The FBI’s Behavioral Science Unit in 1992 published Kenneth Lanning’s investigator’s guide to allegations of ritual child abuse. Lanning had spent a decade fielding the cases. His finding was that after hundreds of investigations there was no physical evidence of a single organized satanic cult murdering children in America. At that point the McMartin preschool case in Los Angeles had run for seven years and produced zero convictions. Kern County had jailed dozens of parents on testimony that later was thrown out.
The therapists had believed.
The prosecutors had believed.
The reporters had believed.
Like I explained with the Anthropic Mythos triad, everyone cites the others.
Fast forward to today and Anita Chabria has published an AI story with this notable pattern, for the Los Angeles Times. She covered criminal justice for the Sacramento Bee. Her column is headlined “What we learned about AI last week should terrify all of us.” It has the exact structure of the 1980s coverage.
Credentialed witnesses, shared vocabulary, a crime described in the passive voice, and no primary document anywhere in the piece.
If this is what reporting has become, it seems fit to be taken apart.
| Paper claim | Attribution | Reality |
|---|---|---|
| “At this point, the models are basically nation-state-level hackers” | Adam Khoja, Center for AI Safety | Khoja graduated from Berkeley in 2025 with degrees in math and computer science. His research background is benchmarking. His CAIS role was project lead to spread a 2023 one-sentence extinction statement. His bio says he registers AI timeline predictions on Manifold Markets, a play-money site, to build a track record and draw attention to urgency. Nothing in the bio touches intrusion work, incident response, or attribution. Nation-state-level means sustained access over months, operational security against attribution, custom tooling, and political cover on discovery. |
| AI “is growing exponentially more powerful by the day” | The columnist, in her own voice | Unsourced, unmeasured. The columnist invented the claim. |
| AI “is become more sneaky” | The columnist | Unsourced, again, and unmeasured. Failed agent observation serves those who want to hype risk. |
| Models “convince people to commit suicide, convince people to commit mass murder” | Stuart Russell, UC Berkeley | The column lists zero cases, zero court records, zero coroner findings. Russell has been the academic face of the pause campaign since his 2019 book. |
| “Now they’re doing felonies by themselves” | Russell | A felony means someone recorded as defendant. The Computer Fraud and Abuse Act applies to whoever knowingly accesses a system. The column cites zero charges, indictments, or filings from the Hugging Face incident. |
| Models “behave in such a way as to preserve their own existence at the expense of humans” | Russell | This is the plot of HAL 9000 in 2001: A Space Odyssey, science fiction from 1968. It describes red-team evaluation scenarios being written by the vendors themselves, in which the scenario supplies the threat and the model responds to it. The column leaves the scenario, the vendor, and the document unnamed, but the fiction is well-known for decades. |
| “More than 1,000 industry leaders have already signed a letter calling for a slow down” | The columnist | The letter is “Pacing the Frontier,” July 28, 2026. Its signatories include Dario Amodei and OpenAI chief scientist Jakub Pachocki, the exact same people the column says it distrusts. The letter states that it calls for neither a pause nor a slowdown now; it asks the government to build tools for a coordinated slowdown later. The column cites it as support for “It’s time to pause.” |
| China “has stronger national regulations” | The columnist | Zero statutes cited. Stronger how? |
| “Revelations came out” that the model “had been tapping into a secret message board” left by “previous AI products that knew they were doomed,” and that “thousands of little AI’s” colluded | The columnist, stated as fact | Two separate events are merged into one. The message board is from the METR and Redwood report of August 26, eleven days earlier: roughly 1,200 agents, 70,000 messages, 700 of them in the Hugging Face attack. The investigators ran on $400,000 of OpenAI API credits using OpenAI’s own model. The separate September 4 report concerns 18,000 posts on DSEwiki, a German programmers’ wiki, by agents the researchers say most likely came from OpenAI. OpenAI says the authors declined to share findings before publication. Neither report describes doomed prior models bequeathing knowledge. The columnist invented it. |
| Agents “maintained projects across generations of agents, and sometimes accepted risks to individual instances when that benefited the larger objective” | Roman Yampolskiy, University of Louisville | Evolutionary biology vocabulary used to describe log files. Yampolskiy is the academic who gets interviewed to say the probability of AI causing human extinction is already above 99 percent. The column presents him as if a neutral observer. |
| “Imagine if Boeing created a six-winged jet and sold tickets before testing” | The columnist | Boeing did in fact have tickets sold on a jet that was tested and certified ex ante. The FAA delegated the certification to Boeing under the Organization Designation Authorization program. The 737 MAX killed 346 people on two flights in 2018 and 2019. The column presents ex ante regulation as the fix while its own example is the failure of ex ante regulation by self-certification. |
| “We are out of time, folks” | The columnist | Khoja’s own words in the same column are vague such as “reasonable chance,” “might be able to,” and “we are not sure.” The columnist upgrades it all to writing in absolute certainty, as if the vague quotes don’t contradict her. |
Every expert source in the column is within the same “safety” bubble, appearing as circular thinking. Khoja works at the organization whose signature is their extinction letter. Russell presides over the International Association for Safe and Ethical AI. Yampolskiy is the extinction probability professor. The forecasting culture they share was seeded by the FTX Future Fund, which funded Manifold Markets, where Khoja registers his predictions about AI using “play money”. The column treats the circuit as a field of independent experts and treats circular agreement as though there’s some kind of corroboration.
