Were Nancy Guthrie's Ransom Notes Written by AI?

No public determination establishes that Nancy Guthrie’s ransom notes were written by AI. In the Pima County Sheriff’s Office investigation into the Nancy Guthrie ransom notes, the available public record supports a narrower and more useful answer: AI authorship is attributed speculation from former investigators and media commentary, while the FBI’s documented basis for discounting certain notes rested on operational signals.

Reuters reported, citing an anonymous source, that the FBI had determined three kidnapping notes were fake; the same report described an operational test in which investigators made a small cryptocurrency deposit into the wallet named in the first note and the money was never withdrawn.[1] FBI Phoenix’s public statement did not go that far. It said some notes were “extortion attempts without legitimacy,” while other ransom demands “may potentially be legitimate.”[2][3]

That distinction matters. A wallet that receives a deposit and is not emptied may help investigators decide a demand lacks practical connection to a kidnapper. A same-IP connection between messages may support treating two notes as linked. Neither fact proves that the words were generated by an AI system.

The evidence status is not the same for every claim

The cleanest legal reading is not that Reuters was wrong or that the FBI statement was evasive. It is that they sit at different evidentiary levels. Reuters supplied anonymous-source reporting and a concrete wallet detail. FBI Phoenix supplied the narrower public position. A lawyer repeating the stronger version needs the underlying record, not a headline approximation.

The same-IP fact is more useful than a stylistic hunch because it identifies something investigators could act on. CBS reported that two February notes were linked to the same IP address.[4] That does not answer who drafted the language. It does support a more limited inference: at least some communications may have shared infrastructure or a sender environment.

The wallet detail works the same way. If a ransom note names a wallet, investigators can test whether anyone with access to the scheme moves funds. Reuters’ reported deposit that was never withdrawn is an operational signal, not a linguistic one.[1] It may help discount a demand without proving whether the writer was a person typing alone, a person using a chatbot, or someone copying text from another source.

How “may be AI-assisted” becomes public shorthand

The AI theory entered the public record through commentary, not through a disclosed forensic report. NewsNation reported former FBI agent Jennifer Coffindaffer’s view that the ransom notes were written by AI, and it also included former FBI agent Andrew Black in the discussion of the notes.[5] That is a real public claim. It is not the same thing as a disclosed FBI authorship determination.

The caution inside the same reporting is easy to lose. KOLD news director Jessica Bobula was quoted asking, “is it assisted by AI?” and adding, “It’s very difficult to tell.”[5] That is not a finding. It is a caveat about uncertainty, and it should remain a caveat when the claim moves into a memo, filing, headline, or client alert.

Former-agent commentary can be valuable. Investigators develop instincts about threats, hoaxes, pressure tactics, and linguistic staging. The problem begins when an attributed instinct is laundered into an institutional conclusion. “A former FBI agent believes the notes were AI-written” and “the FBI found the notes were AI-written” are not paraphrases of each other. One preserves attribution. The other invents a forensic status the public record does not supply.

This is the same evidence-label problem that appears in reverse when parties blame AI for inconvenient evidence. In the Fukuoka bribery matter, the relevant risk was treating an AI-deepfake defense as more than the record supported; here, the risk is treating an AI-authorship accusation as more than the record supports. The direction changes, but the discipline is the same: preserve the status of the claim. See the site’s analysis of the Fukuoka bribery AI-deepfake defense for the mirror-image pattern.

The case timeline, only as far as it bears on the note question

CBS has separately tracked the broader disappearance timeline and the release of ransom-note materials in the Nancy Guthrie case.[4][6] For present purposes, the important point is not the full emotional arc of the investigation, although that remains the setting: Guthrie is missing, the case is unresolved, and ransom communications are being assessed under conditions where a false public label can misdirect attention.

The Pima County Sheriff’s Office’s July 31 release, as covered by CBS, referred to the writer’s “unique linguistic style” and included a redacted line in the released material.[6] Linguistic style can be a legitimate investigative lead. It is not, by itself, a public AI-authorship finding.

The Department of Justice plea also needs to stay in its lane. Derrick Callella pleaded guilty in the District of Arizona to harassment-related counts in connection with ransom-demand text messages sent on February 4, with sentencing scheduled for September 10.[7] The DOJ release concerns those February 4 text messages; it does not resolve who authored the media notes or whether any note was generated by AI.[7]

That is not a technicality. Pleas often become magnets for adjacent theories. Once someone has admitted sending a ransom-related communication, it becomes tempting to let that plea clean up unresolved authorship questions around other communications. The public DOJ record does not do that work here.

