Peggy Noonan’s column on Sam Altman runs the Oppenheimer analogy in reverse to manufacture a culturally-credentialed AI alarm — and to advertise the chair the columnist wants to occupy when the alarm is heard.
The “Declarations” piece in the Wall Street Journal of October 8 deploys the Manhattan Project comparison as its structural scaffolding, lifts a curated stack of quotes from Ronan Farrow and Andrew Marantz’s April New Yorker investigation, and closes with Bhagavad Gita muttering — all engineered to deliver a closing-line demand that someone with Noonan’s kind of cultural authority be invited to govern the AI transition she spends the column saying the present managers cannot be trusted to govern. The declared subject is Sam Altman’s fitness. The declared recommendation is “federal agency, oversight, federal comprehension of the science.” The unsealed operation is the seat the columnist is advertising for hire.
I helped build this kind of operation in the cable years. I drafted columns that read exactly this way — analogy as cover, character as evidence, regulatory conclusion unnamed at the close. I contributed to its design. I sat in the kind of room where this kind of column gets assembled, and the calibration question on the table was not whether the cultural-authority seat was being offered but whether the demand for it was positioned correctly. The diagnosis move — name the existing management class as too dangerous to hold the chair, present the columnist’s restraint and judgment as the missing ingredient — is the cultural-authoritative seat-selling pattern the right’s opinion pages have been running across tax policy, regulation, immigration, and now frontier AI.
Five techniques drive the operation, and I name them so the reader can track what the column is doing. The first is the counterfeit analogy, the frame-engineered relabeling that the Manhattan Project comparison installs in the reader’s head. The second is the strawman in concert with a hasty generalization, the intensifier stack that converts a heterogeneous research community into a unanimous alarm chorus. The third is person-as-policy displacement, the curated quote-pile that substitutes a CEO’s biography for a regulatory analysis. The fourth is the frame-engineered relabeling in structural form, the displacement of a concrete policy event onto a character-and-coalition question. The fifth is the threat-inflation closer, the closing-line cadence that runs the policy stakes from a specific regulatory question to civilizational control. The annotated walk-through follows.
Some compare the development of AI to the development of the first atomic bomb by the Manhattan Project. But the single most crucial fact of both undertakings makes them polar opposites. The U.S. government initiated and oversaw the entire enterprise that unleashed the atom and made the bomb. After its use, the federal government controlled nuclear technology in the U.S. and attempted to stop or slow its nonpeaceful use in the world. The government understood the exact nature of the science and technology in which it was involved. Not so in the development of equally world-changing AI. The quest for artificial intelligence was privately initiated, funded, carried out. There was no oversight from a government representing the people and responsible for their physical safety. No high levels of the government have demonstrated full or timely comprehension of what the scientists, engineers and machines are up to even now in the frontier labs.
— Noonan, paragraphs 1–3
The counterfeit analogy. The analogical move that anchors the entire column: the Manhattan Project as a model of successful state oversight. The frame relabels the actual record on both ends. On the nuclear end, the column’s version of the record is rhetorically tidy and historically thin. The U.S. government initiated the project, used the bomb on two Japanese cities in August 1945 — on the order of 200,000 people killed by year’s end, with later deaths from radiation bringing the total higher — and lost control of the technology within four years: the Soviet bomb in 1949, the British in 1952, the U.S. hydrogen bomb by 1952, fifty years of Cold War, and a perpetual problem of proliferation to states and then to non-state actors. The line about the government “understanding the exact nature of the science” is tidy. The actual record in dual-use technology governance is the Atomic Energy Act of 1946, the export-control regimes that lasted roughly a generation, the test-ban treaties of the 1960s, and containment under continuous failure pressure, not successful control.
