Science's Thorp: AI made publishing slower, worse, costlier

The Editor-in-Chief of the Science family of journals just published an editorial titled "AI in scientific publishing: Slower, worse, and more expensive", and the punchline is exactly the inversion of the old management cliché. H. Holden Thorp writes that the increase in volume and the behaviour of LLMs is not making the work "faster, better, cheaper" but "slower, worse, and more expensive."

Thorp's argument is grounded in three decades of watching technology sales pitches fail their own math. He lists laptops in the classroom, MOOCs like Coursera and EdX, and electronic health records like Epic as prior "this time is different" moments where the promised staff savings never appeared. Of those pitches he writes that "every single time, the salespeople convincing me about this stuff told me I would be able to do more and save money because less human effort would be required. Every single time, they were half-right." On AI he sees the same pattern already: more submissions, more AI-generated errors, more human oversight needed to catch them.

He gives AI credit where he can. LLMs, he notes, have "revolutionized protein structure prediction and materials discovery," and inside publishing they help Science catch plagiarism, altered images, and missing supporting data. But his line on the tooling is the load-bearing one: "although AI is helping Science catch errors that can be corrected or elements that are missing from a paper [...] its use and the evaluation of the output require more human effort, not less." The tools are useful; they are also a new cost line rather than a replacement for a copy editor.

The honest caveat is that this is one editor's editorial, not an audited P&L. Thorp does not publish a per-paper cost figure, does not name the specific screening vendors Science is using, and does not tell you how the calculus differs at a mass-scale journal versus a highly selective one. He also asserts that AI agents "are more likely than humans" to cherry-pick data and hallucinate references, a claim worth interrogating rather than accepting on the masthead's authority.

What the piece does do, and the reason it will get shared, is give operators cover to push back on ROI decks that assume AI cleanly replaces human editorial labour. If the Editor-in-Chief of Science is saying human-curated science is "getting more important, not less," premium journals gain a differentiation story and smaller publishers get a warning about cutting human editors to fund AI screening.