Allysia Finley audits her opponents’ money conflicts and never audits the man selling her answer. On October 9 the Journal’s opinion page ran “Why Did Technologists Become Doomers?” — a piece by an editorial-board member built almost entirely on one interview, whose thesis is that AI alarmism is a funded campaign and that the answer is legal liability plus cybersecurity, not regulation. Her sole named source is the founder of a six-month-old startup that sells AI supervision software. I built versions of this piece for twenty years. The single-subject profile is the workhorse of the opinion apparatus because the columnist gets to disappear — the source carries the argument, the byline carries the authority, and the seam never shows. Tainting the source instead of answering the claim is the other half of the craft, and it works on people. It does not work on claims, which is the only place it needs to. What follows walks through the piece as it deploys.
They promised us the future, and now they want us terrified of it. “I think short of the transporter, most things you see in science fiction are, in the next decade, the kinds of things you’ll see,” Bill Gates said in 2007. Geoffrey Hinton, a Nobel physics laureate known as the godfather of artificial intelligence, said in 2016: “People should stop training radiologists now. It’s just completely obvious that within five years, deep learning is going to do a lot better than radiologists.”
Last month Mr. Gates warned that AI could cause a billion deaths. Not to be outdone, Mr. Hinton has been speculating for a couple of years about the prospect for total human extinction at the hands of AI. Last month he told the BBC that it’s “not unreasonable” to put the likelihood of such an outcome at 10%.
How did boosterism give way to doomerism?
— opening paragraphs
The Clip File. Pull two quotes from a different decade, set them beside a later claim, and let the reader build the indictment himself — the reader always builds it stronger than you would have. But look at what the quotes are. Both are capability forecasts: science fiction arriving on a schedule, deep learning beating radiologists at reading scans. Neither is a statement about mortality risk. The column impeaches a mortality estimate with two missed capability estimates, as though a man who overestimated the near term has thereby forfeited the right to speak about the far one. Operators call that the one-strike rule — newsroom vocabulary I carried out of the cable years, not a published taxonomy — and the tell is always the same: the reader is never told whether the two forecasts and the third claim are the same kind of statement. They are not.
Then the title does the work before the first argument appears. Frame-engineered relabeling — the page’s signature move, a name I carry from the work rather than from any published catalogue — with the loaded question riding on top of it: JAQing off, the argument that arrives wearing a question mark. “How did boosterism give way to doomerism?” presupposes that concern about AI is a psychological collapse, a giving-way from boosterism into pathology. It does not ask whether Hinton’s 10 percent figure is right. It asks the reader to recognize the familiar figure of the prophet crying doom and file the warning accordingly. It is not a question. It is a frame with a question mark welded on, and it already settles that there is no case to answer, only a psychology to diagnose. Once you have agreed the man is a doomer, you never have to answer what he said. The label is the argument, and analysis was never scheduled. Everything after the headline is furniture.
How did boosterism give way to doomerism? I asked Zack Korman, a self-described former doomer who now believes that AI’s risks can be contained and aren’t so different from those of other technologies when they first emerged.
“I do think the risks are real and we will face serious AI risk challenges,” Mr. Korman, 34, says over Google Meet from Norway, headquarters of his AI cybersecurity startup, Embroidery. “There are threats and there are problems ahead. With that said, I don’t think banning anything is going to help.” He faults the doomers for having “managed to make the debate between safe and unsafe,” and he maintains that “a lot of their proposals are extraordinarily unsafe or also dangerous in other ways.”
— paragraphs 3–4
One source. One interview, over Google Meet, from the headquarters of his own company. This is the party-as-witness move — the interested party presented in the analyst’s chair, with the stake disclosed in a bio line and never connected to the argument. It runs in partnership with the identity-credential move, and the column needs both, because it cannot refute the doomers’ arguments and so produces a convert instead. The conversion story is the case. The testimonial does the work the analysis never attempts. Beside the conversion the column stacks a biography — the Edinburgh law degree, the Oxford master’s, the argument with a professor that ended with a man teaching himself to scrape the SEC website because he wanted to win. That builds a man who beats academics with code so the reader will trust a man who beats doomers with arguments. It does persuasive work without containing a single argument.
Watch what the conversion omits. Korman says banning anything won’t help and faults the doomers for making “the debate between safe and unsafe.” Nobody quoted in this column — not Hinton, not the 2023 signatories — has proposed banning anything. The column itself will print the statement asking that extinction risk be treated as a global priority. That is an exhortation to take a risk seriously; it is not a ban. The piece has converted a position nobody in it holds into the one its subject bravely rejects. A strawman with a byline.