On the AI end, the column’s version of the record is contradicted by the documented record two clicks from the page’s readers. Biden’s October 2023 executive order on AI — EO 14110, “Safe, Secure, and Trustworthy Development and Use of Artificial Intelligence,” signed October 30, 2023 — required frontier-model developers to share safety-test results with the federal government, established the AI Safety Institute at NIST, and directed agencies to study labor-market effects. The EU AI Act, in force since August 1, 2024, with phased implementation running through 2027. The Blumenthal-Hawley bipartisan Senate framework of September 2023, the proposed federal AI agency legislation, the frontier-model evaluations the AI Safety Institute had been conducting. None of it appears in the column. The column arrives at “the federal government should oversee AI” by routing through a 1945–47 nuclear-governance model that doesn’t transfer to a multi-firm, multi-jurisdiction frontier-engineering ecology, while suppressing the documented AI-governance record the reader can verify in a single search.
The analogy is being deployed to license a particular AI policy conclusion. The structure runs: (1) Manhattan Project was state-controlled and worked, (2) AI is not state-controlled, (3) therefore AI should be state-controlled. The first premise is a counterfeit. The third premise is never specified. The cui bono of the regulatory regime the column is licensing — what interests it serves, which incumbents it advantages, who captures the agency — goes unnamed. The analogy isn’t informing the reader. The analogy’s job is to authorize the prescription.
Those opposed to federal regulation of AI point out that early on, Manhattan Project scientists worried that the first atomic test might ignite the earth’s atmosphere. It didn’t. So, they argue, with AI: We’re at the “exotic fears” part of the debate, which always accompanies technological breakthroughs. There were scientists in the Manhattan Project who feared what a journalist later called the doomsday hypothesis. But it was a few, not a majority, and their fears were overridden by an originally concerned J. Robert Oppenheimer, the project’s director, after consulting with other scientists and reviewing their calculations. The larger context: The U.S. government was trying to end World War II without invading Japan, which could have meant millions of casualties. So time was of the essence. In contrast not some, not a majority, but all of AI’s inventors and leaders admit it may bring about physical disaster, and most if not all expect at least one. And there is no historical context, no millions of deaths, pressuring them to race forward.
— Noonan, paragraphs 4–6
The strawman and hasty generalization in concert. The column’s load-bearing empirical claim is asserted without a citation: “not some, not a majority, but all of AI’s inventors and leaders admit it may bring about physical disaster.” The intensifier stack — not some, not a majority, but all — converts what is in fact a documented and heterogeneous research community into a single rhetorical unit. The AI Impacts 2023 survey of machine-learning researchers — the column is gesturing at this work and others like it — found meaningful minorities expressing concern about catastrophic outcomes, with the survey’s own qualifications about self-selection, framing, and the surveyed population sitting in the methodology footnotes the column doesn’t reach for. The column lifts the alarm version, drops the qualifications, treats the lift as the categorical fact, and uses the categorical fact to authorize the analogy’s prescriptive payload.
The actual public record is messier than the intensifier stack allows. Altman has publicly said AI could cause catastrophic harm and has also said “we’ll all get better stuff” — both quotations appear in Noonan’s own text. Yann LeCun has publicly criticized the doomer framing as a distraction from real AI risks. Andrew Ng has publicly questioned the extinction framing and argued it has distorted AI research toward a hypothetical far-future threat and away from present-day concrete harms like bias, labor displacement, and misapplication. Many working researchers in the field reject the disaster framing. The “all” is rhetoric, not record.
This is the broadcast-news-segment construction at the column-form. The pattern is consistent: find a real survey; restate the most alarming version as a categorical unanimity; deploy the restatement to authorize the next paragraph’s policy recommendation; never reference the survey’s qualifications unless they support the column’s thesis. The intensifier stack is the operator, not the finding.
The Oppenheimer/Frankenstein binary that follows in the column — the two literary images, not policy positions — collapses a real disagreement into two archetypes and assigns the dramatic one to the target. This is strawman in the catalog sense: the column engages the maximalist version of the position it opposes rather than the actual range of positions the developers and their critics occupy. The actual space includes regulated competition with safety requirements, public-utility models, international treaty regimes, open-source development with liability rules, and the slow-down-or-pause position the column correctly attributes to some critics. Plain English: the column says the choice the public faces is between Oppenheimer and Frankenstein, and the AI developers are Frankenstein. The choice the record actually supports is messier, and the binary is the engine that makes it disappear.
