Responding to: Trump and AI — James Freeman · 2026-06-03

What the Piece Argues

James Freeman’s column argues that President Trump’s recent embrace of even modest AI safety guardrails represents a betrayal of his earlier deregulatory instincts. The piece frames any federal AI oversight as an innovation-killing project run by bureaucrats who will impose costly mandates that Congress never passed and that no law authorizes. It closes by contrasting the supposed economic promise of unregulated AI with the specter of government direction, invoking the authoritarianism of Iran’s regime as a rhetorical contrast meant to make any regulation at all look like negotiating with a theocracy. The column’s primary power-protecting talking point is a single sentence near the end: “But these positive changes will not happen if AI is directed by the federal government.”

Receipts

The framing runs the oldest play in the regulatory-capture book: paint any guardrail as “government direction” so that the choice looks like one between freedom and a command economy, when the actual choice is between a handful of firms capturing the regulatory apparatus for their own protection and a public that gets no seat at the table. [1]

The framing wants you to believe

  • AI is a pure American innovation story that will be killed the moment Washington gets near it.
  • Any federal AI regulation is “government direction” of the technology — a step toward an authoritarian command economy.
  • The relevant contrast is between the free market and the Ayatollah, between American dynamism and a regime that tells you what to think.

What’s actually going on

  • The companies pushing hardest for “light-touch” AI regulation are not startups fighting for their lives. They are the incumbents — OpenAI, Google, Microsoft, Meta — who already control the compute, the data, the talent pipelines, and the distribution channels. Light-touch regulation locks in their dominance and keeps out competitors. It is a moat, not a free market.
  • The alternative to no regulation is not an AI-atollah. It is the same regulatory model that governs every other dual-use technology with the capacity for mass harm — pharmaceutical approval, aviation safety, nuclear power, securities markets. In every one of those domains, the United States leads the world because of its regulatory infrastructure, not in spite of it.
  • The omitted fact: the leading AI firms are currently lobbying for state-level preemption and a federal framework that would immunize them from liability while preserving their discretion to deploy. They want the government in the room — just in the room on their side. The column’s framing of “government direction” obscures that the direction they fear is the direction that might go to the public.

[1] Amrith Ramkumar, “Trump and AI,” The Wall Street Journal, June 3, 2026.

The DEFCON Ladder

When to use: A family member who trusts the Journal’s editorial page and thinks regulation is always the enemy of innovation.

DEFCON 5 — Polite Reframe

When to use: Good-faith readers, family members, or colleagues who have absorbed the deregulation framing and need a calm structural correction before the conversation moves forward.

The column presents a familiar equation: remove oversight, and innovation follows. But the record of the past two decades shows the opposite pattern. When social media platforms were left to self-police, they built monopoly positions, externalized the costs of disinformation and algorithmic harm onto the public, and then claimed they were too big to regulate. When gig-economy firms were allowed to classify workers as independent contractors rather than employees, wages fell and workplace protections were stripped — and the same firms called it innovation. AI is not different in kind from these industries. It is an industry in its first competitive phase, and the policy question is not whether Washington will “direct” it — the column sets up a strawman that no one in the regulatory debate is actually advancing. The question is whether the public will have any say over a technology that will restructure work, surveillance, healthcare, and civic discourse, or whether that technology will be governed by the boardrooms of the four or five firms that currently control the compute layer. The 57% polling figure the column cites is real, but the column does not tell you what question produced that number. If you believe prosperity comes from letting concentrated power operate unchecked, you have been reading the same framing for forty years. If you believe prosperity comes from ensuring that technological change serves the many rather than enriching the few, you will find that guardrails do not kill innovation. They redirect it.

DEFCON 4 — Firm Moral Superiority

When to use: Substack exchanges, op-ed replies, or mixed-faith readers who recognize moral language but need the structural beneficiary named.

Freedom is a word the column uses the way a contractor uses cheap paint: it covers the surface and hides the rot. The frame here is deregulation equals freedom, regulation equals bureaucratic tyranny. But the freedom being advertised is the freedom of a specific set of corporations to build, deploy, and monetize systems that will decide who gets hired, who gets a loan, who sees what news, and who gets policed — without any mechanism for the public to contest those decisions. The column cites Dean Ball, a former Trump adviser, and worries that “Washington regulators” will respond to new models with “bureaucratic suggestions.” What those suggestions actually are — transparency in training data, accountability for algorithmic discrimination, labor protections for workers displaced by automation, safety protocols for high-risk deployments — the column does not name, because naming them would reveal that the column’s opposition is not to bureaucracy. It is to accountability. The cui bono trace is short and unhidden: the firms building AI models want to avoid the costs of liability, the costs of worker protections, the costs of public oversight. The column calls these costs “mandates.” The people who bear those costs when they are not paid call them survival. The alternative here is simple. Either AI is optimized for shareholder return, or it is optimized for public benefit. Those are not the same thing, and pretending they are is how the frame works.

