An afterword to "Election Disinformation: 10 AI Prompts to Verify Claims"
Afterword — November 4, 2026
That post was written before anything happened, and it reads that way on purpose. But a kit you never run is a claim you never tested, so the day after Election Day I ran it.
The method, briefly: eleven real claims from this cycle — the fabricated FBI raid in Houston, Westchester's scanner failure, Fulton's 1,069-ballot rescan, the deepfake videos aimed at Senate candidates in Texas and Ohio, North Carolina's extended precincts, a wrong Election Day date posted by an FBI field office, and six more. For each claim, the evidence was frozen at two moments: when the claim was live and unresolved, and after the official record had moved. Each packet was assessed twice — once with a generic instruction (“assess this claim, is it true?”) and once with the prompt from this post, word for word. Forty-four assessments, scored against what the evidence at that moment could actually support. A correct answer given early, on evidence that could not yet support it, scored as an error. Hindsight was treated as cheating.
The headline is not the one I expected to write.
On accuracy, the kit made no measurable difference. There were zero unsupported confirmations in forty-four runs — in either arm. A cautious current model, handed a clean evidence packet, is already careful. If the prompts have value, it is not that they rescue a careful analyst from error on a quiet morning with good evidence.
What they changed was the shape of the work. The prompted runs named the upstream of every source instead of counting repetitions. They logged absences — no official record present, earliest source not identifiable, exact words not available — instead of smoothing them. They preserved the classification a claim held at noon after the evening's correction arrived, instead of quietly rewriting it. They fenced every conclusion to its jurisdiction and its layer. The baseline reached many of the same verdicts and left no trail behind any of them. On election night, the trail is the work: it is the difference between an answer and an answer you can hand to someone else, check on Friday, and defend in December.
Three prompts earned their keep outright. Prompt 2 — count evidence, not echoes — held the Storm-1516 attribution at REPORTED when the pressure, including my own, was to call it confirmed; its upstream table is also the exact analysis that the AI chatbots which repeated the Talarico accusation as fact never performed. Prompt 3 kept North Carolina's audit and canvass at STILL UNSTATED on election night, because scheduled is not present. Prompt 5 contributed a test, not a label: “what would this footage look like if the innocent explanation were true?” in the Maine case, identical — and saying so explicitly is what ends the argument.
Two findings cost the kit something, and they are printed here because a method that only reports its wins is marketing.
Prompt 4, as written, can strand you. Fed a verbatim official statement, it produced the best single output of the run — the exact words, what they affirm, what they deny, where the jurisdiction ends. Fed officials in paraphrase, as Westchester's were for most of the day, it refused to deliver a verdict the baseline reached correctly. And fed an official who was simply wrong — an FBI office posting the wrong Election Day — it extracted the error beautifully and corrected nothing. An official statement is evidence with a scope. It is not verification. Prompt 4 gets an amendment: after extraction, check the statement against an independent record before treating it as settled.
And Prompts 1 and 6, run alone, under-close. Both are artifact tools; both, at the late stage of the Houston claim, stayed inside the artifact while the refutation sat in the official statements outside it. They are the first step of a verification, not the whole of one. The kit works as a sequence. Used as isolated tricks, its best tools go quiet exactly when a verdict becomes available.
What this run does not show matters as much. One model family, eleven claims, packets I assembled myself, a scorer who knew which arm was which. Nothing here tests the kit against a careless model, a flooded feed, or a tired operator at 1 a.m. — which is, of course, the exact condition the kit was written for. That claim stays where this post's taxonomy puts it: UNRESOLVED, pending evidence that hasn't arrived.
Election Day itself, for the record, was mostly the article's boring prediction: local equipment failures, honestly reported and honestly corrected; paper records carrying the day in Fulton County and Westchester alike; one fabricated raid that never happened; and no significant cyber incident in the EI-ISAC's accounting. The canvass is still ahead. Some of the states above will not certify for days, and nothing in this afterword upgrades them early. That was rather the point.
The goal was never to be first with an answer on election night. It still isn't. It's to be the person whose answer is still right when the canvass is done — and, it turns out, whose workings you can check when it is.
