Damage Done: Wikipedia Leaves Biased Edits Unchecked Despite Editor Bans

Executive Summary

More than a year after ADL published “Editing for Hate: How Anti-Israel and Anti-Jewish Bias Undermines Wikipedia’s Neutrality,” exposing a coordinated, multi-year campaign by approximately 30 Wikipedia editors to inject antisemitic and anti-Israel narratives into entries on Israel, Hamas, Zionism, and Jewish history, the underlying bias remains embedded in the online encyclopedia’s most-read pages on the conflict, and the work of even sanctioned editors is still largely intact and unreviewed.

The Wikipedia community banned some of the editors identified in the original report from the topic area, or the site entirely, with sanctions beginning in January 2025 and afterward amid sustained external scrutiny from media, Congress, and federal officials. The Wikimedia Foundation, under new leadership, signaled commitments to anti-abuse mechanisms. But the edits themselves were never systematically reviewed, and much of that original damage is still standing. 

This matters beyond Wikipedia as these articles feed into AI training pipelines, Large Language Model (LLM) outputs, and search engine results in a matter of minutes, carrying the distortions with them and shaping how millions of readers encounter this information.

Wikipedia’s “Gaza war” article, for example, was stripped of the forensic record of the Hamas-led attacks on October 7, 2023, by a since-banned editor who removed sections covering forensic documentation that victims were bound with wire and burned, evidence of sexual violence, and video evidence of an attempted decapitation. What remains is a single line about how terrorists “reportedly engaged” in such violence during the attack. That same banned editor worked on the “Hamas” article, which does not even identify the Palestinian Islamist extremist group as a terrorist organization designated by the U.S. and other governments until the very end of the lead section; everything before that describes Hamas in terms consistent with a governing authority in Gaza, not the terrorist group that led the worst antisemitic massacre since the Holocaust. 

Wikipedia's "Zionism" article, meanwhile, defines Zionism as achieved “through colonization” and states, as fact, that Zionists sought a state with "as few Palestinian Arabs as possible" — a characterization Wikipedia's own editors have internally flagged as disputed. As “Editing for Hate” reported, six editors who were later topic-banned were among those who installed that language, and a February 2025 moratorium on further discussion locked it in place. The moratorium has expired; the central framing remains.

At the same time, biased editing across Wikipedia has continued through quieter, subtler techniques than those documented in 2025, even as the visible coordination decreased significantly. A newly identified cohort, working alongside some members of the original 30 who avoided consequences for the coordination, now carries this work forward. 

This follow-up report by the ADL Center for Technology and Society (CTS) presents four main findings:

• Finding 1: Past biased edits have not been undone. No banned editor's body of work has been systematically reviewed to determine which edits should be corrected. On Wikipedia's "Hamas" article, one banned editor's changes remain in place a year after the account was banned. Across six flagship articles examined in this report, roughly 91% of that editor's edits are still in the live text, with no review since the ban.
• Finding 2: Sustained external pressure and enforcement have driven a measurable decline in coordinated activity. Talk-page coordination within the original 30-editor group fell 95%, and same-page tandem editing fell 88%, following congressional, federal, and platform scrutiny.
• Finding 3: Biased editing continues, quietly. A July 2026 review found 19 accounts actively editing Israel- and Palestine-related articles: eight editors from the original 30 who avoided sanction, and 11 newly identified editors. Their techniques are more subtle: undoing other editors' changes to hold a framing in place, hedging factual claims, and rebuilding articles slowly enough that no single edit draws attention.
• Finding 4: The new cohort now sets the frame. On the 229 articles where these 19 accounts overlap, the 11 newly identified editors are far more active recently and hold the most recent edit, effectively a page's current text, on roughly six in ten of them. Several of these editors are among the principal authors of Wikipedia's pages on October 7 sexual violence and have installed and defended a skeptical framing for more than two years.

ADL's recommendations below are addressed to the Wikimedia Foundation, to search engines and AI developers, and to policymakers and the public who rely on Wikipedia.

Introduction

With more than seven million entries in its English edition alone and roughly 15,000 new articles added each month, Wikipedia is among the most-visited reference sites on the internet, and for many users it is the first stop for information on any topic.

Its open-editing model rests on the assumption of good faith among contributors and on a stated editorial policy of verifiability and neutrality. That neutrality policy, known on the platform as NPOV (neutral point of view), requires articles to represent all significant viewpoints published in reliable sources, in proportion to their prominence. It is the standard against which the editing documented in this report is measured, and the standard the editors themselves invoke when defending that editing.

Wikipedia is the foundational infrastructure for the broader online information ecosystem. It ranks at or near the top of search results for most substantive queries, its content is excerpted in Google’s knowledge panels and similar features across other platforms, and survey research finds it is used as a starting reference in classrooms, in newsrooms, and in everyday research workflows even where official guidance recommends against citing it directly. Wikipedia is also a core training corpus for most major LLMs in production today, including the systems behind ChatGPT, Gemini, Claude, and the AI search features now embedded in Google, Bing, and Meta’s products. 

The framing that dominates a Wikipedia entry propagates outward. Antisemitism and anti-Israel bias on the platform move into all of those channels at once, which is what makes neutrality at Wikipedia a public-interest question rather than an internal Wikipedia matter.

In March 2025, ADL’s “Editing for Hate” documented the systematic coordination and evasion of Wikipedia’s neutrality policies by roughly 30 editors to manipulate content related to Israel, the Jewish community and history, the Gaza war, and the Israeli-Palestinian conflict. This campaign included the removal of citations to reputable sources, tandem voting on contested talk pages, and a sustained effort to downplay Palestinian terrorism and antisemitism while foregrounding criticism of Israel.

