James Taranto rigs the human-ghostwriting analogy to license undisclosed AI use in op-eds. In a piece published September 18 in the Wall Street Journal’s opinion section, the paper’s editorial features editor argues that disclosure of AI assistance is unnecessary, that detection tools are unreliable, and that anyone who thinks otherwise is peddling “bad advice.” The piece deploys at least six distinct rhetorical operations across its roughly 1,500 words; this column walks through them in the order Taranto uses them.
I worked that beat. I wrote talking points like his. I sat in rooms where the disclosure question came up and the answer was “don’t,” then watched the answer become the policy.
“I use artificial intelligence when I write articles, including this one. I run partial or complete drafts through a chatbot and ask what it thinks. It spits out copious suggestions, some of which inspire me to make revisions and improvements. AI is also an immensely powerful research tool, able to find useful information via queries too vague or abstract to be effective in an ordinary Google search. For this essay, I used it to track down a magazine article I dimly remembered reading almost 40 years ago. As you’ll see, that old piece turned out to contain a killer quote.” — Taranto, opening graf
Motte-and-bailey — Shackel 2005 in the Bad-Faith Catalog — is anchored in the disclosure the column spends its length arguing against. Taranto opens by admitting he uses AI for everything from “suggestions” to “research.” The admission is doing two jobs at once. It establishes credibility — see, I am being candid. And it preempts the disclosure question — by mentioning it, Taranto performs transparency so he can spend the rest of the column arguing transparency is unnecessary in the absence of a specific topic requiring it. The opening paragraph is the bait; the column is the switch. The move is the disclosure-as-defense maneuver out of the cable greenroom: the operator confesses the thing being discussed so the discussion can be steered toward “we should not have to do this for every piece.” Taranto’s own opening is the cleanest example on the page of the device the rest of the column is built around.
Operative in the same paragraph, before the column moves on: frame-engineered relabeling — WSJ §4.1; Bad-Faith Catalog: frame_engineered_relabeling. Taranto says he “runs partial or complete drafts through a chatbot”; the chatbot “spits out copious suggestions, some of which inspire” revisions; AI is “an immensely powerful research tool.” Read the verbs: spits out, suggests, inspires, improves. The productive act is the AI’s. The credit act is the writer’s. Drafting assistance is relabeled as research, so what is substantively generative work travels under the conventional legitimating cover of journalism. The opening paragraph is doing the relabeling; the rest of the column lives inside that frame.
“I ensure the integrity of my own work with a simple rule: I never paste AI output into a piece I’m writing (except when AI transcribes text from an image or recording, and then I check the quote against the source). That has been my practice since I started using AI, and it is consistent with the Journal editorial page’s policy for staff, which my boss, Paul Gigot, announced last month: ‘No letting AI draft a column or editorial.’ Several months ago, Mr. Gigot and I discussed whether to adopt a policy on AI use by outside writers, as other newspapers have done. The New York Times op-ed contract provides: ‘As a freelance contributor to The Times, you may not submit any material for publication that contains content generated, modified or enhanced by gen-A.I. tools.’ I argued that attempting to police outsiders would be too cumbersome and that we don’t yet know enough about AI to make sensible uniform rules governing its use. And although I believe professional writers have an ethical obligation to write our own words, I don’t think that’s true of every op-ed contributor.” — Taranto, second section
Narrow rule, broad practice — operates here through a rule drawn narrowly enough that it cannot be falsified, deployed as cover for a practice the column never names. Taranto does not paste AI output into his pieces. Taranto does run partial or complete drafts through a chatbot and incorporates the “suggestions.” The narrow rule is defensible in a way the broad practice is not, and the column’s defense rests entirely on the narrow rule staying narrow while the broad practice continues.
The outside-contributor carve-out is the second half of the same structure: WSJ has a rule for staff; WSJ has no rule for the building’s other tenants; the no-rule zone is where the bylines that drive the page’s traffic actually live. The rule always covers the masthead. The carve-out always covers the bylines. That is how opinion pages have always worked, and the column is not telling you this is new; it is telling you this is fine.