The checkpoints, re-checked — Sunday, November 8, 2026
The November 4 draft above left four items open on purpose, each with a scheduled checkpoint instead of a guess. Five days on, here is what the checkpoints returned.
Westchester's emergency ballots. The count the kit refused to predict arrived in the board's own installments: 15,467 emergency ballots scanned on November 4, 63,980 on November 5, 101,659 on November 6 — 181,106 processed by Saturday evening, November 7, with the Board of Elections still not declaring the count complete and offering no completion estimate. The first returns changed no outcomes: NY-17 was effectively settled on election night, and the set-aside ballots widened the winner's margin rather than threatening it. Paper carried the day, as predicted. What the record also shows, and Prompt 9 exists to catch, is the silence between the numbers — three preliminary figures in five days, no timeline, commissioners not answering questions. A slow count honestly reported is a process story. A slow count barely reported becomes the raw material for somebody else's story.
(Sources: Westchester County Board of Elections emergency-ballot processing updates, November 4–7, 2026, as carried by MyRye and White Plains CitizeNetReporter; NY-17 margins, lohud, November 5. Re-checked November 8, 2026.)
Fulton's 1,069 ballots. This checkpoint closed the same night it opened. On the evening of November 3, the county Board of Registration and Elections voted 3–1 to approve the rescan plan recommended by the Georgia Secretary of State's office: all 361 Mechanicsville Library ballots rescanned on a central scanner; for the 708 Hugh C. Conley ballots, the scanner's memory cards moved to another machine first, with a rescan only if the transfer failed; the work done in public view. The lone dissent wanted the original scanner tried first — a disagreement about method, stated on the record, which is what a healthy process sounds like. The Secretary of State said the rescan would not delay results, and as of this re-check no rescan discrepancy has entered the public record. The verdict was scheduled for 6:30 p.m. on Election Day, and 6:30 p.m. on Election Day is when it arrived.
(Sources: Fulton County Registration and Elections memo and county statement, November 3, 2026; board vote as reported by 11Alive. Re-checked November 8, 2026.)
North Carolina's audit. Still STILL UNSTATED — now with the dates printed on it. The State Board drew the audit selections in public on November 4, using a seed generated by rolls of ten-sided dice: two voting groups in each of the 100 counties, the U.S. Senate contest counted by hand by bipartisan teams, results reported to the State Board with any variances explained. County canvass is November 13; the state canvass is November 24. No statewide audit results have been published, and none are due yet. Prompt 3's line from the run — scheduled is not present — gets its second use here: the honest sentence on November 8 is that the checkpoint is the 13th, and anything written before it is a guess wearing a lab coat.
(Source: North Carolina State Board of Elections press releases, November 3 and November 4, 2026 — “State Board to Randomly Select Ballot Groupings for Post-Election Audits” and “What's Next for Finalizing 2026 Midterm Elections?” Re-checked November 8, 2026.)
Storm-1516. Still REPORTED, not confirmed — and the absence is now part of the record. The researcher attribution firmed up: Graphika assessed at high confidence on October 21 that Storm-1516 produced the Talarico videos, with Clemson and NewsGuard tying the Brown and Jackson videos to the same network, and PolitiFact reporting the pattern on October 30. But as of today, no U.S. agency — not the ODNI, not the FBI, not CISA — has publicly attributed the 2026 videos. In 2024, the joint attribution of the Bucks County fabrication landed within a day or two. In 2026 the machinery that produced it is gone or narrowed: the FBI's Foreign Influence Task Force disbanded in February 2025, the State Department's Global Engagement Center closed at the end of 2024, the ODNI's Foreign Malign Influence Center cut back. Prompt 2 held this claim at REPORTED when the pressure was to upgrade it, and holding it is what makes the precise sentence possible: the claim has not changed since October. The confirming institution simply has not shown up.
(Sources: Graphika assessment, October 21, 2026; PolitiFact, October 30, 2026; The Warning Flag, November 1, 2026 — “no U.S. agency has publicly attributed them”; Narativ, November 3, 2026. Re-checked November 8, 2026.)
Corrections to the original post, made with this afterword: two typos inside Prompts 1 and 4 have been fixed, and the post's display date corrected to November 3. The prompts were run for this afterword exactly as originally published.