A verification standard for lawyers using an “AI-generated” label

An “AI-generated” label is an evidence claim. It should be handled like one. Before that label appears in a legal filing, risk memo, client alert, headline, or internal investigation report, the reviewer should identify the level at which the claim exists.

• Public forensic finding: a disclosed lab, agency, or qualified examiner states the method and conclusion.
• Operational investigative basis: investigators acted on wallet activity, IP linkage, payment behavior, device logs, account records, or other non-linguistic signals.
• Attributed expert speculation: a named expert, former investigator, consultant, or analyst gives an opinion without a disclosed forensic method.
• Media commentary: a report summarizes, paraphrases, or sharpens the claim for public consumption.
• Unresolved conflict: public sources point in different directions or speak at different levels of certainty.

The Guthrie notes currently fall into several boxes at once. Some notes have been publicly described by FBI Phoenix as illegitimate extortion attempts.[2][3] Reuters reported a stronger all-three-fake conclusion through an anonymous source.[1] The wallet and IP facts are operational indicators.[1][4] The AI-authorship theory remains attributed commentary.[5]

That sorting exercise is not pedantry. It determines what can be said without overstating the record. “Investigators had operational reasons to doubt certain notes” is a different sentence from “the notes were AI-written.” The first can be supported from the public materials. The second cannot, unless the writer has access to additional investigative or forensic material not currently public.

Ask what method produced the AI conclusion

If the source of the AI claim is a detector, the next question is not whether the detector sounds confident. It is whether the method is identified, whether the text sample is adequate, whether there is a comparison set, whether human authorship alternatives were tested, and whether the tool’s error profile is known.

The University of San Diego Legal Research Center’s AI-detector guide collects documented false-positive and false-negative concerns in AI-detection tools.[8] That record does not mean all detector use is worthless. It means a detector result, if one exists, is not a self-authenticating fact. It needs corroboration from metadata, account records, drafting history, device evidence, admissions, or some other record-based support.

The same verification discipline applies to generative-AI outputs more broadly. In the site’s Hank Green analysis, the risk was unverified AI-assisted research being treated as reliable legal support; the lesson carries over here because a detector-adjacent claim is still a claim that must be independently checked. See Hank Green’s ChatGPT research red flag.

Do not let operational proof smuggle in authorship proof

Operational corroboration can justify discounting a document without identifying how the document was drafted. A ransom wallet that sits untouched is a payment-behavior clue. An IP overlap is a network clue. A “unique linguistic style” may be a drafting-pattern clue. None of those facts, standing alone in the public record, identifies an AI system as the author.

This is where legal writing often goes wrong. A cautious sentence says the notes were reported as fake, or that investigators had reasons to question them. A careless sentence says AI wrote them. The careless version sounds modern, decisive, and searchable. It also adds a fact that has not been publicly established.

Keep allegation-only status visible

A status flag should travel with the claim every time it is repeated. If the source is NewsNation’s former-agent commentary, say that. If the source is Reuters’ anonymous-source report, say that. If the source is FBI Phoenix’s public statement, use the narrower language FBI Phoenix used. The point is not to drain the sentence of meaning; it is to keep the reader from mistaking one category of information for another.

That discipline is also why allegation-only records should be labeled as allegation-only even when the story is high-profile. The site’s Ariana Grande hacker lawsuit monitoring record uses the same kind of status control: it does not imply an AI tool is involved where the pleaded or public record does not support that implication.

The safe formulation as of Q3 2026

A careful formulation would read: public reporting and former-investigator commentary have raised the possibility that Nancy Guthrie ransom notes were AI-written or AI-assisted, but no public forensic determination establishes that conclusion. The FBI’s public statement and the reported investigative details support a narrower point: some ransom communications have been treated as illegitimate or suspect based on operational indicators, while the broader kidnapping investigation remains unresolved.

That is the standard lawyers should apply unless the underlying investigative or forensic record changes. “AI-written” remains an unverified assertion in the public record, not a finding.

References

1. FBI determines Nancy Guthrie kidnapping notes to be fakes, source says — Reuters, July 1, 2026
2. FBI determines 3 ransom notes in Nancy Guthrie disappearance were fake: report — KTUL
3. FBI Confirms Some Nancy Guthrie Ransom Notes Are Fake — But Others Are Still Being Investigated — Entertainment Weekly
4. Timeline of Nancy Guthrie's disappearance as search intensifies — CBS News
5. Former FBI agent: Guthrie ransom notes written by AI — NewsNation
6. Nancy Guthrie ransom notes released — CBS News
7. California Man Pleads Guilty to Harassment in Connection with Ransom Demand — U.S. Attorney’s Office, District of Arizona
8. Generative Artificial Intelligence (AI): AI Detectors — University of San Diego Legal Research Center