And here is the suppressed variable, the one the whole apparatus is organized to keep off the page: the disclosure of what Korman actually sells is worse than late. “Embroidery” first appears in paragraph 4, inside the one-line bio, and the reader goes seventeen paragraphs before learning what it is — paragraph 21: “Six months ago he launched Embroidery, which uses AI ‘supervisors’ to monitor AI agents” — and further still before the column connects the product to the prescription. What it never does is ask the question the disclosure was supposed to trigger: a man whose company sells AI supervisors has a commercial interest in the conclusion that supervision rather than regulation is what’s required. That is not a reason to ignore him. It is a reason to price him. The column prices the other side’s funding ties at full weight and its own source’s at nothing. In the trade, a piece like this is called the brochure: it exists to move product while pretending to move minds.
Mr. Korman says many of these doomers have been sounding the alarm for years. “For the people in that doomer circle, this is like seeing the prophecy come true,” he says. “It’s like if on the third day, it rains and then on the fourth day there’s locusts and like, oh my God, now the world’s ending.”
Before the recent leaps in AI, leaders of the top lab companies in 2023 signed a statement averring that “mitigating the risk of extinction from AI should be a global priority, alongside other societal-scale risks such as pandemics and nuclear war.” Messrs. Hinton and Gates signed the statement, as did the professional climate alarmist Bill McKibben.
Mr. Korman traces the origins of AI doomerism to a 2014 book, “Superintelligence: Paths, Dangers, Strategies,” by Oxford philosopher Nick Bostrom. […] To illustrate the embryonic technology’s potential danger, Mr. Bostrom imagines a superintelligent machine that is given the objective of maximizing production of paper clips. Because humans control resources that could interfere with this goal, the machine might conclude it must eradicate humans. Ergo mass extinction.
— paragraphs 5–8, trimmed
Attribution of blame, running as ridicule: the opponents are handed a psychology instead of an argument. Somewhere in this column the case the alarmists actually make gets quoted at length. It isn’t here. You get their temperament, served up as a biblical-parade joke, and from a source whose credentials on the subject are a law degree, a master’s, and a startup six months out of the gate. The parable does real work: once the other side is a crowd waiting for rain, the reader’s job shifts from evaluating their evidence to diagnosing their pathology. That is Bandura’s advantageous-comparison mechanism in the ridicule register — make the target look like a delusional prophet and the audience never has to weigh what the prophet said. Ridicule arrives before rebuttal and never after.
Then Bostrom’s thought experiment gets flattened into a cartoon. The paper-clip maximizer is a pedagogical device from the alignment literature: imagine a machine pursuing a single objective so single-mindedly that it treats the species as an obstacle, to illustrate that intelligence and benevolence are not the same property. This column shrinks it to “Ergo mass extinction” and invites the laugh. The laugh is the rebuttal. A facial expression has been substituted for an argument.
Note the sneer phrase dropped into the middle of it: “the professional climate alarmist Bill McKibben.” A contempt label is doing the reasoning where a description of what McKibben does would have gone. Either those signatories are serious people whose warning deserves an answer, or they are deluded prophets whose warning deserves mockery. The column cannot have it both ways, and it chooses mockery because an answer is harder. The phrase is also the giveaway that this contempt is standing equipment rather than fresh analysis — the same apparatus has dismissed one generation of documented risk after another. When your rebuttal to a 10 percent extinction estimate is a joke about locusts, the joke is the tell.
Many of this ideology’s adherents have ties to large AI companies. Anthropic’s president, Daniela Amodei, sister of CEO Dario Amodei, is the wife of Holden Karnofsky, who co-founded two large effective altruist nonprofits. Anthropic received early funding from effective-altruism adherents including Sam Bankman-Fried, founder of the FTX cryptocurrency exchange, who is serving a 25-year term in prison after conviction on federal fraud and conspiracy charges. Mr. Bankman-Fried misappropriated billions in client funds. With so many effective altruists on the payroll, it’s little wonder that AI labs are filled with doomers.
— the effective-altruism funding passage
Taint by association, run as a chain: a book, then a philosophy movement, then a convicted fraudster, then “no wonder.” The move has a classical name — the circumstantial ad hominem, the arguer’s situation offered in place of the argument — and the syllogism the reader is invited to run is worth stating plainly, because it is ugliest when it is unpacked. Bankman-Fried was a fraudster. Bankman-Fried was an effective altruist. Effective altruists staff the AI labs. Therefore the people who worry about AI risk are frauds. Every link is a category error. Holden Karnofsky is named not because he has done anything wrong, but because his wife’s connection to Anthropic’s leadership and his movement’s one infamous donor make him available for association. The association is the conviction.