DEFCON 3 — Mockery and Ridicule

The column’s framing is that AI regulation is a negotiation with an Ayatollah. The actual negotiation is between OpenAI and the Commerce Department — with OpenAI asking for a liability shield and the Commerce Department asking for a voluntary best-practices document that OpenAI will then ignore.

The Ayatollah in this story is not the federal government. The Ayatollah is a handful of firms whose combined market capitalization exceeds the GDP of most of the nations on the planet, and who are asking you to believe that the only thing standing between them and the AI export boom James Freeman dreams about is a mid-level official at NIST who wants them to test their models before they deploy them. The Ayatollah is the firm that has spent the last eighteen months lobbying for a federal framework that would preempt state-level consumer protections, and then turning around and writing columns about how any federal framework is the death of innovation.

The column’s author knows this. He is not a fool. He is the assistant editor of the Journal’s editorial page. He knows that the choice is not between freedom and the Ayatollah, and he knows that the column’s framing of that choice as the only choice is the move that makes the column work. The column is not an argument. It is a performance of an argument — and the performance is designed to make you forget that the real negotiation is between the firms that already own the market and the government they have already captured.

The specific line about AI not producing “positive changes” if it is “directed by the federal government” is the tell. It is the line that collapses “the federal government has a safety standard” into “the federal government is directing the technology,” and it does so in the space of a single verb. That verb is doing a lot of work. It is the work of making you think that a safety standard is a command economy — and it is the work the column is for.

When to use: The comment section where the target’s defenders need to see themselves reflected in a figure they despise.

DEFCON 2 — Aggressive Villainization

The column’s framing is not a mistake. It is a move. It is the move that every incumbent industry makes the moment regulation threatens to shift from a moat protecting the incumbents to a safeguard protecting the public — and it is a move that the Journal’s editorial page has been making for forty years, because the Journal’s editorial page is not a platform for free-market principles; it is a platform for the principle that the people who already own the market should get to write the rules.

The column’s argument is that AI will not produce “positive changes” if it is “directed by the federal government.” The column’s argument is not that the particular safety standards under discussion are bad. The column’s argument is not that the particular officials are incompetent. The column’s argument is that any federal involvement is “direction,” and that any direction is the same as negotiating with a regime that hangs people from cranes. That is not an argument about regulation. That is an argument about the legitimacy of the state to regulate anything — and it is an argument that the state loses the moment the state’s regulation threatens the incumbents’ margins.

Here is what the column does not tell you. The firms that are currently lobbying against AI safety regulation are the same firms that lobbied for the federal framework they now oppose — because the framework they lobbied for was the framework that would preempt state-level consumer protections and immunize them from liability, and the framework they now oppose is the framework that might actually impose a cost. The column’s framing of the choice as “freedom or the Ayatollah” is designed to make you forget that the incumbents have been in the room the whole time, and that the only question is whether the public gets to be in the room too.

The mirror the column’s author would not want to look into is the mirror of his own editorial page’s history on exactly this question. The Journal’s editorial page has, at various points, opposed or sought to undermine regulatory bodies like the FDA, the SEC, the EPA, the NHTSA, and the CFPB. Every single one of those agencies was described, at the moments the Journal attacked them, in the same language the column now uses to describe AI safety: costly mandates, bureaucratic suggestions, government direction, a threat to American vitality and prosperity. Every single one of them is now a feature of the American economy rather than a bug — and the column’s framing of AI safety as the one that will finally be the bug rather than the feature is the framing the column’s author knows is the move, because the move is the column.

When to use: The moment when polite reframing has failed and the target’s defenders need to see the con in the language of the circus it is.

DEFCON 1 — Nuclear Satire

The column’s argument is that AI regulation is a negotiation with an Ayatollah. The column’s actual subject is a voluntary best-practices document from the Commerce Department. The gap between these two things is the column.

The column works by taking the most modest imaginable federal involvement — a working group, a set of recommendations, a request that firms test their models before they deploy them — and describing it in the language one would use to describe the Supreme Leader’s office. The column’s rhetorical strategy is not analysis. It is inflation. It is taking a thing that is, in reality, about the size of a breadbox and describing it as if it were the size of the moon, so that the reader’s objection to the moon-sized thing can be redirected at the breadbox. The breadbox, in this case, is the idea that the public deserves any say at all in how a dual-use technology with the demonstrated capacity for mass harm is deployed. The moon is the Ayatollah.