“We instituted no policy, and that saved us from embarrassment by allowing us to stand behind a (very good) piece by the investor Stanley Druckenmiller that Pangram’s detector tagged as ‘AI generated.’ Mr. Druckenmiller answered media inquiries with an admirable nonchalance. ‘Of course I used AI,’ he told the news site Notus (now the Washington Sun). ‘I was a B student in English, but an A-plus in economics,’ he said to a Journal reporter. ‘These are my ideas and I’ve been speaking about them for over 15 years, as anyone who knows me knows. Would I prefer that I was a great writer? Yes, but that’s not who I am.’” — Taranto, on the Druckenmiller false positive
One false positive, total exemption — operates here through one false positive generalized into a license for the entire practice. Taranto’s column needs Pangram to be unreliable, because if Pangram is reliable then the disclosure question becomes operational and the page has a problem. So Taranto offers the Druckenmiller piece as proof: here is a case where the detector misfired, therefore no detector can be trusted, therefore no disclosure regime can be built on detection, therefore disclosure is unnecessary. This is the manufactured-controversy move run in reverse — not the construction of doubt about a consensus but the construction of certainty about unreliability from a single case. The mechanism is always the same: one false positive, total exemption. The mechanism never has to scale.
Layered in the same passage: moral justification plus displacement of responsibility — Bandura mechanisms 1 and 4; the WSJ §4.10 common-sense-versus-elite pivot operates along the same axis. “This is bad advice, and I won’t follow it. The repetition would quickly become boring, and my readers deserve better.” Three sentences and the substantive question is gone. What does disclosure actually protect readers from? What are the epistemic stakes when a journalist runs drafts through a generative tool? The paragraph does not engage those questions. It substitutes the boredom frame and the “deserve better” frame. In editorial-page rooms of this kind, the third-party-foil critique is the apparatus’s most reliable way to kill a transparency rule without engaging what the rule would actually do. You don’t fight the substance; you discredit the loudest advocate and let the silence on substance do the rest. Taranto runs the play.
“A related and similarly candid acknowledgment came from Kyle Crichton, deputy op-ed editor of the New York Times: ‘If we insisted that every politician write his own piece, we’d either get terrible pieces or none at all.’ But Mr. Crichton, who has since moved to another position at the Times, wasn’t talking about AI. His quote appeared in the April 18, 1988, issue of the New Republic, in an article by Ari Posner about ghostwriting. In September 1987, a young senator had ended his presidential campaign owing to a plagiarism scandal. Joe Biden lifted not only words but biographical details from speeches by the British politician Neil Kinnock. … The headline notwithstanding, Mr. Posner allowed that ‘obviously there is a spectrum of respectability here.’ Nobody expects the president to write his own speeches, even if nowadays we can count on plenty of improvisation. ‘At the other extreme, people do have the right to expect that a journalist’s byline means that he or she actually composed the words that follow.’ There’s no obvious reason the same distinctions shouldn’t apply when the ghostwriter is a machine rather than a human.” — Taranto, on Ari Posner’s 1988 New Republic essay “The Culture of Plagiarism”
Advantageous comparison — Bandura mechanism 3; the operators call it false analogy — operates across the column’s longest stretch. Taranto retrieves a 1988 New Republic essay on plagiarism to argue AI disclosure is unnecessary because human ghostwriting has never required it. Ari Posner’s “spectrum of respectability” provides the gradient. The conclusion follows: “There’s no obvious reason the same distinctions shouldn’t apply when the ghostwriter is a machine rather than a human.”
The advantage Taranto is taking is the conflation. A human speechwriter working from notes — who can be sued for libel if the work product is defamatory, who collects a paycheck, has a name, a reputation, an editor, a point of view — is not the same operation as an opaque large language model trained on the entire copyright-bearing corpus of human writing, generating text the user then evaluates, accepts, rejects, and modifies. One ghost is a person. The other is a statistical machine that synthesizes from millions of people’s work without their consent. The cases are not analogous — but the analogy’s advantages for Taranto’s argument are real. If 1988 ghostwriting was fine, 2026 AI assistance is fine.
The operator’s-eye-view, briefly: the historical-analogy play is the apparatus’s most reliable way to launder any novel practice that needs to look familiar. Luntz’s climate-communications memo did it with “climate change” for “global warming”; the WSJ editorial page does it with “tax relief” for tax cuts; Taranto does it here with 1988 ghostwriting for 2026 AI assistance. The technique is not new; the practice the technique is laundering is novel. Operators of this kind reach for historical analogy because the analogy’s job is to launder a practice that cannot stand on its own.
