Grokipedia Stopped Listening.
The Machines Kept Citing It.
On April 24, 2026, the encyclopedia xAI built to be written by machines quietly stopped processing corrections from people. Thirteen thousand and two edits now sit frozen “in review.” No announcement. No public explanation. And the AI systems that cite it as a reference — roughly 356,000 citations by Ahrefs’ March 2026 count — have shown no sign of noticing. Neither, most likely, have you.
On October 27, 2025, xAI launched an encyclopedia written by a machine. Grokipedia arrived positioned as the corrective to Wikipedia: the reference layer rebuilt for the AI era, drafted by the model at machine scale. Human beings kept exactly one lever: the correction channel. Find an error in the article about your field, your firm, or you; submit the edit; the record improves. On April 24, 2026, that lever quietly stopped working.
The finding is specific, and it is not a rumor. Renée DiResta of Georgetown and Ronald Robertson of Stanford published the investigation in Lawfare on August 5th: across 34,519 Grokipedia pages, 13,002 human-submitted edits sit frozen “in review,” and the freeze dates cleanly. On Grokipedia’s own entry for ChatGPT, twelve edits submitted on April 24th were approved the same day, while every edit submitted afterward — May, June, and July — still sits “in review.” The pattern, the researchers write, holds across topics and traffic levels. Futurism corroborated the freeze independently the next day.
Two details turn an outage into a structural finding. First, the audit trail no longer describes the record: a mass rewrite in mid-March broke the text anchors corrections attach to, and Grokipedia’s logging system retroactively reclassified previously accepted edits as rejected — while, in cases the researchers reviewed, the changes “appear to have been incorporated into the articles anyway.” Second, the freeze changes nothing downstream. An Ahrefs analysis published in March counted Grokipedia surfacing in roughly 356,000 citations across AI systems, most often in ChatGPT and Google’s AI Mode. That is a fraction of Wikipedia’s footprint, and enough to reach rooms that will never visit the site itself. Nothing in an engine’s behavior suggests it audits a reference’s governance before borrowing from it. Nothing suggests the systems citing Grokipedia in August are citing it any differently than they did in March. Edition No. 037 filed a record the machines can’t read: authority sealed in pixels. This is the inversion: a record they read fluently that no one can currently fix. And somewhere in that citation graph, the machines may already be wrong about you. This is the queue your correction would enter.
What a frozen channel actually breaks
A filed correction is a receipt, not a result. The submission interface kept working for months after the review behind it stopped, still producing timestamps and the status “in review.” Thirteen thousand and two receipts currently stand where results are supposed to be. Whoever filed the first of them in April is still waiting. Not rejected. Not approved. Waiting.
Citing systems inherit the freeze without knowing it. They see a page, a timestamp, a citation count, and never the review queue behind it. Whatever an entry got wrong in late April rides into every downstream answer that leans on it, corrected at the source or not. The fix never converts into machine-readable trust.
The log is not the event. Edits recorded as rejected persist in live articles; edits recorded as pending have sat unreviewed for three months and counting. When the record of the process contradicts the process, checking the log stops being a defense. The only reliable test of what a system says is what it says — read directly, on a schedule.
The mechanism, not the motive
Lawfare and Futurism both stop short of a motive, and so does this edition. In the researchers’ own words: “We do not know whether xAI is preparing a major revision or has altogether abandoned active development of the project.” As of this edition’s release, no public explanation for the pause had been identified. A company is entitled to rebuild its review pipeline. The people behind 13,002 pending corrections were simply never told the pause was a pause. The mechanism is the story: a correction channel died silently, and the citation layer above it gave no sign of noticing.
The general law is not about xAI. Every reference layer you don’t control carries a correction channel you don’t control, and that channel’s health is invisible from your side of the counter. What the machines say about you is not installed once and finished; it drifts, and the drift doesn’t announce itself. That is why ongoing signal maintenance is the sixth phase of Answer Engine Authority™, not a support plan bolted on afterward. The signals you control are the only ones with a working maintenance schedule, and detection, knowing what changed and when, is the only guarantee anyone can honestly sell. Anyone whose record was wrong on Grokipedia in April is still wrong in every answer that borrowed it since. And if your record is wrong somewhere you don’t control, filing the correction proves you tried. It doesn’t prove the record changed.
Renée DiResta (Georgetown) and Ronald Robertson (Stanford), “Grokipedia Stopped Reviewing Edits in April. It Didn’t Tell Anyone.” — the primary investigation: 34,519 pages, 13,002 edits frozen “in review,” the April 24 cutoff, and the retroactive reclassification of accepted edits — Lawfare, August 5, 2026. Independently corroborated the following day — Futurism, August 6, 2026.
The ~356,000-citation figure is from an Ahrefs analysis published March 2026, as reported within the Lawfare investigation — the count is Ahrefs’, pre-dates the April freeze, and is a fraction of Wikipedia’s citation footprint by the same analysis. Grokipedia launched October 27, 2025. No public explanation for the review pause had been identified as of this edition’s release.