Since its release, the report’s findings have been the subject of a U.S. House Committee on Oversight investigation, multiple disciplinary proceedings before Wikipedia’s Arbitration Committee (ArbCom, the elected panel of volunteer editors that handles the platform’s most serious conduct disputes), and substantial coverage in mainstream and trade media.

This follow-up report assesses what has changed, in policy, in editor behavior, and in measurable on-platform activity, and what work remains.

Finding 1: Past Biased Edits Have Not Been Undone

The legacy edits of the campaign documented in “Editing for Hate” remain largely intact across Wikipedia’s most-read entries on Israel, Zionism, Hamas, and Jewish history. Banning an editor does not restore the articles they distorted, and to date no banned editor's body of work has been systematically reviewed to identify which edits introduced bias and should be corrected. A ban is a reason to review an editor’s work edit by edit, and the audit ADL recommends below is designed to do that. As of October 2026, the articles widely relied upon to summarize key events and concepts in the Israeli-Palestinian conflict remain shaped by the non-neutral edits documented by ADL more than a year ago.

Case Study: Iskandar323

The account of Iskandar323 shows what that looks like in practice. The account was part of the original group of 30 editors documented in ADL’s 2025 report, later site-banned by ArbCom, and was independently examined in a May 2026 academic analysis by Dr. Shlomit Aharoni Lir of Tel Aviv University’s Institute for National Security Studies (INSS). 

ADL's review shows a years-long editing history with a sharp concentration in the post-October 7 window. The account made 485 edits across six articles, referred to throughout this report as flagship articles, over its full editing history on those pages, spanning a period well before the Hamas-led attacks of October 7, 2023. Of those, 225 came after the attacks, and 85 were in October 2023 alone. All six are among the highest-trafficked English-language Wikipedia entries on the topic: “Israel,” “Gaza war,” “October 7 attacks,” “Hamas,” “Zionism,” and “Jews.” 

This is not the record of an occasional contributor. According to ArbCom’s public proceedings, the site ban was imposed because Iskandar323 continued editing on Palestine/Israel-adjacent topics even after the original topic ban took effect, specifically on articles such as “Hellenistic Palestine” and “Yahweh,” where ArbCom found that adjustments to ancient historical language constituted proxy engagement in the modern conflict. A topic ban alone could not close that workaround. Only the site ban could.

The persistence bars in the graphic show how much of the account’s work outlasted the ban.

Persistence varies across the six articles. It runs highest where Iskandar323’s changes were absorbed into the article’s own language, and lower where later editors reworked the surrounding text. Of the account’s 485 edits to these articles, 398 could be checked against the live text as of August 2026. The remaining 87 were page moves, Wikipedia’s built-in process for renaming an article, which carries its content and edit history to the new title and leaves a redirect at the old one (where persistence cannot be checked); edits to revisions Wikipedia has since deleted; or changes that could not be reliably traced. Of the 398 checkable edits, 362, roughly 91%, remain in place.

“Softening” Hamas: An example of lasting damage

Iskandar323’s edits to Wikipedia’s “Hamas” article are the most consequential single set of changes in this review. That article surged to nearly one million views per day on October 7–8, 2023 and traffic has remained high since. 

Three days after the October 7 attacks, Iskandar323 moved the bulk of the main article's material on the group's antisemitism, anti-Zionism, hate speech by its representatives, use of human shields and child combatants, restrictions of political freedoms in Gaza, and human rights abuses into "Criticism of Hamas," a far less-read companion article, leaving only a brief generalized reference behind. Nothing was deleted. But a reader arriving at "Hamas" in the days when the page drew close to a million views daily would no longer encounter that material unless they followed a link to a second article. The mechanism is illustrated below.

Six weeks later, on November 21, 2023, the same account removed roughly 17,000 characters of text and citations from “Gaza war,” deleting a section detailing the forensic evidence presented by Israeli authorities of the violence by Hamas-led terrorists on October 7, 2023, including an attempted decapitation, torture, and sexual violence. Sections that were removed include descriptions of a clip screened for foreign journalists in Israel showing a “partially burnt female corpse” in a state of undress with “underwear missing,” a forensic analysis of victim remains showing “charred hands with marks indicating the victim was bound “with metal wire before being burned alive,” and the discovery of a “large charred mass” showing “the remains of a parent and child bound together” before being killed. Two days later, this account returned and cut roughly 7,400 more characters from what remained, including witness accounts from Israeli soldiers and the first-responder organization ZAKA.

The edit summary describes the section as covering “confirmed or otherwise” events of October 7.

Iskandar323 was eventually removed from Wikipedia altogether, but not for these edits: the account was topic-banned in January 2025 for its conduct in the topic area, and site-banned a year later for continuing to edit in defiance of that ban.

Neither the Wikimedia Foundation nor Wikipedia’s volunteer community has, to date, systematically reviewed or assessed the account’s changes to the “Hamas” article, even though the editor responsible was banned. 

That review is what a ban should set in motion. Reverting a banned editor’s entire body of work would risk sweeping up legitimate contributions along with biased ones. A process that examines the work edit by edit, restoring removed sourced material and balanced wording where bias is documented, would correct that record without that risk. Wikipedia already has a similar process in place for copyright issues: When an editor is found to have repeatedly added copyrighted text, volunteers can open a contributor copyright investigation and check that editor’s past additions, article by article. Wikipedia has no comparable process for editors sanctioned over biased editing. ADL’s recommendations below call for one.

