AI Stories Scored Higher. Human Authorship Still Mattered.
Dan PetersonSynopsis — AI-drafted from Dan's notes
Grove-White, who runs AI workshops for professionals, starts from a tension in a new study: readers rated AI-written stories higher for quality and absorption, yet still favoured stories they believed a person wrote. He calls the discount the “Authenticity Tax”. Readers apply it when they suspect the named writer did not shape, review or stand behind the work, and it costs reach, trust and reputation. The bar is higher for anyone whose value is their judgement.
His distinction is authored versus unowned. AI-assisted work is still authored when the writer can explain the evidence, defend the judgement and take responsibility for the result. Unowned work has a name on it without enough of the person inside it. He grants that the backlash is partly earned: low-effort output moves the work downstream to the reader (a vague reply to his questions about a volunteer role left him doing the thinking). But “this sounds like AI” is a lazy answer to a sound argument, and the usual tells only show an editor where to look.
He sees platforms drawing the same line. LinkedIn now limits the reach of generic AI-looking posts and lets readers report them as slop; Substack lets readers scan a post for its human/AI share and lets writers publish a “How I make this” statement, which he prefers. Publishing is in a “gotcha” stage: Hachette pulled Shy Girl over AI allegations, and a deal worth over $2 million for Jerry Falade’s novel was withdrawn by his own agents when they could not show how the manuscript evolved. Both writers deny that AI wrote their books.
His own method is a chain of approved decisions: the model interviews him, transcripts recover what was really said, he picks the reader and the outline, and a “linguistic signature” from his older writing keeps the sound. Substack’s Pangram scan still called the finished draft 100% AI. For him that proves the point: a detector reads the page, and the hours of choosing, rejecting and restoring sit behind it where no scan can see them.
Two details are looser than the sources. Only the first of the study’s three experiments (Sears and Weisberg, Villanova) measured quality and absorption, and in one follow-up readers did worse than chance at spotting the AI story. Falade’s deal was withdrawn by his agents, not cancelled by the publisher.
Where I land
It seems Grove-White and I are very much aligned in our view of AI-assisted work and when and where it should be deployed. Based on my own experience with debates via social media, I think the problem most people have is a lack of transparency.
Also, I would note that writing at least a large share of the prose yourself would reduce the Pangram score, as it preserves your authorship.