots/tools use heuristics and digital activity traces of other patrollers to facilitate these quick decisions--e.g. Huggle's triage feed default to prioritizing edits by editors who have multiple talk page warnings for review, or the disproportionate rate of deletion of articles by newly-registered editors as "suspected hoaxes".

False positives are taken for granted

Prioritization of speed and reliance on heuristics (both mental heuristics on the part of patrollers, and heuristics formalized into algorithms) result in false positives. Algorithms don't handle exceptions well, and can be biased (e.g. ORES predictions rate IP contributions as higher likelihood of vandalism, all things being equal). Vandal fighters may assume that good faith contributors who are erroneously reverted will follow up to contest the decision. Time-consuming judgement calls may be rejected, reverted, or deleted by default.

Patrolling workflows can be complex, contingent, and brittle

There are temporal dependencies encoded in the software of the workflows: a vandalism report to AI/V allows an admin watching that feed to use a custom userscript (ResponseHelper) to block that user; the block triggers a bot who posts the notification of the block.

A lot of these tools (bots, scripts) have different maintainers, some may be not have a dedicated maintainer at all, or the maintainer may be unresponsive. Forums like AI/V need to be monitored constantly by someone with the requisite permissions to act on reports. Many workflows require background research, judgement, comparative analysis, multiple rounds of communication, manual steps that must be completed by a single editor in a specific order.

Different wikis have different rules, roles, and tools

Not everyone has access to the best available tools for patrolling. Sometimes this is a policy decision--some tools are only available to people with certain userrights, which must be applied for or attained through meeting some minimum threshold of activity or age. More importantly, most tools are designed to work with a particular process, or a particular wiki.

Degree of tool support (either bots, apps like Huggle, or userscripts) is not distributed in a need-based way. Stewards, who address a wide variety of high-importance community issues, including extreme vandalism, across (theoretically) all wikis have some of the lowest levels of tool support, and their work tends to be among the most complex, and therefore expertise and labor-intensive.

That said, the vast majority of research I was able to review for this project is based on analysis and reports from English Wikipedia. I was able to find relatively little direct information about vandal fighting on other wikis. Therefore, the decision to focus on non-English Wikipedias for this project is well-motivated.

Key questions for interviews

What tools/rules/roles re:anti-vandalism are different on your home wiki than on other wikis you know of?
What tools/rules/roles are most salient to your particular anti-vandalism work and why?
What kinds of vandalism are most concerning to you? Which are most prominent on your wiki?
How collaborative is the process you follow when you patrol?
To what degree does your work involve collaboration/coordination with autonomous agents like bots?
In your opinion, what are the biggest existential threats, if any, that vandalism can present to Wikipedia?
What kinds of vandalism is hardest to spot? What do you look for when encountering these kind of vandalism?
Have you encountered any of the following: hoax articles, coordinated attempts to insert mis/disinformation across multiple articles, attempts to use non-credible sources to support dubious claims
What's the longest time you take on any particular vandalism case? What kind of vandalism is that?
(If your wiki has vandalism-detection/prediction algorithmic tools) How much faith do you put in these algorithmic judgements? Have you noticed cases where they work particularly well, or particularly poorly?
What metadata about vandals or vandalism edits is most useful/necessary for determining or dealing with vandalism, that is currently not accessible to you or hard to access?