"Recasting" Al Jazeera: A second example of lasting damage

The problem is not limited to banned editors.

Cinaroot is one of 11 accounts newly identified in a July 2026 review by the ADL Center for Technology and Society of Israel- and Palestine-related editing, examined in Finding 3 below. The account registered in 2022, went dormant, and reactivated two weeks after the October 7 attacks. 

To date, nearly 200 of its edits are to the Wikipedia entries on Qatari state broadcaster Al Jazeera: 122 to the article "Al Jazeera Media Network” and 71 to the article on the network's influence, with two further Al Jazeera articles created outright. A recent joint investigation by the Foundation for Defense of Democracies and NPOV, a private company that investigates coordinated influence campaigns on Wikipedia and other online platforms, described Cinaroot as “among the most influential editors on all of Wikipedia” who worked to recast Al Jazeera as "a democratizing global force" and, according to that report, to methodically frame the network as an independent newsroom. In August 2026, weeks after that investigation named the account, it exercised Wikipedia's courtesy vanishing option and no longer appears under that name.

Finding 2: Wikipedia Enforcement, External Pressure Coincide with a Decline in Coordinated Activity

Since the publication of “Editing for Hate,” the coordinated activity of the 30-editor group documented in that report has declined sharply, coinciding with sustained, multi-front external pressure on the platform. That decline is measurable and runs across multiple metrics. 

The pressure came from several directions: congressional scrutiny, including a House Committee on Oversight investigation and a bipartisan Congressional letter from 23 members, federal inquiries from the Department of Justice and the Senate Commerce Committee; enforcement work by Wikipedia’s NPOV-policy editors, ADL reporting, and broader media scrutiny. 

This analysis draws on Wikipedia's public revision API, comparing two equal-length windows of six months bracketing the March 2025 publication date of “Editing for Hate”; cross-validation against the original datasets confirms the same direction and comparable measures of change. The full methodology is located in the appendix.

Part of the decline in coordinated activity reflects a smaller roster: six editors from the original 30 were topic-banned; three were site-banned, and several others have gone inactive. But the sharpest declines are in measures of behavior.

The clearest effect of external pressure is on coordination rather than volume. Total editing by the documented group fell 27%, but the group itself shrank by a third over the same period, so the editors who remained were individually more active than before. Coordination fell far faster than the roster: same-page tandem editing dropped 88%, and talk-page coordination fell 95%. The sustained scrutiny affected the visible machinery of coordination, not the editing.

Some important caveats apply. The visible signs of coordination did diminish significantly: Talk-page coordination is by definition the visible form of cooperation and its near-disappearance can mean editors have stopped working in concert. But it can also mean they have moved off-platform or stopped leaving traces where the messages are publicly indexed.

Sanctions by Wikipedia’s Volunteer Community

Since “Editing for Hate,” some of the bad actors who had been operating in plain sight retreated. Wikipedia’s Arbitration Committee, or ArbCom, handed down topic bans, site bans, and procedural restrictions in the Israel-Palestine topic area beginning in January 2025 and continuing since. 

The sanctions differ in scope and severity: a topic ban prohibits an editor from contributing to a specific subject area while allowing them to continue editing the rest of the encyclopedia. A site ban prohibits the editor from contributing to Wikipedia entirely, on any topic. 

Of the approximately 30 editors named in "Editing for Hate," at least six have to date been topic-banned for non-neutral editing or repeated rule violations: Iskandar323, Levivich, Makeandtoss, Nableezy, Nishidani and Selfstudier. Three were banned from the site (including Iskandar323) and several more have gone inactive on Wikipedia altogether.

Site bans are ArbCom’s most serious editor-level remedy and remain rare across the platform. They are generally reserved for repeated violations after lesser sanctions have failed. An editor who has been site-banned has been judged by Wikipedia’s highest internal authority to be unsuitable to participate in the project at all. 

In April 2025, ArbCom imposed site bans on two editors implicated in a separate coordination campaign on the off-Wikipedia "Tech for Palestine" Discord server, where a dedicated Wikipedia workgroup coordinated editing across more than 100 articles on the conflict. In early 2026, ArbCom voted to site-ban Iskandar323, bringing the bans up to three accounts.

Much of the evidentiary work that supported these sanctions came from Wikipedia's own volunteer NPOV-policy editors, who compiled specific edits, mapped coordination patterns, and presented evidence at ArbCom proceedings.

ArbCom also passed a topic-ban extension motion in January 2026. Where an existing topic ban for the Israel-Palestinian conflict has been found insufficient to prevent disruption, administrators may now extend that ban to topics related to Israel, Israelis, Jews, Judaism, Palestine, Palestinians, Islam, and Arabs. The restriction closes a gap that topic-banned editors had been able to exploit. A ban on the Israeli-Palestinian conflict "broadly construed" left room to argue about its edges, and editors barred from the conflict itself continued editing articles on Jewish history, religious figures, antisemitism, and the reliability of sources, then contested enforcement on the grounds that those pages fell outside the ban. Iskandar323's site ban followed exactly this pattern.

The coordination declined; the editing did not. The damage from the original campaign remains despite the bans and other measures. The findings that follow describe what that editing looks like without coordination and who is now doing that work.

Finding 3: Biased Editing Continues, Quietly

ADL’s ongoing monitoring shows that biased editing has continued to reshape Israel- and Jewish-related entries across Wikipedia through mechanics less visible than those documented in 2025. The techniques below are ones a single editor can apply alone, leaving no coordination signature to assess. The group sustaining them is only partly the same: eight of the 19 accounts examined here appear in "Editing for Hate," and 11 do not.