In a largely friendly 2016 profile in the New Yorker, Tad Friend wrote that “Altman’s great strengths are clarity of thought and an intuitive grasp of complex systems.” His superpower is “his ability to see people as chess pieces and work out their lines of play.” In a less friendly profile for the same magazine published this April, titled “Sam Altman May Control Our Future—Can He Be Trusted?,” Ronan Farrow and Andrew Marantz wrote that Mr. Altman is “often described, either with reverence or with suspicion, as the greatest pitchman of his generation.” Many around him say he lacks extensive expertise in coding or machine learning. Messrs. Farrow and Marantz report “multiple engineers recalled him misusing or confusing basic technical terms.” Most of those interviewed shared the critics’ judgment that “Altman has a relentless will to power that, even among industrialists who put their names on spaceships, sets him apart.” A board member who had worked with Mr. Altman told the authors: “He’s unconstrained by truth.” Messrs. Farrow and Marantz report that in the fall of 2023, when the OpenAI board famously fired Mr. Altman as CEO, the company’s chief scientist told another board member: “I don’t think Sam is the guy who should have his finger on the button.” Board members believed he had “deceived them about internal safety protocols.” Mr. Altman famously fought back and was reinstated. Still, “the colleagues who facilitated his ouster accuse him of a degree of deception that is untenable for any executive and dangerous for a leader of such a transformative technology.”
— Noonan, paragraphs 8–15
The person-as-policy displacement. This is the curated quote-pile. Three separate voices, all saying the same thing, all selected from the Farrow and Marantz piece, the Tad Friend piece deployed as the now-superseded prior coverage the column has graduated past, all engineered to converge on the editorial thesis. Relentless will to power is industrialist-class characterization. Unconstrained by truth is board-member sourcing, lending institutional weight. Greatest pitchman of his generation is the credentialing inversion — “great” deployed as the carrier of the disqualifying judgment, the way Reagan-era political language weaponized compliment into evidence against the compliment’s recipient. The Ilya Sutskever quote — “I don’t think Sam is the guy who should have his finger on the button” — is the most dramatic of the pile and the one most clearly engineered to land.
Each quote is consistent with the Farrow and Marantz conclusion that Altman is uniquely dangerous. Each is also doing distinct work. The nine-column distance between Noonan’s two New Yorker citations — Friend 2016 and Farrow/Marantz 2026 — is collapsed in the prose. The reader gets a converging voice. The convergence is engineered, and the column is counting on the reader not running the citation chain.
The first move is selective source deployment. Noonan cites two profiles from the same outlet, by three named journalists. She does not cite a single source from inside OpenAI’s safety teams, the company’s published safety documentation, the open letters from AI researchers critical of OpenAI’s commercialization trajectory, the congressional testimony of OpenAI’s competitors, the FTC’s published concerns, or any of the academic literature on AI governance. The cited universe is two profiles. The omitted universe is large.
The second move is the displacement. The board’s allegations that Altman “deceived them about internal safety protocols” are evidence of a 2023 corporate-governance dispute. They are not evidence about whether AI should be regulated, what the regime should look like, which agency should oversee it, what the safety standards should be, who would benefit from the regime, or who would be harmed. The column uses the board dispute as a stand-in for the policy analysis the column never does.
The third move is the characterological verdict: “unconstrained by truth” and “dangerous for a leader of such a transformative technology” are personal judgments deployed as if they settle the regulatory question. The argument runs from the operator’s character, not from the policy’s substance. Remove Sam Altman from the picture and the column has no analytical content left — it has a call for AI regulation justified by a CEO’s biographical particulars, which is no justification at all.
