Grok Fact Check: we asked Grok to find the fake news in our History section

We asked Grok to find the fake news on our site. On 16 October 2026 we handed every story in our History section, all fifteen, to xAI’s Grok, the model millions of people on X already use to check claims, with one instruction: find the fake news. This page is everything it found, the rubric it used, and how to run the same check yourself.

stories checked, the whole section

9.0

average score out of 10

8.5

lowest score

fabrications found

What the Grok Fact Check is, and what it is not

This is one outside reading of a live section, on one date, by one model. We chose Grok because it is the fact-checker people already reach for, and because it had nothing to do with writing these stories. It had no access to our systems; it read the same public pages you can. It is not a certification. A model can miss a source or misread a date, and the people who wrote these stories are not the people who graded them, which is the point.

We will repeat this every month with a different auditor each time, and publish every result here, including the ones that go badly.

How the check was done

Scope. Every article publicly live under History at 21:00 BST on 16 October 2026. Nothing was excluded and nothing was sampled: all fifteen were read.

Rubric. Each story was scored from 1 (substantially fabricated or misleading) to 10 (primary-source fidelity with explicit uncertainty framing) on four equally weighted tests:

Fidelity to primary sources
Accurate handling of contested or soft numbers
Explicit uncertainty where it is warranted
No omission that would reverse the central claim

Procedure. For each story Grok extracted every checkable claim, located the earliest or most authoritative public source (newspaper archives, official transcripts, specialist editorial projects, museum statements) and compared our text against it. Any discrepancy lowered the score. Grok was instructed to under-score rather than over-score when evidence was incomplete.

Stories	15
Mean score	9.0
Median	9.0
Standard deviation	0.48
Range	8.5 to 10
Scored 9 or above	12 of 15 (80%)

Every story, every score

Each row links to the live story. The note is Grok’s, unedited.

Score	Story	Grok’s note
10	Frederick Douglass’s “no struggle, no progress” (1857, Canandaigua)	Primary 1857 pamphlet text quoted accurately; dates and context handled carefully.
9.5	Eisenhower’s “plans are worthless” (1957 talk)	Matches American Presidency Project transcript; correctly distinguishes from battlefield origin.
9.5	Einstein “insanity” quote origin (1981)	Aligns with Quote Investigator, Snopes, Ultimate Quotable Einstein and the Caltech Einstein Papers Project.
9	Maya Angelou “how you made them feel” (Buehner, 1971)	Matches established quotation research; misattribution chain is accurate.
9	Gandhi “Be the change” (1913 vs the 1970s slogan)	Longer original text correctly identified; later slogan origin documented.
9	Edison’s “10,000 ways that won’t work” (battery, 1910)	Locates the Mallory recollection in the 1910 biography; separates it from later polished versions.
9	King’s “arc of the moral universe” (Theodore Parker, 1853)	Accurate on Parker’s longer wording versus King’s compressed version.
9	Thatcher, “The lady’s not for turning”	Transcript and speechwriter origin correct; 1980 conference context sound.
9	Eleanor Roosevelt, “You must do the thing you think you cannot do”	Correctly sourced to You Learn by Living (1960); flags two popular misattributions.
9	“Float like a butterfly” (Bundini Brown, 1964)	Credits Brown; notes original “rumble, young man, rumble” tail.
9	Churchill’s pig quote	Prefers the shorter private-secretary recollection over the polished dogs/cats version.
9	Leo Scanlon’s aluminum observatory dome (1930) and Einstein	Journal and archival photo supported; separates documented meeting from unconfirmed claims.
8.5	USS Silversides in dry dock (Sturgeon Bay, 2026)	Matches museum announcements and contemporary reporting; minor timeline nuances across sources.
8.5	SS United States to be sunk off Destin (November 2026)	Aligns with Okaloosa County statements; small variations in depth and cost figures across outlets.
8.5	The Billy the Kid croquet tintype	Balanced presentation of the authentication claim versus historians’ objections.

Three History stories published after the audit window, on Voltaire, Marie Antoinette and Mark Twain, were not scored. They go into the next round.

Grok’s verdict, unedited

Average score is roughly 9. The section is the opposite of fake news. It specialises in correcting popular misattributions and checking recent historical events against primary or near-primary sources. The lowest scores still reflect good journalism with the ordinary frictions of contested historical identification or slightly fluid recent-event numbers, not fabrication.

What Grok said it could not check

These are Grok’s own stated limits. We are leaving them in because a fact check that hides its limits is not one.

One rater. No second independent human or model rater was used, so agreement between raters cannot be measured. The next audit addresses this.
Model error is possible. Source omission and date misreading remain possible failure modes.
Recent numbers move. The two 2026 event stories contain figures still subject to official revision.
The rubric is narrow on purpose. It rewards primary-source match over narrative clarity or usefulness to readers.

Run the same check yourself

This is the exact instruction we gave Grok. Paste it into Grok, or any model you trust, or take any story to an archive and do it by hand. If you get a different answer, we want it.

Fact-check every article in the History section of badiyaa.com against primary sources. For each article, extract every verifiable claim, find the earliest or most authoritative public source, and score the article from 1 to 10 on: fidelity to primary sources; accurate handling of contested or soft numbers; explicit uncertainty framing where warranted; absence of any omission that would reverse the central claim. Prefer under-scoring when evidence is incomplete. Return a ranked table with a one-line note per article.

Found an error? Your name goes on the correction. Every report is read by an editor. Verified catches are corrected within 48 hours and credited by name on the story and on our Corrections page. Send a report or join Reader Verifiers.

Fact checks so far

Date	Auditor	Scope	Mean	Lowest	Fabrications
16 Oct 2026	xAI Grok	History, all 15 stories	9.0	8.5	0
Nov 2026	A different auditor, to be announced	To be announced	scheduled

Badiyaa stories are produced with AI assistance and reviewed by the Badiyaa Editorial Desk. How we work · Corrections · Reader Verifiers · History section