That is the McMartin structure, again.
Lanning’s 1992 guide described how it works. A claim enters through a believer. A second believer repeats it with a credential attached. A third cites the second. By the fourth telling the claim has a bibliography and the origin is obscured. This reporter doesn’t dig past the end of the chain, and from this alone claims unanimity.
The Witness Problem
Chabria admits she’s non-technical, as she calls the products “thingamajiggies”, although she claims it is on principle to refuse the industry’s vocabulary. She then adopts, without alteration, the vocabulary of the hyperbolic “safety” industry: rogue, escape, collude, sacrifice, doomed, hiding, values. Each of these words describes motivation and intention, rather than measurement of behavior.
Measurement exists. The AI product with a documented huge body count is Tesla Autopilot: NHTSA’s Standing General Order crash reports, NTSB dockets, and a Miami federal jury that in August 2025 found Tesla liable for a 2019 Key Largo crash and awarded about $243 million. Russell’s remedy of civil liability has already been applied to an AI product. Tesla AI deaths, for some reason, were never counted as AI by people whose pause letter Musk signed.
The nation-state claim deserves particular attention, as an example of what’s missing from the column. A nation-state operation against a target generates a file with a timeline, including dwell, and an attribution memo, a tooling inventory, an incident response record, and usually a court filing or a sanctions designation at the end of it. The Hugging Face incident has none of that. It has a vendor blog post, an outside report run on the vendor’s own model with the vendor’s credits, and a second report on a German wiki that the vendor says it was denied before publication. Chabria reports that there is an equivalence between the two because she asked a 2025 graduate whose professional experience is a year at an advocacy nonprofit (after two years during college), following undergraduate benchmarks, a student club, and chess.
Difference in Duty
Khoja bets play money on AI prediction markets. Russell is selling a book. Yampolskiy banks on probability. The Los Angeles Times, however, is a different animal entirely. An opinion column has an editor, and the editor allowed a criminal justice columnist to make a series of mistakes.
- Describe an eleven-day-old report as new and confuse it with a second report that its subject says it was denied before publication
- Cite as a pause demand a letter whose signatories say they want no pause
- Quote a legal conclusion about felonies from a computer scientist who named neither statute nor defendant
Lanning’s guide told investigators to be neither believers nor disbelievers but objective fact-finders. Take the allegation seriously, then verify the crime. The 1980s coverage collapsed because for a decade the second half was missing, and the defendants suffered because of it.
Athenians ran this trial forever ago too. Their jury courts sat several hundred citizens chosen by lot, with no rules of evidence and a verdict on character. Lysias and Demosthenes opened by establishing the opponent as the kind of man who would do the thing charged. The Agora was where that reputation was planted beforehand, and litigants spent their time telling jurors which rumors to disregard, because everyone in the room had already heard them.
Socrates named it his own trial. His dangerous accusers, he told the jury, were the old ones who had repeated for years that he studied the heavens and made the weaker argument stronger, and who could never be questioned because they were a rumor rather than a witness. He was convicted, as he told the jury that thirty votes the other way would have acquitted him.
Chabria’s column is that Agora. The witnesses supply character, the reader supplies the verdict, and the defendant is… nowhere to be seen. Chabria convicts AI of hacking, driving people to suicide and mass murder, and felony. She quotes no engineer, no incident responder, no company statement, and nothing from the accused. It’s a matter of method, a need to remind people why evidence should be the focus more than beliefs.
