A proposed disclosure scale for A.I.-assisted writing could replace the increasingly divisive question of whether a writer used artificial intelligence with a more useful account of how it was used.
The idea follows a dispute involving billionaire investor Stanley Druckenmiller, who acknowledged using A.I. to write a Wall Street Journal opinion article criticising Treasury Secretary Scott Bessent’s interventions in the bond market.
Readers identified what they considered telltale phrasing in the article, while the A.I. detector Pangram gave it a score of 100 per cent A.I. Druckenmiller said he had rejected many of the chatbot’s suggestions and that it had not written the entire piece.
Asked about the article, Druckenmiller said: “Of course I used AI. I’m not embarrassed by it … I write everything using AI now for the same reason I use a calculator when I do math problems.”
Paul Gigot, the Journal’s editorial page editor, defended the decision to publish it, saying: “AI is a fact of modern life.” In a subsequent column, he wrote that “Mr. Druckenmiller’s sin against the media pharisees was admitting its use.”
The episode has highlighted the difficulty of treating A.I. use as a simple yes-or-no matter. A writer who uses a chatbot to suggest ideas or correct grammar has deployed it differently from one that asks it to produce a complete opinion article.
A proposed scale for A.I. writing
The suggested “Who Wrote This?” scale would run from H5, meaning no A.I. was used apart from ordinary spellcheck, to H0, or “A.I. Slop”, where a model produces the ideas, structure and voice from a prompt and receives only light editing.
H4 would describe assistance with research, fact-checking prompts, headlines or grammar, without any rewriting of the author’s sentences. H3 would apply where the writer completed a draft before using A.I. to tighten sentences, condense paragraphs or reorder sections.
At H2, the author and the model would have worked back and forth, with some passages originating from each, while the argument remained the author’s and was substantially edited by them. H1 would mean that A.I. produced a complete first draft from the author’s notes, interviews or reporting, leaving the ideas to the author but most of the wording to the model.
The proposal is not the first attempt to establish such a system. An open-source A.I. Usage Scale has suggested six similar levels, while at least one publisher, The Datatech Times, has begun printing an A.I. level beside each byline.
Supporters argue that disclosure could protect writers who use A.I. for research, editing or brainstorming, while giving readers a clearer understanding of the work in front of them. The central concern, they say, is not necessarily that a writer used the technology, but whether readers were misled about who supplied the ideas and language.
Writers have become increasingly wary of what has been described as an A.I. witch hunt. Some have removed em dashes from their work in an effort to avoid suspicion, while the Journal’s James Taranto has called Pangram a “defamation machine” that contributes to false accusations.
A self-reported system would not eliminate dishonesty, but it could offer readers an explanation without requiring them to submit every article to a detector. Writers who claimed a piece was H4 but were exposed as having relied on A.I. at H0 would face a greater embarrassment than those who had openly acknowledged using a full generated draft.
The debate has echoes of earlier arguments over ghostwriters and speechwriters. Public figures have long relied on other people to craft language published or delivered in their names, and readers are generally aware that such arrangements exist.
The proposed scale is intended to make the same principle more explicit for A.I.-assisted writing: not simply whether the technology was used, but how much of the finished work came from it.