Nineteen Active Accounts

A July 2026 forensic review of editing records found 19 accounts actively editing across Israel- and Palestine-related articles on Wikipedia, eight of them unsanctioned editors from the original 30 referenced in "Editing for Hate.”

The remaining 11 have editing histories that predate ADL's original investigation. They are new to ADL's documentation, not to the online encyclopedia. Several have edited Wikipedia for years, with one account registered as far back as 2012. Their editing records, like those of the original cohort, run mostly through articles that relate directly to the Israeli-Palestinian conflict, among them some of Wikipedia's most contested pages. 

The editing techniques are slightly different, however: more subtle and defensible under Wikipedia's rules rather than overtly biased, while still reshaping what a reader encounters.

Editing Activity

The editing histories of three of the 11 newly identified editors illustrate three distinct methods: rebuilding a definitional article until one account becomes its dominant author, reverting other editors across dozens of shared articles to hold a framing in place, and reframing an article's opening premise.

The editing histories of three of the 11 newly identified editors illustrate three distinct methods: rebuilding a definitional article until one account becomes its dominant author, reverting other editors across dozens of shared articles to hold a framing in place, and reframing an article's opening premise.

DMH223344

Registered in 2023, this account has made itself the principal author of the site's “Zionism” page. It has made 543 edits to the article itself, making it the second most prolific contributor in the page's 24-year history, behind only one other editor. Across those 543 edits, the account added roughly 193,000 characters and removed roughly 188,000, for a net change of under 5,000. Almost nothing about the article's length changed. Almost all of its contents did, on the page that defines the movement for every reader, search engine, and AI system that consults Wikipedia.

The account’s own edit summaries describe the direction of the rebuild. In September 2024, it added some 12,700 characters to “add discussion on zionism and colonialism,” consolidating material on “settler colonialism” that was previously spread through the article's general history, into a section of its own. Consolidation of this kind adds no new argument to an article. It raises the prominence of one already there, giving the material its own heading, its own line in the table of contents, and a fixed position. Two months later, this account edited the page to "make sure transfer as euphemism for ethnic cleansing is covered." In the historiography of the conflict, "transfer" refers to proposals to move the Arab population out of the area, and historians disagree over how central those proposals were to the Zionist movement. The account's edit summary states what the edit was for: to ensure the article treats the term as a euphemism for ethnic cleansing. 

The lead of the "Zionism” article as it stands today defines Zionism as achieved “through colonization” and states, as fact, that Zionists sought a state with "as few Palestinian Arabs as possible." Wikipedia’s own editors have internally flagged this characterization as disputed. Presenting a contested characterization as settled fact, rather than as one perspective among others, runs counter to the neutrality Wikipedia claims to uphold.

As "Editing for Hate" documented, this language was installed through revisions involving multiple editors, including six who were topic-banned for their conduct mere months before the report was released. When other editors attempted to restore the previous, neutral wording, an administrator with no prior involvement in the topic area imposed a 12-month moratorium on reopening the question in February 2025, locking the disputed version in place. The moratorium expired in February 2026, and the lead reopened within a day, but its central framing survived. Editors now dispute only how that framing is sourced and worded, not whether it belongs, a sign it has become the article's settled baseline. None of the topic-banned editors’ contributions to the article have been reviewed. 

DMH223344, which wrote roughly a quarter of the surviving text, has been dormant since December 2025. The account’s rebuild of the article remains the foundation others edit around.

Butterscotch Beluga

This account focuses on gatekeeping through reverts, or undoing edits. Close to 30% of its 500 most recent article edits are reverts, a high share for an account of any age and an unusual one for an account that began editing just two years ago. It has not edited since late June 2026.

Its gatekeeping is clearest on the “Zionism” article. In December 2024, the lead’s statement that Zionists sought a state with “as few Palestinian Arabs as possible” went to a Request for Comment (RfC), Wikipedia’s formal process for drawing the wider editor community into a content dispute. There, the account has argued that the sentence was reliably sourced and should stay. 

The same approach shows up in smaller edits elsewhere. On “German support for Israel in the Gaza war,” it restored wording that the German government “denied” a genocide was taking place in Gaza, which the other editor had changed to “rejected the assertion,” a phrasing choice that implies an established fact and an alleged German denial of that fact. On the article about the slogan “Death to Israel,” it removed the category “Incitement to genocide of Jews,” writing, “Do not treat Jews & Israel as interchangeable groups.”

On pages related to the Hamas-led October 7, 2023, attacks in Israel, the account reverted the work of an editor who had changed the lead of “October 7 attacks” to describe the subject as a massacre. On “Sexual and gender-based violence in the October 7 attacks,” it restored the wording that a rape allegation had been “later discredited” over another editor’s “have been called into question.” Both pages are the subject of a case study in Finding 4 below.

إيان

This account focuses on reframing the opening premise of a specific page. In the Wikipedia article “October 7 attacks,” this account added blockade framing to the lead, rewrote the first sentence so that terrorists “broke out from the blockade of the Gaza Strip,” and labeled that change only “ce” (Wikipedia’s shorthand for a minor copyedit). The account also deleted the clause describing the attacks as “the first invasion of Israeli territory since the 1948 Arab-Israeli War” as a “deceptive claim,” and added a map of the blockaded Gaza Strip. This was done in one sitting in June 2025.

