Frequently asked questions

Last updated: September 2, 2026

Isn’t this a bias detector?

No. The activity score is direction-blind: it measures how much an article’s editing changed relative to that article’s own 36-month past — edit rate, text churn, revert-like edits, talk-page activity, editor concentration, attention-adjusted editing — not which way the text moved or whether anyone should have moved it. A cleanup drive, a flood of new sources, and a dispute all raise it. Controls behave the way controls should: for October 7, 2023, Ancient Rome scores 12 and Photosynthesis 4, while related articles Theodor Herzl (2), Balfour Declaration (1), and Israel (9) all come back typical. Activity is not proof of bias or misconduct. Read the methodology for what the number does and does not measure.

Who edited the article?

HistoryRewritten does not show that, by design. The service computes aggregate editor statistics only — number of unique editors, the share held by the most active one, three, and five, a concentration index, the anonymous share, the revert-like share — and never returns, stores, or displays usernames from revision metadata. The model that writes comparison summaries is instructed not to infer intent, identity, ideology, or misconduct, and the schema it must fill has no place for any of it. Wikipedia’s own history page is the public record if you need it; this tool will not become a way to build a case against a person.

Why is October 7, 2023 the default event?

Because that is where the project started. After October 7 the maker suspected that Wikipedia articles on Jewish and Israeli history were being systematically reframed — and did not trust himself to check that fairly by reading diffs with a conclusion already in mind. So he built a tool that could tell him he was wrong, with a model-free score, user-selectable event dates, and control articles that have to come back typical for the method to mean anything. The public dispute that prompted it — the co-founders’ statements, the ADL decision, the congressional inquiries and the arbitration record — is summarized with sources on the why page. It partly did tell him he was wrong: Zionism scored high, but Theodor Herzl, Balfour Declaration, and Israel scored typical. The event date is a parameter, not the product’s identity: presets include the invasion of Ukraine, the ChatGPT release, and the COVID-19 pandemic declaration, and any date since January 15, 2001 works.

Can I cite a report?

Yes. Every report names the exact revision it analyzed and the algorithm version (activity-v1), and every comparison shows both revision IDs with links to the exact revisions and the official Wikipedia diff. Use Cite this report at the bottom of any report for a plain-text or BibTeX citation that pins those identifiers. Article excerpts are © their Wikipedia contributors under CC BY-SA 4.0; when you quote them, keep the attribution via the revision links. Analyses are re-verified weekly and can change as the article does, so cite the revision and the access date, not just the URL.

Why does Ukraine score 1 after February 24, 2022?

Because the score is a sustained median shift, not a burst detector — and this is its most important limitation, so it comes first. The monthly components compare the median post-event month with the article’s baseline months. Across the many full months since February 2022, the median month for Ukraine does not stand far outside the article’s own baseline by these measures, so the score is tiny — whatever the timeline on its report shows about the weeks right after the invasion. The first-90-day window is reported in every analysis but enters the score only when no full post-event month exists yet. An acute-window component is a candidate for a future algorithm version. Until then: read the timeline and the acute window next to the score, and treat a low score on a famously busy article as a statement about medians, not about the event.

What does “typical” mean?

A score from 0 to 34: the article’s editing since the event looks like its own past. It does not mean nothing happened — edits continued, and the timeline on the report shows every month of them — only that the post-event months do not stand out against the article’s baseline by the six measures. A typical result on an article you were sure about is not the tool failing; it is the tool answering.

Does the AI decide the score?

No. The score is model-free: six components computed with median-based robust statistics against the article’s own baseline, weighted and summed, with every intermediate value returned so you can reconstruct it by hand. A language model is used for exactly one thing — summarizing the changed excerpts when you click Compare. It is never called on page load, has no tools or web access, receives the article text as untrusted data, must return JSON in a fixed schema, and must cite the evidence IDs of the exact before/after pairs it describes. A validator rejects summaries that cite unknown IDs; a deterministic fallback takes over if the model fails, and the response says so. The exact excerpts and the official diff are always shown beside whatever the summary says.

Is my data collected?

There are no accounts, no sign-ins, and no tracking cookies. Analyzing an article sends its title, language edition, and your chosen event date to the service, which fetches public revision statistics from Wikimedia. Traffic is measured with cookieless, aggregate analytics (page views, referrers, load timings) with no personal identifiers. The Chrome extension’s keyword matching runs entirely in your browser, and in silent mode it sends nothing until you open the popup. The privacy page has the full picture, including the short-lived, salted rate-limit counters that are the only IP handling.

How do I report a wrong finding?

Open a GitHub issue with the report URL and, for a comparison finding, the evidence ID (it looks like chg_0007) of the pair that is mis-aligned or misdescribed. Every report has a Report an issue link that opens the issue prefilled with the report URL, page ID, analyzed revision, and algorithm version. Confirmed alignment defects get a fixture test and a diff-pipeline version bump, which recomputes every cached comparison. The linked Wikipedia revisions and diffs are always the authoritative record. Issues: github.com/cheesejaguar/historyrewritten.org/issues.

Is HistoryRewritten affiliated with Wikipedia?

No. HistoryRewritten is an independent, unfunded, open-source project. It is not affiliated with or endorsed by the Wikimedia Foundation or Wikipedia. It reads the same public APIs anyone can, following Wikimedia’s published API etiquette (an informative User-Agent, paced and serialized requests, honoring Retry-After and maxlag, and caching so a popular article costs a handful of calls a week). Article text is reused under CC BY-SA 4.0 with attribution to the exact revisions. More on the about page.

Articles with fewer than 12 months or 20 edits of pre-event history say “insufficient baseline” instead of pretending to a verdict; the methodology explains both gates. The Chrome extension (Chrome Web Store) shows the same analysis on the article you are reading. Anything else: open an issue.