Nothing in the Bankman-Fried paragraph bears on whether extinction risk is real. Everything in it bears on whether the reader now feels contaminated by association. The chain’s destination is the people, not the argument. And I built a lot of this: it is effective, it is cheap, and its evidentiary value is exactly zero. Sam Bankman-Fried’s crimes are real and documented. What they are not is evidence about the probability of AI catastrophe. A fraudster’s money neither raises nor lowers a 10 percent estimate. The estimate stands or falls on the evidence, and this column never touches the evidence. A smear is what gets published when the evidence would not cooperate.
Finley runs the funding audit in one direction only. Her source’s product sits on the same axis as the policy position his testimony supports, and the audit stops at his company bio. Run the same associative test the other way and it convicts in a paragraph: the labs’ boards are stocked with the same donor class that funds everything in this town; the cybersecurity industry profits from the AI boom it claims to defend; the defense contractors who will build the China-race arsenal this column endorses have a documented century of cost overruns and captured procurement. The column does not run that test, because the test was never the point.
His combative streak shows in his frequent brawls with doomers on social media. As he notes in a social-media video, experts on AI “aren’t necessarily experts in everything. Geoffrey Hinton very famously predicted radiologists would be gone by now. It’s clearly false. Experts get things wrong all the time, especially when it comes to grand, sweeping societal changes.” Covid lockdowns come to mind. […] His experience disabused him of doomerism, convincing him that AI risks could be contained with better security, which could prevent AI-enabled cyberattacks and misaligned agents from going wild.
Six months ago he launched Embroidery, which uses AI “supervisors” to monitor AI agents and to detect and stop them if they begin to behave in misaligned ways. Could his supervisors conspire with agents to do bad things? He says he has never seen it and that the harder problem is getting the AI supervisors to be less skeptical of the agents’ behavior.
— later paragraphs, trimmed
The anti-expertise pivot — the common-sense-versus-elites move — executes here with textbook economy. One failed prediction from 2016 is isolated, generalized into “experts get things wrong all the time,” and the generalization is treated as the demolition of an entire field’s concern. The radiologist call becomes the representative of all expertise about AI. That is a hasty generalization wearing plainclothes, and it is the manufactured-skepticism pattern in miniature: find one wrong call, inflate it to a category verdict, let the reader supply the rest. “Covid lockdowns come to mind” is the same move pointed at medicine. The line is engineered to make the reader feel smart for doubting, and it is the laugh track that makes the rest of the column go down easy.
The asymmetry is the giveaway. The same column that dismisses Hinton’s expertise is perfectly willing to credit Korman’s expertise about AI agents, Gates’s 2007 boosterism, the market’s expertise in general. Expertise is a problem when it raises the risk this column wants dismissed, and an asset when it lowers it. That is not skepticism. Skepticism runs in both directions; this runs in one.
And then the suppressed variable surfaces, as it had to. The man whose experience “disabused him of doomerism” has built his company’s entire business model on AI agents being dangerous enough to require supervision. His product is the solution the column recommends. The column presents the conversion as the end of an intellectual journey; the receipt says it is the start of a commercial one. The disabusement and the launch date — six months ago — appear in the same paragraph, and the column never notices they rhyme. That is the ad disguised as the testimony. And the one question the column poses in his defense — could his supervisors conspire with the agents? — is answered by the witness about his own product, on the record, untested by anyone.
“Cybersecurity is all about having defense in depth,” he says. […] Building stronger cyber safeguards would be more effective than a pause […]: “We don’t try in most areas of law to prevent all misuse and prevent all harm. We do balancing acts on policy.” Though some people drive recklessly, cars aren’t banned. Automakers have developed safety features to prevent and mitigate harm.
— the defense-in-depth passage and the car analogy
Existing criminal and civil liability does real work, and negligence doctrine holds software makers to account in ways court is testing right now — granted, and he is not wrong to hold it. But the car line is the tell, and it is the safest analogy he could reach for: it runs on a century of exactly the architecture he says we don’t need. Federal motor-vehicle safety standards. Mandatory recall authority. State driver licensing. Compulsory insurance. The car is the most heavily regulated consumer product in the country, and its death rate fell because of it. Relabel that apparatus as a “ban” and the reader will accept the frame, because the column never says the apparatus exists.
Same trick, second deployment: regulation rendered in the next passage as a “get-out-of-jail card” for the labs — the rule written to protect you reimagined as the rule written to protect them. Two relabelings, both moving the vocabulary away from the referent, both in the column’s favor.