Another editor reverted the changes some six hours later. On the talk page, the account defended the rewrite by citing Democracy Now, the anti-Zionist outlet Electronic Intifada, and Al Jazeera, describing October 7 as a “jailbreak” from Gaza. 

The revert undid the edit, but not its effect. The article's opening sentence, as of September 2026, describes the attacks as incursions from the "blockaded" Gaza Strip, the framing this account introduced. The invasion clause was restored, though it now appears further down the lead rather than in the first sentence. A single edit was rejected, and the framing it carried was absorbed, which is how this editing works: the lead of the encyclopedia's article on October 7 now opens on the blockade.

"Editing for Hate" Editor Activity

Researchers also reviewed the continued activity of four of the eight still-active editors from the original group of 30, finding each consistently advancing a non-neutral framing across multiple articles between the publication of “Editing for Hate” in March 2025 and July 2026. 

Some examples of the editing activity from these editors within this time window follow below. 

Skitash, the most prolific editor in that cohort, reverted the description of Marwan Barghouti, a Palestinian figure serving five life sentences for murder, from “convicted terrorist” to “elected legislator,” and has repeatedly edited language on the “Hezbollah” article to frame the U.S.-designated Lebanese terrorist organization, an Iranian proxy, as a “Lebanese Shia Islamist political party with an active paramilitary wing.” These edits remain on the pages.

Vice_regent has focused its editing on casualty framing, defending Gaza Health Ministry figures against efforts to qualify them. In the weeks after the bodies of Israeli hostages Shiri Bibas and her two young sons, Kfir and Ariel, were returned to Israel in February 2025, Vice_regent made 64 edits to the article on their abduction and killing, and its discussion page. The edits worked to keep the manner of their deaths open: removing a statement that the family had been killed because “source don’t yet say the bibas family was mass murdered, as autopsy on their bodies is pending,” adding that “neither Israel nor Hamas have provided evidence to back their claims,” and opening a discussion-page section headed “Evidence of either claim?” Israeli forensic authorities concluded that Shiri Bibas and her children were murdered in captivity and found no evidence they had been killed in an airstrike, contradicting Hamas's account.

Since “Editing for Hate,” these four original-cohort editors, who were never sanctioned, have made roughly 1,300 edits to Israel- and Palestine-related articles.

Finding 4: The New Cohort Now Sets the Frame

This finding shows how much of the topic area's current framing belongs to the group of 19 accounts, and to the newly identified 11, in particular. The 19 accounts were identified by a forensic review of editing activity across Israel- and Palestine-related articles, then checked against the roster of the 30 editors documented in “Editing for Hate.” Eight appear in both. The remaining 11 are newly identified.

To measure, researchers identified 229 shared articles, pages edited by two or more of these 19 accounts, and collected 565 records of recent edits, each traceable to a specific, dated, public revision. The set emerged from the accounts' own histories rather than advance selection: Each account's most recent 500 article edits were pulled from Wikipedia's public records, and any page touched by two or more accounts entered the dataset.

Of the 229 shared articles, 167 relate directly to the Israeli-Palestinian conflict, among them some of Wikipedia's most contested pages. Shared editing on these pages is not, on its own, proof of coordination; active editors will naturally overlap on the encyclopedia's most contested articles. But the scale of that editing points in one direction: Recent activity across the shared articles in the topic area belongs disproportionately to the 11 newly identified accounts.

An edit counts as recent if it was made within the 12 months before the July 2026 survey. By that measure, 95% of the newly identified editors' records were recent, compared with the original cohort’s 83%. The typical most recent edit by a newly identified editor dated to late April 2026; for the original cohort, to mid-March 2026. The original cohort's records stretch back to 2023, while the newly identified group's reach back no further than late 2024.

Newly Identified Accounts Have Last Word

That recency gap matters because of how Wikipedia works. The most recent edit to a page is also its current text, and shapes how it reads. Across the 229 shared articles, a newly identified editor holds that position on 144, roughly six in ten, and on 95 of the 142 articles both groups have actually edited, two in three.

The pattern sharpens on 17 articles identified as handoffs: pages where every original-cohort editor’s most recent edit is more than a year old, but a newly identified editor has edited recently. The reverse, where an original-cohort editor is still active on a page the newly identified group has gone quiet on, happens once. 

Case Study: Controlling the Frame on October 7 Sexual Violence

These techniques are clearest on the Wikipedia pages covering sexual violence during and after the Hamas-led October 7 attack.

ADL pulled the complete revision histories of four pages and their discussion pages: “October 7 attacks,” “Sexual and gender-based violence in the October 7 attacks,” “Screams Without Words” (on the New York Times investigation) and “Screams Before Silence” (on the documentary). That is 9,576 revisions in all: 4,597 to the articles themselves and 4,979 to the discussion pages where their wording is negotiated. Editors documented in this report made 832 of those revisions, across the articles and their discussion pages alike, roughly one in every 11.

The finding is the pattern: hedged language installed in article leads and defended for over a year, statements in Wikipedia’s factual voice converted to attributed claims, and the resulting framing protected through at least 89 reverts in activity running from November 2023 to within days of the data pull in July 2026.

Reverting and reframing

A review of the “Sexual and gender-based violence in the October 7 attacks” Wikipedia article found that the most prolific contributor in the entire history of the page is Pincrete, an account in the research behind “Editing for Hate,” with 94 edits. The third most prolific, with 77 edits, is Raskolnikov.Rev, an account that surfaced when the ADL Center for Technology and Society examined this page and others on the topic. This account also ranks third all-time on “Screams Before Silence.”

