Wikipedia:Deferred changes

Deferred changes is a technical way to mark for review and if necessary delay the visibility of suspicious edits by new or unregistered users until they have been manually reviewed by an editor. The objective is to increase our reactivity to vandalism and other clearly inappropriate edits. An edit would be deferred for review when it matches an edit filter set to defer, when requested by a bot via the API, or when it is considered highly damaging by ORES. When an edit is deferred, all prior uninterrupted edits by the same user to the page are deferred for review as well. There are two types of 'deferrals'. If 'passive', the revision displayed to readers is still the latest revision as usual, but if 'active', the immediately preceding revision (the same revision that rollback reverts to) would be displayed as a preventive measure. The implementation is based on the flagged revisions extension and integrated with pending changes. All deferred edits would appear at Special:PendingChanges, which would provide the option to list only those actively deferred or from pending changes-protected pages, and could be reviewed by pending changes reviewers. When the page has a single author, and the edits can therefore not be deferred, it would be listed on a dedicated special page.

Rationale
The objective in introducing this new technical tool is to reduce the visibility of damaging edits to the public, both on low traffic articles where few people see it but where it can persist for a long time, and on high traffic articles where those are reverted faster but are viewed by more people. A January 2012 study published on the SignPost suggests that about 10% of damaging edits are viewed by more than 100 persons, many of them to high visibility articles; the authors state that new solutions are needed to address this issue. Not much has changed in this area since then. The persistence of vandalism and BLP violations on low traffic biographies of living people is a lingering problem. Despite anti-vandalism bots and semi-automated tools, a substantial proportion of those damaging edits is not identified and reverted in a timely manner.

Current proactive tools, the spam blacklist and the edit filter, offer only one preventive action: disallow, and in order to keep false positives low, only the most extreme types of edits are prevented. Deferring for review, even actively, provides a much less stringent option that can therefore be used more widely. This can also mitigate the issue of filters being evaded, especially in previous edits by the same user on the same page. The edit filter provides the ability to tag edits, but no efficient reviewing system exists for tagged edits. As a result, many are missed, sometimes for days, defeating the point of flagging those edits, so tagging alone isn't particularly efficient at highlighting suspicious edits to recent changes patrollers.

If those edits had been actively deferred by the edit filter, a bot, or ORES their edits would not have been visible to the public. Even if only passively deferred, marking them for review would ensure that they are taken care of, which would be of particular interest on lightly watched articles. Although not all 'bad' edits would be detected and subsequently deferred, since we still need to keep relatively small false positive rate and backlog of unreviewed edits, this should represent a significant improvement.

Impact analysis
The impact on the backlog of unreviewed delayed edits should be small, provided the edit filters and bots prompting active deferral as well as the ORES damaging threshold are kept in check, which would be the responsibility of administrators. All deferred edits would be listed at Special:PendingChanges, ensuring strong visibility. From experience with pending changes, the pending edits that are vandalism, spam or BLP violations are dealt with quickly. In contrast, on pending changes protected articles, despite being among the articles most subjected to this type of edits, a very large portion of all edits are still acceptable, and they end up forming the bulk of the backlog.

At the moment, most pending changes are reviewed within one hour, and it takes very rarely more than a few hours, this small increase can be managed. If this becomes an issue nonetheless, the backlog can be reduced by tightening the criteria used by automated processes or setting an expiry to the deferrals. The impact on general editors should be minimal. Unlike pending changes protection which remains on articles continuously for significant periods of time, and thus is visible and potentially interfering during that time, the ephemeral nature of a deferred change ensures a very limited distraction. The false positive rate will be maintained low to minimize the risk of good-faith editors having their change deferred and unnecessary reviews.

Edit filter
Currently, numerous filters target vandalism, spam or BLP violations, the user is usually either prevented from performing the edit or warned, in the latter case they can still confirm and save the edit, which is simply tagged. Those edits are often not reverted in a timely manner, in addition the users often find ways to elude the edit filter. The situation would be markedly improved by flagging these edits for review, along with previous ones by the same user, and delaying visibility for the most troublesome.

Bots
Bots would be able to defer for review edits via the API, actively only if the bot has rollback rights.

For example, works by calculating a probability that an edit is vandalism, and reverts when it is above a specific threshold, determined to match a specified false positive rate, currently of 0.10%, with exceptions such as when the user has already been reverted by the bot on the same article in the last 24 hours. The bot may be instructed to passively defer edits with a probability of vandalism where the associated false positive rate is between 0.20% and 0.15%, and actively defer those between 0.15% and 0.10% as well as repeated edits by a single user on the same article.

ORES extension
The mw:Extension:ORES could defer edits that it considers highly likely to be damaging, with a higher threshold for active deferral.

Comparison table
EF : Edit filter, SBL : spam blacklist, B : bot, BR : bot with rollback, R : rollbacker, A : admin