On August 24, the Wall Street Journal published an opinion piece by Stanley Druckenmiller, one of the country’s best-known investors, criticizing Treasury Secretary Scott Bessent, who once worked for him, for buying back long-term government bonds to keep interest rates down. Readers noticed something about the prose. An AI detection tool called Pangram scored the piece as AI-written. When a reporter asked, Druckenmiller said, “Of course I used AI.” He compared it to using a calculator for math, and he said that his name is on the piece and the message is his. Paul Gigot, who runs the Journal’s opinion pages, said the piece was Druckenmiller’s genuine opinion and broke no rule, and no changes or notes were added to the article.
Eleven days earlier, the Financial Times published a column from Harvard economist Ricardo Hausmann about tariffs. After learning he had used AI to shorten the column, the paper added a note that it prohibited the use of AI in the writing process.
Two prominent contributors. Two respected newspapers. Two different answers to a question writers and readers increasingly face: What is going on?
Each paper applied a different test. The Journal asked whether the argument belonged to the author and whether the author had the standing in their field to make it. Call that a standing test. The Financial Times asked whether AI was involved at all in the writing process. Call that a process test.
Each test has a weakness. A process test is difficult to enforce. Readers cannot see how a piece was made. They can only guess from the style, so violations surface as scandals, after the fact, through detection tools that are probabilistic at best. A standing test can be enforced, but it protects only people who already have standing. Gigot said the Journal had published Druckenmiller for many years and nobody could doubt his opinion was his own. That is true. It also does nothing for a first-time contributor, a student, or a resident writing a letter to their paper, whose text is the only evidence of standing they have.
Neither paper asked the author to tell readers what he did. That absence is the story. We solved this problem once before. For most of the last century, an executive dictated a letter, a secretary typed it, and the executive signed it. At the bottom of the page sat a small notation, something like PAR:jms. The capital letters belonged to the person whose ideas and signature these were. The lowercase letters belonged to the person who put the words on the page. George Mason University’s Writing Center still explains this convention in its business-letter guidance. Those initials made assistance visible without requiring an essay about the production process. The signature remained prominent. Someone else’s contribution was recorded.
The arrangement was imperfect. A few lowercase letters could hardly capture an assistant’s judgment, corrections, or tactful rewriting. And typing a dictated letter differs from composing an argument. The harder case is ghostwriting, and it is still contested. When Theodore Sorensen drafted much of Profiles in Courage, nobody doubted the book expressed John Kennedy’s convictions. The argument was about the Pulitzer Prize, which was awarded for the writing. The typist supplied labor. The ghostwriter supplied words that did some of the thinking. The trouble has always been the line between the two. Still, the convention recognized something our AI arguments sometimes obscure: a person can take responsibility for a document that other hands helped produce. Readers can also be told about that help.
AI divides labor harder to describe because it can move so easily from correcting a sentence to changing what the sentence means. Consider a memorable line in Druckenmiller’s column. Discussing the possibility of a 5.5 percent yield on a 30-year Treasury bond, he writes, “that isn’t a crisis. It is an invoice.” The word ‘invoice’ does a lot of work. It suggests a bill coming due, an obligation incurred, a payment that should be expected. It encourages readers to see higher borrowing costs as the consequence of earlier choices. A language model gives you a way of seeing the question, and you may not have chosen it. Whoever chose the word ‘invoice’ made an argument. If the machine chose it, then the ideas and the wording were not as separate as Druckenmiller says.
Here is, perhaps, the hardest part. Disclosing AI use costs something. Oliver Schilke and Martin Reimann ran thirteen experiments with more than 3,000 people, in settings from classrooms to investment advice to office email. In every setting, people who said they had used AI were trusted less than people who said nothing. Two other findings matter more. Being caught was worse than disclosing. And the penalty shrank when people believed the practice was common. This is playing out in a country where only 25 percent of people say they trust the news at all. Under the current arrangement, nobody discloses, and detection tools decide. That arrangement produces the worst result for everyone except the people with the most standing.
That brings me to this column. It is about AI in society. You are entitled to ask how I made it, and I would rather answer than be tested. I regularly use AI in three ways. It helps me find sources. I test ideas against it in conversation. AI helps me with revisions and suggests wording. The argument is mine. The judgments are mine. The responsibility is mine. I would rather pay the trust cost up front, and I would rather readers judge the argument knowing how it was made.
The problem in August was never that two writers used AI. The problem was that two newspapers decided afterward, based on who the writer was, what their readers were owed. A small line at the bottom of a page settled that question for executives and secretaries a century ago. It can settle it again, if writers go first.
P. ANAND RAO is the Director of the Center for AI and the Liberal Arts at the University of Mary Washington.