On “Sexual and gender-based violence in the October 7 attacks,” Raskolnikov.Rev has worked to preserve the article’s framing that Israeli women and girls were “reportedly” subjected to sexual violence during the October 7 attacks. Between August 2024 and November 2025, Raskolnikov.Rev reverted the work of at least four different editors who removed that qualifier, restoring it each time. The account’s public edit summaries argue that “these are not weasel words... removing them in fact violates NPOV,” a reference to Wikipedia’s neutral point of view policy. The word at issue is “reportedly,” in the opening sentence of the article on sexual violence during the attacks. Whether it stands determines whether a reader meets that violence as established or as claimed. Defending it for 15 months against four different editors is how one account settled that question for everyone who reads the page.

The account also framed the same article as describing “a debate over the existence of weaponized sexual violence.” These edits occurred in March 2025, 17 months after the attacks, by which point the European Union had sanctioned Hamas units over sexual violence and a United Nations report was on the record. In the revisions summary, Raskolnikov.Rev said the debate “includes whether there was any weaponized sexual violence as part of the attacks by militants at all.”

Smallangryplanet, one of the 11 newly identified accounts, has made 30 article edits and 51 discussion-page edits on “Sexual and gender-based violence in the October 7 attacks,” placing the account among the top five participants in that page’s discussions.

Tagging, then removing

A second technique appears in Smallangryplanet’s editing of “October 7 attacks” over five weeks: flag a passage as unsupported, then delete it because it is unsupported. The sequence began on December 13, 2024, when the account removed all mentions of the Israeli police from the article’s account of the attacks. Two days later, it flagged a sentence in the sexual violence section with a “by whom” tag, and nine days after that it deleted the sentence it had tagged. The following day, it removed additional material, and in January 2025 it deleted an entire section, citing Wikipedia’s rule that material should appear in proportion to its prominence in reliable sources.

The sequence matters more than any single edit. Flagging a passage as unsupported and deleting it days later for being unsupported produces a record in which every step looks like maintenance. The effect over five weeks was evident: the removal of every reference to the Israeli police from the article’s account of the attacks, a sentence attributing sexual violence that the account had itself tagged nine days earlier, and finally an entire section. A reader arriving at Wikipedia’s “October 7 attacks” page today finds an account of the attacks roughly 3,300 characters shorter, with the police findings and the attribution of sexual violence no longer there.

What this case study does and does not claim

The finding here is that editors documented in this report are among the principal authors of the pages defining October 7 sexual violence, and that, across almost three years, they have installed and guarded a consistent skeptical framing. In Wikipedia’s factual voice, an article states that something happened, while its attributed voice states that someone says it happened, inviting the reader to weigh the source. Applied consistently to one side's account of an event, that shift turns documented fact into contested claim.

Conclusion

More than a year after “Editing for Hate,” Wikipedia's most-read entries on Hamas, Israel, the Gaza war, and Zionism remain shaped by years of non-neutral editing, and as Finding 1 shows, that damage remains largely unrepaired.

The coordinated mechanics of the campaign have visibly diminished, and the Wikimedia Foundation has signaled an intention to invest in anti-abuse infrastructure. That decline is real, but it is not repair. The campaign now edits more quietly, not less effectively, and the text it already wrote has never been systematically reviewed. Neither response — individual bans nor promised infrastructure — closes the governance gap that allowed a coordinated campaign to operate for years. Reviewing and repairing the distortions is the work that has not yet begun. As long as it remains undone, every search query, AI-generated answer, and knowledge panel drawing on those articles carries the distortions forward.

The four structural recommendations ADL issued in 2025 that remain unimplemented all aim at the same thing: ensuring that Wikipedia's integrity does not depend on perpetual external pressure to enforce its own stated values. This report renews those recommendations, calls for a structured content audit to address the legacy-edit problem, and asks search engines and AI developers to address how Wikipedia's contested content reaches their users. The path forward is structural reform, and ADL will continue to advocate for it. 

Recommendations

The Wikimedia Foundation, the nonprofit that hosts and administers Wikipedia, enters this moment with new leadership. Bernadette Meehan, a former U.S. ambassador to Chile, succeeded Maryana Iskander as CEO in January 2026. ADL welcomes the arrival of a new CEO at a critical moment for the platform.

The Foundation’s 2025–2026 Annual Plan and the accompanying Product & Technology OKRs (Objectives and Key Results) already make public commitments to anti-abuse infrastructure: a working incident reporting system across all wikis, strengthened sockpuppet and ban-evasion detection, and an AI-assisted moderation tool, Automoderator, rolling out across multiple language editions. These are meaningful investments, but they address a different vulnerability than the one this campaign exploited. Of ADL's five structural recommendations from 2025, only one, evaluating Wikipedia's tools against coordinated inauthentic behavior, has a partial counterpart in this announced work. 

The four recommendations below remain the priority.

For the Wikimedia Foundation: 

• Institute private identity verification for editors operating in designated contentious topic areas, where the identity of these editors would be shared with Wikimedia directly but not made more visible publicly. This could take the form of only collecting the information that is needed to confirm an account belongs to one real person and putting in place other controls such as deleting information once confirmed or limiting access to a small number of staff. This would be designed to deter foreign state actors and coordinated groups whose anonymity is currently the central asset of any influence campaign on the platform. 
• End anonymous IP-address editing on sensitive pages, where the cost of bad-faith activity is highest and editor accountability lowest. Requiring a registered account would still preserve user privacy, given the pseudonymity of all usernames on the platform, while also giving every edit on these pages a persistent record that administrators can review and assess for evidence of ban evasion, if any.
• Require merit-based closure of controversial content disputes by vetted closure editors, rather than the de facto majority-vote outcomes that have allowed organized minorities to drive consensus on contested pages; 
• Create a dedicated Foundation task force on antisemitic bias, modeled on the Foundation’s successful 2021 intervention against Holocaust revisionism on the Croatian and Serbian Wikipedias.