“I have also ghostwritten op-eds. My first full-time job was in the PR department of the Heritage Foundation, a Washington think tank, which I joined two months after Mr. Posner’s article appeared in 1988. Among my duties was to produce one op-ed column a week for the foundation’s president, Ed Feulner. I would select a topic, draft a formulaic piece about it, and send it upstairs, where Feulner would edit it (in most cases minimally), approve it, and send it back. It was then printed and sent by mail to newspapers across the U.S., which could publish it free of charge. I was somewhat uncomfortable with this practice, and Heritage found it embarrassing enough that my boss there once told me I was forbidden to mention ‘ghostwriting’ on my résumé. But in retrospect I don’t think it was all that bad.” — Taranto, on his own Heritage Foundation ghostwriting
Complicity as credential — anchored here through the use of past participation in the practice as evidence of standing to defend the practice. Taranto ghostwrote for Ed Feulner. He tells you so. The disclosure is doing the work of an admission — except the admission is supposed to bind the speaker’s authority, not establish it. Taranto’s disclosure does the opposite: I did it, therefore I know what I’m talking about, therefore my defense of the practice is expert testimony rather than interest testimony. The circularity is the trick.
Layered in: diffusion of responsibility — Bandura mechanism 5. What the operation Taranto describes actually was: a Heritage Foundation op-ed-production machine writing weekly columns for the foundation’s president under his name and mailing them free to newspapers across the country. That is precisely the institutional message-discipline work the apparatus was built to do. Taranto, when he confesses the practice, is confessing one of its early assembly-line workers. He says he was “somewhat uncomfortable.” He says he was told not to put “ghostwriting” on his résumé. And he says, “in retrospect I don’t think it was all that bad.” The reader is asked to credit the disclosure as license. The disclosure does the editor’s 2026 case the work of an alibi. Taranto’s own complicity is the standing.
“AI is merely a tool, and a versatile one. It can be used to automate an assembly line, to compensate for a gap in the user’s skills, or to augment and sharpen individual creativity. Pangram won’t tell the difference. That requires old-fashioned human judgment.” — Taranto, closing
“Merely a tool” — operates here through the engineering of a closing sentence designed to be quoted as the take-home, on a foundation the column has spent its length making impossible for the reader to evaluate. The closing-line cadence is the WSJ board’s signature move (WSJ §3.5): declarative, short, quotable, designed for retransmission. The “merely a tool” formulation places AI in the same category as a calculator or a wrench — a category the column has spent its length showing it does not actually fit.
Read the verbs in parallel to the opening: automate, compensate, augment, sharpen. Four productive-sounding actions. The clause that does the work — “or to augment and sharpen individual creativity” — borrows the rhetoric of artistic creativity to license whatever the “augmenting” covers. The opener’s verbs assigned productivity to the AI; the closer’s verbs assign discretion to the writer. The two halves of the column are stitched at the same seam. “Pangram won’t tell the difference” closes the loop: the opacity of distinction between uses is assigned to the third-party detector, not to the writer whose use the detector would be detecting. The actual opacity is at the writer’s table — was this paragraph drafted, edited, or merely audited by the writer? — and that opacity is the substantive question Taranto’s column is engineered to keep in place. The “old-fashioned human judgment” Taranto reserves for himself is the judgment that concludes the reader does not need to know.
A wrench is a tool. A wrench does not write op-eds. A chatbot is a tool that writes op-eds, and the disclosure the column argues against is the disclosure that would tell readers it wrote this one.
A tool is what the WSJ editorial page has called everything it wanted to keep from disclosure — taxpayer money is a tool, regulation is a tool, the editorial page itself, whose declared purpose is “free people and free markets.” The apparatus knows what “tool” means in its own usage; the apparatus has engineered the word for decades. AI is now a tool in the same sense, and the word does the same work. Tools can be misused. Tools require operator discipline. Tools are exactly the kind of thing that benefits from disclosure of who is using them, for what, under whose authority. Taranto’s piece does not engage that question. The piece is engineered to look like it engages the question while doing other work — the “research” frame, the Pangram critique, the 1988 essay, the 1988 confession, the closing line. The substantive argument for non-disclosure never lands because the substantive argument is not what the piece is built to deliver.
Taranto opens by relabeling what his chatbot produces as “research.” He kills the disclosure principle by attaching it to an unreliable detector. He pulls a 1988 New Republic essay on plagiarism across to license machine assistance as part of a long-standing “spectrum of respectability.” He confesses his own past as a Heritage Foundation ghostwriter for Ed Feulner, then deploys the confession as cover for the editor’s 2026 discretion. He closes with the apparatus’s most reliable closing-line cadence: a tool, a tool, a tool.
The chat drafts that “inspire” revisions are the tool. The “research” label is the cover. Pangram is the foil. Posner is the analogy. Heritage is the license. “Human judgment” is the editor’s seal.
The defense of an editor’s standing does not need to engage the question of whether readers deserve to know. The standing is the answer. Taranto’s column is the operational evidence for the architecture, deployed by the features editor whose byline runs the actual page — and the documentary record his own column leaves is what this column walks the reader through.
— Phukher Tarlson