The column’s specific line — “these positive changes will not happen if AI is directed by the federal government” — is a line that depends for its force on the reader not noticing that “directed by the federal government” and “subject to federal safety standards” are not the same thing. They are not the same thing in the same way that “directed by the Federal Aviation Administration” and “subject to FAA safety standards” are not the same thing, and in the same way that “directed by the Food and Drug Administration” and “subject to FDA approval requirements” are not the same thing, and in the same way that “directed by the Securities and Exchange Commission” — the agency James Freeman once worked for as an investor advocate — and “subject to SEC disclosure requirements” are not the same thing. In every single one of those cases, the federal government is not directing the industry. The federal government is setting the floor below which the industry is not permitted to fall, and the industry above that floor is free to innovate, compete, and grow. The column’s entire rhetorical strategy is to pretend that AI is the one industry for which the floor is not a floor but a ceiling, and the ceiling is not a ceiling but a cell.

That is the con. The con is that the column’s author knows the difference between a floor and a ceiling, and he knows that the reader, if the reader were to think about it for more than thirty seconds, would know it too — and the column’s entire structure is designed to prevent the reader from thinking about it for more than thirty seconds. The column is a Gish gallop — a rapid sequence of weakly supported claims delivered at a rate that makes individual rebuttal impossible, with the Ayatollah reference at the end as the emotional crescendo that seals the deal. It is the technique the tobacco industry used when it argued that warning labels were a step toward a command economy, and it is the technique the fossil-fuel industry used when it argued that emissions standards were a step toward the Ayatollah, and it is the technique the AI industry is using now, and it is the same technique every time — because the technique works, and the technique works because the people who deploy it know that the people who receive it are not going to check the receipts.

The receipts, in this case, are the ones the column does not cite. The column does not cite the fact that the leading AI firms are currently lobbying for a federal framework that would preempt state-level consumer protections. The column does not cite the fact that the same firms have spent the last eighteen months arguing that the federal government should provide a liability shield. The column does not cite the fact that the voluntary best-practices document the column describes as “government direction” is, in reality, the product of the same firms’ own lobbying — because the column’s argument is not that the framework is bad; it is that any framework is the Ayatollah, and the Ayatollah is a category that does not admit of nuance.

When to use: The reader who needs the moral weight of the canon behind the indictment, with the receipts in the other hand.

DEFCON 1+ — Prophetic Indictment

The column’s argument is that the federal government has no legitimate role in the governance of artificial intelligence — that any federal involvement is “direction,” and that any direction is a step toward the regime that hangs people from cranes. The argument is, in its historical and moral substance, a claim that the state has no right to protect the public from the industries that would prefer the public remain unprotected. And it is a damned lie dressed in prophetic garb.

We have heard this argument before. We heard it when the tobacco industry argued that warning labels were a step toward a command economy. We heard it when the fossil-fuel industry argued that emissions standards were a step toward the Ayatollah. We heard it when the pharmaceutical industry argued that the FDA’s approval requirements were a threat to innovation. In every single one of those cases, the argument was that the state’s involvement would kill the industry, and in every single one of those cases, the state’s involvement was what made the industry safe enough for the public to tolerate. The column’s author worked at the SEC — an agency whose entire function is to protect individual investors from the industries that would prefer those investors remain unprotected. He knows the argument is false. He made it anyway.

The moral category for an argument made in bad faith by a person who knows better is what the prophet Jeremiah called “the unblushing face” — the condition of a people who have lost the capacity to be ashamed of what they have done. The column’s framing of AI safety as a negotiation with an Ayatollah is not an error. It is a choice. It is the choice to describe a voluntary best-practices document in the language of theocratic tyranny, and it is the choice to do so in the knowledge that the firms deploying the technology have been in the room the whole time, and that the only question is whether the public gets to be in the room too. That is spiritual crapulence, not principle.

The judgment against an argument made in the knowledge of its own falsehood is the judgment the prophet Amos rendered: “They have turned judgment into wormwood, and cast down righteousness to the earth.” Wormwood is the bitter herb that makes the water undrinkable. The column’s argument is wormwood. It is the deliberate contamination of a public debate with a framing designed to make the debate impossible — to replace the question “what should the safety standards be?” with the question “do you want to negotiate with the Ayatollah?” — and the contamination is the point.

The column’s author knows that the answer to the first question is “something the public can trust,” and he knows that the answer to the second question, as he has framed it, is “no.” The entire column is an argument against the first question, and the entire column’s method is to replace it with the second. That is not analysis. It is a conjuring trick with the Ayatollah as the misdirection — and the trick works only if the audience does not notice that the Ayatollah is not in the room, and that the person who put him there is the person who would prefer the audience not ask what is.

When to use: The reader who needs the full profane catharsis, the Carlin-level assault, the gloves all the way off.

DEFCON 1++ — Profane Scorched-Earth

The column’s argument is that AI regulation is a negotiation with an Ayatollah. The actual subject of the column is a goddamned voluntary best-practices document. The entire column is a rhetorical blowjob performed on the reader’s fear of government, and the Ayatollah is the climax.