On Remediation: 

• Commit to a structured content audit of articles substantially edited by editors who have been topic-banned or site-banned for non-neutral conduct, beginning with the six flagship articles identified in the Iskandar323 case study in Finding 1 above. The audit should assess removed source material and altered framing and should restore sourced content and balanced wording where CTS, INSS, or independent researchers have documented bias. Sanctioning an editor addresses future harm; without article-level remediation, that editor's distortions continue to seed search results, AI training data, and downstream reference works indefinitely. After an editor was banned in 2012 for systematically misrepresenting sources across history-of-science articles, editors compiled every article edit the account had made and traced, page by page, which of its material survived, so each could be checked and corrected. The same standard should apply here.

For search engines and AI developers: 

Wikipedia powers a large share of search-result snippets, knowledge panels, voice-assistant answers, and LLM pretraining corpora. ADL urges search engines and AI developers to:

• Disclose where Wikipedia is the source of an answer.
• Consider editorial caveats on contested topic areas.
• Refresh articles in their indexes after documented remediation rather than relying on stale captures.
• Develop bias-detection benchmarks specific to topics where coordinated manipulation has been documented.

The right form of these interventions is an open question we believe the AI and search industries should be addressing now, in dialogue with platforms like Wikimedia and with civil society.

For policymakers: 

• Continue oversight hearings on bias and antisemitism on Wikipedia.
• Create a standing task force to monitor coordinated content manipulation on user-generated platforms.
• Convene academics, technologists, Wikipedians, and civil society leaders to develop concrete policy options.

For the public and the press: 

• Recognize that a chatbot's answer, an AI search summary, or a knowledge panel may be repeating Wikipedia's framing on a contested topic without flagging it as such.
• Verify citations and quotes rather than assuming a paraphrase or an AI-generated answer has preserved accuracy and context.
• Document biases, where seen, on the ADL Antisemitic, Bias or Discriminatory Incident reporting page.

Methodology

This report draws on five datasets, each built from Wikipedia’s public records and each answering a different question. All five were assembled between July and August 2026 and are described below, with the capture date for each.

Sources and provenance

Revision data was retrieved through Wikipedia’s public MediaWiki API at en.wikipedia.org/w/api.php. Each pull was logged to a manifest recording of the query issued, the retrieval timestamp, the number of revisions returned, and a SHA-256 hash of the payload, so any figure here can be traced to the pull that produced it.

Sanction status of accounts was checked against Wikipedia’s public block logs, published Arbitration Committee decisions, and the Wikipedia:Editing restrictions page. Statements about what an article says today were checked against the live article, and the date of that verification is given in the text.

The persistence audit

CTS retrieved the full public contribution history of the account Iskandar323 and isolated its edits to six articles: “Israel,” “Gaza war,” “October 7 attacks,” “Hamas,” “Zionism,” and “Jews,” the highest-trafficked English-language entries in the topic area.

The audit covers the account’s full editing history on those six articles, not only its post-October 7 activity: 485 edits. Of these, 225 fall after October 7, 2023, 85 of them in October 2023 alone. Of the 485, 398 were assessable: an edit counts as assessable when its diff (from the word difference) can be retrieved and compared against the current article. Page moves, edits to revisions since deleted, and edits whose diff could not be resolved were set aside.

Each assessable edit was compared against the live article in August 2026. An edit counts as persisting when the change it made is still present: text the account added is still in the article, or text it removed is still absent. Across the six articles, 362 of 398 assessable edits persist, roughly 91%. Persistence is reported per article as well as in aggregate because it varies between them.

An analysis published in May 2026 by Dr. Shlomit Aharoni Lir of Tel Aviv University's Institute for National Security Studies examined the same account by a different method. The two studies measure survival differently and report different aggregate rates. They are reported here as separate findings, not combined into one.

The two-window comparison

This analysis compares two equal windows of six months, one ending and one beginning at the March 18, 2025, publication of “Editing for Hate,” across the 30 accounts identified in the research behind that report. Four measures were computed for each window: total article-namespace edits to Israel- and Palestine-related pages; tandem edits, meaning two or more documented accounts editing the same article within one hour of each other; group messaging, meaning user-talk messages exchanged between documented accounts; and average daily edits by the five most prolific accounts.

Total editing fell 27%. Tandem editing fell 88% and group messaging fell 95%.

The 27% fall in total editing reflects a smaller group, not less activity by its members. Several accounts were banned in early 2025, and others stopped editing, so the group lost roughly a third of its members; those who remained were editing slightly more than before. Tandem editing and group messaging fell much further, by 88% and 95%, more than a smaller group can explain.

The continued-activity pull

Figure 2 rests on a separate pull covering four accounts from the original 30 that were never sanctioned, from March 2025 to July 2026. Roughly 26,000 edits were retrieved. Of these, 2,442 were in the article namespace, matched a topic rule and fell inside the window; 1,145 of those were classified as maintenance and excluded, leaving 1,297 substantive edits.