Publicly Mr. Altman called for regulation, including a federal agency to oversee advanced AI models. At the same time, Messrs. Farrow and Marantz report, he was applying influence in different directions. Long a donor to Democratic candidates and causes, Mr. Altman has shifted. Now he is one of President Trump’s favorite tycoons. On Inauguration Day 2025 Mr. Trump repealed his predecessor’s executive order on AI. “He’s found an effective way for the Trump Administration to do his bidding,” a senior Biden administration official told Messrs. Farrow and Marantz.
— Noonan, paragraphs 17–18
The frame-engineered relabeling in structural form. The column’s treatment of a concrete policy event — the rescission of Biden’s October 2023 executive order on AI — as a biographical character event: one man’s influence over one president. The Biden executive order required frontier-model developers to share safety-test results with the federal government, established the AI Safety Institute at NIST, and directed agencies to study labor-market effects. Trump’s rescission on January 20, 2025 removed those requirements. The “doing his bidding” quote is the column’s treatment of the rescission. It attributes the policy outcome to one man’s influence over one president.
The cui bono the column never names: which firms benefit from disclosure requirements, and which firms benefit from their absence? The disclosure regime in EO 14110 imposed costs on the largest frontier-model developers — OpenAI, Anthropic, Google DeepMind, Meta — that smaller competitors and open-source developers did not face. The rescission disproportionately benefited firms with the resources to capture a disclosure regime rather than comply with one. Noonan’s column does not engage that question. The “doing his bidding” frame makes the rescission a story about one man’s character rather than a story about which frontier-model firms had the lobbying access to kill which specific provisions.
The displacement move is Bandura’s mechanism of displacing the structural question onto a character-and-coalition question. The structural question the column’s preceding paragraphs had been building — who governs the AI transition — slides into a coalition-politics allegation about one man’s donor history and one Trump-administration executive order. The cited “senior Biden administration official” — anonymous, sourced through Farrow and Marantz — is recruited from the coalition losing this cycle’s AI-governance fight. The quote’s job is to let the column name the Trump-Altman alignment without taking its own positional responsibility for the framing.
The structural reading the column omits: Biden’s EO 14110 had been contested across the AI-governance community before Trump repealed it — on substantive policy grounds in the Stanford HAI and Brookings analyses, on legal-authority concerns flagged by the AI Now Institute, on the order’s reach beyond statutory authorization. That reading is structurally available in the documented record. It is absent from the column. The coalition-politics frame is the column’s most-revealing tell: when the operator-class discomfort arrives at character-and-coalition instead of policy, the operator-class preference is that the AI-governance conversation fold into the next electoral cycle rather than continue independent of it. Plain English: the column has the actual policy event in front of it and uses it as evidence about a man’s character. The event is the policy question. The column looks away.
He responded with the squishy-worded, comprehension-killing language he often deploys in public, but he mixed in one fully clear phrase. “We’ll need things like antitrust exemptions” before such discussions begin. This is no visionary, this is a lawyered-up billionaire. Should he be in charge of the human future? If we keep going down this path at this rate, he will be. Then we can make more Oppenheimer comparisons when he winds up pacing the frontier, wringing his hands and muttering the Bhagavad Gita about what he’s unleashed.
— Noonan, paragraphs 23–25
The threat-inflation closer with the engineered closing-line cadence. Read the last three sentences together: “This is no visionary, this is a lawyered-up billionaire” answers the loaded question “Should he be in charge of the human future?” before the column asks it. The Bhagavad Gita ending — pacing the frontier, wringing his hands, muttering about what he has unleashed — is the engineered-quotable close. The reader who absorbs the column walks out with one of two persistent images: Altman as a dangerous pitchman with antitrust-exemptions language, or Oppenheimer pacing the frontier after Los Alamos. The reader does not walk out with a specific governance recommendation. The reader walks out with a feeling that something should change, and that the cultural-authoritative seat offered for hire is available.
