Who wrote this? It's the wrong first question

In August, an AI detector helped turn one op-ed into a public argument about who wrote it. Investor Stanley Druckenmiller had published a Wall Street Journal piece criticizing Treasury Secretary Scott Bessent's bond-buying strategy. Readers ran it through a detector called Pangram, which scored it 100% AI, and critics called it "Claudeslop". Druckenmiller said he wasn't embarrassed and likened AI to a calculator. The Journal's opinion editor, Paul Gigot, defended the piece, saying what counts is whether it carries the author's original argument and credibility (that's Semafor's account of his position). Whatever you think of the piece, the episode shows how hard it is to say what "AI-written" even means.

Policies don't agree on the answer either. Semafor's reporting says the New York Times bars AI in developing and drafting guest essays. The Washington Post asks writers to confirm their submissions weren't created or manipulated with it. The Journal ran Druckenmiller's piece and stood behind it. The same essay can clear one editor's standard and fail another's.

Here's my position, and it's an opinion: who typed the sentences matters less than who answers for the claims and whether readers are told. I mean cases where a person's attention, decisions, taste, and reasoning went into the final piece. In those cases I don't think it matters much who typed the prose or ran the searches. I don't know how Druckenmiller worked, but I know how I work. So here is my own case, and you can judge whether the test holds.

What the AI did, and what I did

Most of the words in my last project were Claude's, and I'd rather say so plainly. I recently used Claude to develop the thesis for an article series. I gave it some thoughts on what I might write about, and it proposed ideas tied to current news, ran searches, summarized sources, and wrote most of the prose in our planning document. My part is editing for my own voice and style and, mostly, making sure I agree with the meaning behind the words and with the piece as a whole. So the real question is what I contributed besides typing.

What I contributed was the frame, the rules, and a steady habit of asking whether things were true. The frame was AI handling the timing and coordination between a company's different functions, with Goldratt's The Goal as a reference. The rules were to cite data when it exists, to give my reasoning and say so when I'm speculating, to put an AI-use disclosure on every article, and to keep the final conclusions mine alone. Rules only matter if they catch something, and these did.

The rules caught a claim that Claude had stated as fact. Its first draft said coding got faster while requirements, review, testing, and release didn't, so work now waits in front of them. I asked whether that was defensible. The sources didn't support it. It became a cited claim plus a labeled speculation. Left alone, that sentence would have gone out as fact under my name.

Claude made smaller errors too, and I caught those as well. It called a news story single-sourced when it was Bloomberg's original report relayed by other outlets and unconfirmed by Apple. It used a roundup's 93% adoption figure when the primary source said 91%. It described two figures as conflicting when they came from different reports measuring different things. None of those were typos, and each would have gone out under my name.

The thesis changed because I changed it. I believed product management and risk tolerance were the bottleneck. The research found no data testing that, and one vendor survey pointed the other way. I changed the thesis instead of defending it, and I ended up dropping the claim that any single place is the bottleneck. The model told me what the evidence said, and the decision was mine.

Why I anchor on accountability

Accountability is the right anchor because that's where institutions have already put it. The ICMJE and COPE say AI can't be an author because it can't answer for the work, and humans stay responsible and must disclose. In Mata v. Avianca, lawyers were sanctioned $5,000 for filing six fabricated cases. The judge said using AI as a research aid wasn't itself improper. The failure was filing without checking and not correcting the record. In both, responsibility follows the person who signed.

The strongest objection

"A human was involved" can be a fig leaf. In a study of 27 radiologists (obviously a medical task, not writing), accuracy fell to roughly 20–45% when the AI's suggestion was wrong, from about 80% when it was right. In Johnson v. Dunn, five hallucinated citations got three lawyers reprimanded, disqualified, and referred to the bar, including two who signed the filings without drafting or checking them. So "I reviewed it" proves nothing by itself, because I can list the errors I caught but not the ones I missed.

The second objection: is it still my view?

An AI can shape what you think, not just how you say it. In a study of 1,506 people, writers used an AI helper that was slanted toward one side of an issue. As a result, they wrote essays that leaned that way, and afterward they agreed with that side more. In another experiment, writers using one AI model wrote essays that sounded more alike. So the tool can steer both what people argue and how much they sound like each other, and I can't just assume "the conclusions are mine."

My test is whether I've changed my mind against the model's grain. That's opinion, and it rests on one anecdote. I dropped my own bottleneck thesis when the research went against it. That isn't proof, but it's the check I can actually run.

Does disclosing cost you?

Disclosing does cost something. Across 13 experiments, people who disclosed AI use were trusted less. But the label effect on headlines was small and grew as readers assumed the AI did more. In a news study, listing sources offset much of the drop. My opinion is that you answer readers' guesses about how much the AI did with specifics: what it did, what you did, and where the claims come from.

Where reasonable people disagree

Reasonable people disagree about where the line goes. The Times says don't draft with AI at all, and the Journal's editor says judge the argument. Medium requires labels only for mostly-AI text with little editing. Elsevier exempts grammar tools. No single rule here is obviously right.

The objections run in both directions. Some will say drafting is thinking, so outsourcing the sentences outsources part of the author. Others will say a detector score tells you little (its maker claims a 0.01% false-positive rate, which I can't verify). Others will say readers deserve disclosure even when an editor doesn't ask. I land on accountability plus disclosure, but I won't claim it doesn't matter who wrote the words. It matters to some readers, for reasons I respect, but I think accountability and disclosure should come first.

How this was written: I use AI (Claude) to develop ideas, outlines, and draft text. I edit, correct, and direct it throughout, and the conclusions are mine alone. Facts link to their sources. Where I'm speculating or giving my opinion, I say so.