Topic and maintenance classification were applied mechanically by keyword matching, against the page title and the editor’s own edit summary respectively. The expressions are recorded in the figure codebook alongside an edit-level file recording every classification decision, so any subset can be re-coded by hand. Talk-page edits are excluded. The two classifiers err in opposite directions: the topic filter misses some relevant edits, while the maintenance filter counts some vaguely described maintenance edits as substantive. The plotted totals are therefore approximate, and the edit-level file allows either error to be checked.

The convergence survey

Findings 3 and 4 rest on a forensic review of editing records conducted for this report, captured on July 9, 2026.

The article set was not chosen in advance. For each surveyed account, the 500 most recent article-namespace edits were retrieved, and any page edited by two or more surveyed accounts entered the dataset as a shared article. The map of overlap therefore emerged from the accounts’ own histories.

The surveyed cohort is 19 accounts: eight editors referenced in “Editing for Hate” who were never sanctioned and remain active, and 11 accounts not previously documented by ADL. Two accounts present when the dataset was first built were under topic bans at the time of the survey and were removed. Because a shared article is defined by the presence of two or more surveyed accounts, that test was re-applied after the removal; ten articles then carried edits from only one remaining account and fell out. The final dataset is 229 shared articles and 565 editor-article records, each traceable to a dated, public revision ID.

Each of the 229 articles was hand-coded for topic. An article is core when its subject is the Israeli-Palestinian conflict or its history directly; adjacent when the subject is Israel, Jewish history, or the wider region without being about the conflict; and unrelated otherwise. 167 of the 229 are core.

The October 7 page histories

For the case study in Finding 4, CTS retrieved the complete revision histories of four pages and their discussion pages: “October 7 attacks,” “Sexual and gender-based violence in the October 7 attacks,” “Screams Without Words” and “Screams Before Silence.” That is 9,576 revisions, 4,597 to the articles and 4,979 to the discussion pages. Contributor rankings cited in that section are counts of edits to the named page across its full history, not within a window.

Definitions

Substantive edit: an edit that changes article content, as distinct from a maintenance edit repairing a link, adjusting a category, fixing a typo or applying a template.

Shared article: a page edited by two or more of the 19 surveyed accounts.

Recent: an edit made within the 12 months before the July 9, 2026 capture.

Last word: holding the most recent edit to an article among the surveyed accounts. It is not a claim that no other editor has touched the page since.

Both-cohort article: a shared article edited by at least one account from each group. There are 142.

Handoff: a both-cohort article where every original-cohort account’s most recent edit is more than 12 months old while a newly identified account has edited within 12 months. There are 17. The reverse pattern occurs once.

Wikivoice: text written in the encyclopedia’s own factual voice, as distinct from text attributed to a named source.

Original cohort and newly identified: the original cohort is the eight accounts from the research behind “Editing for Hate” that also appear in the 2026 survey. Newly identified means new to ADL’s documentation, not new to Wikipedia; several of these accounts have edited for years, one since 2012.

Verification of named accounts

Named accounts carry the highest risk of error in a report of this kind. Sanction status was checked against the public block log and the relevant arbitration decisions for each. Every diff cited in the text was opened and read against the live article rather than relied on from a summary or a byte count.

Accounts whose activity could not be verified to that standard, or whose editing on inspection proved predominantly maintenance work, were dropped from the profiles even where they remain in the aggregate counts. One surveyed account exercised Wikipedia’s courtesy-vanishing option during the reporting period; its edits remain in the public page histories under a renamed-user handle and remain in the dataset.

Limitations

1. Shared editing is not proof of coordination. Active editors in a contested topic area will overlap on the most contested articles. The overlap map describes where the surveyed accounts are active. It is not evidence that they act in concert.
2. The survey is a snapshot of July 9, 2026. Accounts active then may since have stopped; accounts dormant then may since have resumed.
3. The 500-edit cap compresses high-volume accounts. An account making thousands of edits a year is observed over a shorter span than a less active one, so recency measures are not directly comparable between accounts of very different volumes.
4. Keyword classification errs in both directions. Topic-area articles whose titles carry none of the listed terms are missed, which undercounts, and maintenance edits described vaguely are counted as substantive, which overcounts.
5. A decline in visible coordination is not necessarily a decline in coordination. Talk-page messaging is by definition the observable form. Its near-disappearance is consistent with editors having stopped coordinating and equally consistent with their having moved to channels that leave no public record.
6. Persistence measures survival, not correctness. That an edit still stands says nothing about whether it was a good edit. The finding is that the encyclopedia has not revisited this work.
7. Sanction records are incomplete by design. A topic ban is a conduct restriction and leaves no block-log entry, so an absent entry cannot establish that an account was never restricted. Counts of sanctioned editors here are minimums.
8. The two windows were built to comparable but not identical specifications. The comparison re-runs a dataset constructed for the earlier report; cross-validation confirms the same direction and comparable magnitudes, not identical construction.

This report documents editing behavior and its effect on article text. It does not assert that any surveyed account is acting on instruction, is coordinating off the platform, or is anything other than a volunteer editor, and it makes no claim about the identity, nationality, employment or affiliation of any account. It does not claim that any individual edit is illegitimate under Wikipedia’s rules. Many of the edits examined here are defensible under those rules, and several cite genuine policies correctly. The finding is the pattern the edits form and the state in which they have left the articles.

The analysis scripts, per-edit output files, and codebooks underlying each figure exist in full, including a record-level file for every dataset above in which each edit, its classification, and the reason for any exclusion can be inspected. These materials, the list of 229 shared articles, and the provenance manifests for each pull are available to researchers on request.

Appendix