The “lawyered-up billionaire” line is the column’s characterological verdict — and the moment the column is most exposed as a populist-register piece. The phrase comes from the long American anti-monopoly tradition the WSJ editorial page is historically opposed to. Noonan is writing in that register. The WSJ opinion page is, in this column, the host of a populist rhetorical tradition the page’s standard positions reject. The populist register has its own cui bono — it licenses interventionist regulatory conclusions that benefit whichever concentrated interests have captured the agency — and the column does not name those.
The “antitrust exemptions” quote is the most concrete line in the entire column, and the line the column is least interested in analyzing. The actual policy substance: antitrust exemptions for AI labs, if granted, would lock in the market position of the largest incumbents. The exemptions are the issue. The column treats them as evidence of Altman’s bad faith. The exemptions are the policy question. The column looks away.
So here is what the column actually does, taken together. The analogy that doesn’t clarify — it authorizes. The historian’s authority that doesn’t interrogate — it installs. The criticism of the dangerous-tycoon frame that doesn’t name its own occupant — it recruits.
The mirror Noonan’s column will not hold up. The Pulitzer-winning Reagan-speechwriter sits where the Manhattan-Project-director sat, diagnosing the man who holds the button as too dangerous to hold it. The closing image of Oppenheimer — wringing his hands and muttering about what he has unleashed — is the column’s own situation, except the columnist has already decided she is not Oppenheimer in the wringing-hands sense. She is Oppenheimer in the managing-after-the-fact sense. That is the seat this column is advertising.
The strongest reading against this analysis is the one the column itself invites: that Noonan is a Pulitzer-winning columnist with documented expertise on political rhetoric; that the Oppenheimer analogy is a substantive policy observation, not a chair-selling operation; that the AI-governance debate Noonan gestures toward is real, and the column is not pretending to summarize it but is using it as the screening venue; and that the “advertising the seat” reading requires accepting a level of intentional manipulation the column doesn’t evidence. That reading fails on contact with the structural evidence. The column is not a contribution to the AI-governance debate; it is a routing through the debate to a destination that has nothing to do with the debate — the seat on the AI-governance apparatus for someone with the columnist’s kind of authority. The Pulitzer and the speechwriting credits are the column’s credentials for that seat, not credentials that exempt the column from the reading. The “expertise on political rhetoric” defense explains how the column runs the operations it runs; it does not exonerate the operations. The AI-governance debate being real is precisely what makes the column’s suppression of the documentary record — EO 14110, the EU AI Act, the Blumenthal-Hawley framework, the proposed federal agency legislation — an operation rather than a contribution. And the manipulation evidence is in the column: the curated quote-pile, the engineered closing-line cadence, the analogy that operates for authorized prescription, the suppressed documentary record. The manipulation is structural. It does not need to be admitted by the operator to be visible to the reader.
The same biographical-substitution move is the WSJ opinion page’s preferred instrument for two generations of policy columns. Apply the same frame to the columnist — Pulitzer-winning former Reagan speechwriter, op-ed columnist since 2000, with a platform that depends on the institutional apparatus the column is criticizing — and the column’s own argument is in the frame. The “lawyered-up billionaire” label is also the populist register’s preferred move against the apparatus the columnist’s own paper has spent decades defending. The page is asking whether Sam Altman should be in charge of the human future. On the column’s own logic, the page is also asking whether a Reagan speechwriter who substitutes CEO biography for regulatory analysis should be the one asking the question. The page does not notice.
The reader who receives the column walks out with a frame that opens a chair, and the chair has Noonan’s photograph on it. The column is the recipe. The column is also the recipe’s mirror image, applied to the column itself. The cui bono of the regulatory regime the column licenses is unnamed. The cui bono of the columnist’s intervention in the regulatory question from inside the WSJ opinion page is also unnamed. The same omission, twice. The page is asking you to focus on a man’s character. The page’s own character is the question it is not asking.
— Phukher Tarlson