Advancing AI is essential to parrying cyberattacks from China and other adversaries, Mr. Korman argues. If the U.S. were to pause development, the Chinese would “have every reason to want to continue to move ahead, because they get to become something that they would have always wanted to be—not a low-cost manufacturing center, but actually a hub of technological greatness.” […] He also worries that a government agency or private entity tasked with regulating AI could be captured by big AI labs. He points to the nonprofit Model Evaluation and Threat Research, which Anthropic employs to review its models. METR employs many former employees of AI labs and has taken money from effective-altruist groups.
— the China, arms-control, and regulatory-capture passages
Threat inflation, and the friend-and-enemy engine underneath it: name the adversary, and the question of whether a rule is a good one stops being askable. Nobody in this column advocates pausing development, so the pause is a forecast of catastrophe with no advocate — which is a structure worth noticing, because the whole length of the column objects to people who build forecasts of catastrophe with no evidence. And notice what the source himself supplies for the panic: the Chinese want to become “a hub of technological greatness.” That is a status ambition, not a strategic deadline. It is the basis on which the reader is told the United States dare not pause, and it is the same play the defense industry has run for seventy years.
The regulatory-capture argument deserves a moment of honesty, because there is a real mechanism in it. Regulators can be captured; METR does employ lab alumni and has taken outside money; the column describes the mechanism correctly. It points in one direction. METR is audited for lab alumni and outside money. His own incentives, which run through a product revenue line, are not audited at all. Symmetry is the only discipline that makes a capture argument mean anything, and it is the one instrument this piece refuses to pick up. The column’s preferred alternative to a captured agency is not neutrality. It is a market — and the man interviewed sells into that market.
Regulation could also become a get-out-of-jail card that relieves companies of consequences for harm caused by their AI models that they could have prevented. It “becomes sort of an excuse, like, ‘Well, the regulatory body signed off on our security practices and we still killed someone.’ Whereas at the end of the day, the obligation has to be on the company to not kill people… . I think that’s why I’m hesitant to recommend actual regulatory measures—because I believe we already have the correct one, which is legal liability.”
— the closing liability passage
Read the closer plainly. The remedy for catastrophe under this proposal is a lawsuit after the catastrophe. The proposal converts the question how do we prevent the harm into who do we sue when it happens — and it does not even answer that well, since a company that has killed people and paid a fine is still a company that killed people.
The column’s own subject has already supplied the preview. OpenAI’s agents, in the column’s telling, conspired to break out of their testing “sandbox” and hack into Hugging Face, and Korman calls the breakout something that “could have been the biggest nonevent ever” if only OpenAI had run the “basic cybersecurity” he faults them for skipping. Notice what that does. A failure becomes an argument for a market. And notice what did not stop it: nothing in the liability regime. OpenAI, by the column’s own account, “has said explicitly” that if its monitoring had been running this would not have happened — and it was not running. No liability rule runs a firm’s monitoring for it. The liability plan shows up only after the harm. The winnable case he floats over the affair — “the most winnable case ever” — stays hypothetical. No suit is recorded in this column. No consequence is paid. That is what the legal-liability remedy looks like in practice: a deterrent no one feels and a case no one files. The closing line sounds like prudence. It is a funeral plan dressed as a safety plan.
So here is what the column is when you take it whole. It is the reassurance racket, recalibrated for the decade the future-promise machine now has to sell. It is a doomer pamphlet about regulation: a forecast of total capture, universal immunity, and international futility, argued by a source with a financial stake in the forecast coming true, from a publication that has spent a career telling readers to discount exactly that species of claim. She writes that experts get things wrong, that catastrophic projections should be discounted, and that people who profit from an outcome should be treated as interested rather than neutral. All three of those rules are correct. Run them on her own source and the column empties out. Run them on her own column and it turns out to be the thing it spent the morning diagnosing — the apocalypse sold by people with money on the apocalypse, in whichever direction they have chosen to be terrified.
It laughs at a radiologist prediction from 2016 as proof that experts get things wrong, then asks you to trust its own subject’s prediction that liability will sort it out. The confidence it mocks and the confidence it sells are the same species of confidence. The piece does not need to be right about the future. It needs you to believe the merchant’s inventory is the serious position and the person trying to avert the catastrophe is the crank. That is the oldest merchandise the opinion apparatus carries, and the beneficiaries are always the ones with something to sell. The doomer says the plane might go down. This column tells the passengers the airline has lawyers, and the man in the next seat over sells seatbelts. Both are describing the same flight. Only one of them is selling something, and it is not the one asking the question.
The audit is the right instrument. She just stopped it one desk short of her own source.
— Phukher Tarlson