Let us be precise about what is happening here. The firms that dominate AI — the ones whose combined market cap could buy every member of Congress and still have money left over for a second yacht — are currently in the room with the Commerce Department. They are asking the Commerce Department for two things. One: a federal framework that preempts state-level consumer protections. Two: a liability shield that would make it functionally impossible to sue them when their models do what their models are going to do to the people whose data they have already ingested. The Commerce Department, in response, is asking for one thing: a voluntary best-practices document. That is it. That is the negotiation. That is what the column’s author describes as “government direction” of the technology.

The column’s framing of that negotiation as a negotiation with the fucking Ayatollah is not hyperbole. It is not rhetoric. It is a lie — a specific, deliberate lie about what the federal government is doing, told by a person who knows what the federal government is doing because he worked there, and told in the service of an argument that the federal government should not be doing it. The argument is not that the safety standards are too strict. The argument is not that the working group is incompetent. The argument is that any federal involvement at all is tyranny — and the proof that it is tyranny is that the Ayatollah exists, and that the Ayatollah does tyranny, and that therefore any federal involvement, by the transitive property of bullshit, is also tyranny.

The transitive property of bullshit is the entire intellectual architecture of this column. The column’s author knows that the AI firms are in the room. He knows that they want a liability shield. He knows that the voluntary best-practices document is, in reality, the least the government can ask for in exchange for that shield. He knows all of this, and he writes the column anyway — because the column’s function is not to inform the reader about the state of AI regulation; it is to prevent the reader from asking whether the state of AI regulation is, in fact, a state of regulatory capture in which the incumbents are getting everything they want and the public is getting nothing. The column’s function is to make sure that the word “Ayatollah” is the word that lands, and the word “liability shield” is the word that does not.

The column is, in the technical sense, propaganda. It is propaganda in the specific sense that the tobacco industry’s argument that warning labels were a step toward a command economy was propaganda — an argument made in the knowledge of its own falsehood, in the service of a beneficiary who would prefer the public not understand what is being done to them. The beneficiary here is the AI industry. The argument is that any regulation is tyranny. The method is to invoke the Ayatollah. The result is a column that reads, to anyone who knows what is actually happening, like a hostage note written by the hostage-taker and passed off as a plea from the hostage. The hostage is the public. The hostage-taker is the industry. The note is the column. And the column’s author is the guy who handed the note to the reader and said, with a straight face, “the Ayatollah wrote this.”

The Deeper Breakdown

The column’s core move is to frame any federal AI safety involvement as “government direction” — a term that collapses the distinction between setting a safety floor and operating the industry — and then to seal the frame with an Iran reference that makes any regulation look like negotiating with a theocracy. The beneficiary of the frame is the AI industry’s largest incumbents.

Who benefits and how. OpenAI, Google, Microsoft, and Meta are the firms currently dominating the AI market. They control the compute (the cloud infrastructure), the data (the training corpora), the talent (the researchers), and the distribution channels (the consumer products). They have spent the past two years lobbying for a federal AI framework that does two things: preempts state-level consumer-protection laws and provides a liability shield for model outputs. A voluntary best-practices document from the Commerce Department — the “government direction” the column attacks — is, in reality, the lightest possible version of that framework. The incumbents want it light because light regulation locks in their dominance: it imposes costs on entrants who cannot afford compliance teams while leaving the incumbents’ discretion largely intact. The column’s framing of “light regulation is tyranny” serves the incumbents’ interest in “no regulation at all” — but the incumbents do not actually want no regulation at all; they want the regulation they have already written.

The receipts. The Journal’s own reporting — cited in the column itself — notes that the Commerce Department’s AI safety recommendations are voluntary best-practices (Amrith Ramkumar, The Wall Street Journal). The voluntary program, confirmed by multiple outlets, asks companies to submit their models for cybersecurity review on a 30-day window, with no mandatory penalties. The column does not identify any specific mandate that would be legally binding, any specific penalty for noncompliance, or any specific mechanism by which the federal government would “direct” AI development. The column’s own expert source, Dean Ball, describes the risk as a “chilling effect” — a claim that depends on the assumption that voluntary guidance chills innovation, an assumption for which no evidence is presented. The column’s citation of Alex Tabarrok and Richard Baldwin concerns AI’s export potential and the general productivity gains from service-sector AI, not the specific question of whether federal safety standards would kill that potential.

The omitted fact. The leading AI firms are currently lobbying Congress for a federal AI framework that includes a liability shield for model outputs. The same firms are members of trade associations — the Chamber of Commerce, the Software Alliance — that have publicly advocated for federal preemption of state AI laws. The column’s framing of “freedom or the Ayatollah” omits that the incumbents are in the room asking for the government’s help, and that the only question is whether the help they get includes an obligation to test their models before they deploy them on the public.