According to science.org, the journal Science has established specific policies governing the use of AI in research it publishes. The rules allow the use of large language models for certain tasks like editing text for clarity or gathering references without any disclosure. However, any use of AI to draft manuscript text must be declared, and the use of AI to create figures is completely prohibited. All authors remain fully responsible for all content, including AI-assisted material. On its own end, Science uses AI tools like iThenticate and Proofig to detect plagiarism and altered figures, and it collaborated with DataSeer to evaluate data sharing. That evaluation of 2,680 papers published between 2021 and 2024 found 69% shared their underlying data, and a DataSeer-powered reproducibility checklist is now being integrated into the journal’s protocols.
The Human Firewall
Here’s the thing: the journal’s stance is fundamentally about maintaining a human firewall. They’re not Luddites; they’re actively using AI as a tool for vigilance. But they see the core risk clearly. The big worry isn’t job loss for editors—it’s the degradation of the scientific record itself, what they pointedly call “AI slop.” Allowing AI to generate figures or undisclosed text is a fast track to a literature filled with plausible-looking but potentially meaningless or fabricated content. So their policy is a pragmatic mix: use the machine to help spot problems, but keep the core creative and evaluative work firmly in human hands. I think that’s the only sane approach right now.
More Work, Not Less
And this leads to a crucial point the article makes that often gets lost in the hype: using AI well requires more human effort, not less. Think about it. An AI tool might flag a hundred potential issues in a paper—but someone has to scrutinize every single one of those flags. It’s creating a new layer of work: AI auditing. The promise of total automation is a fantasy. The real outcome is that the human role shifts from brute-force checking to sophisticated judgment calls on AI-generated alerts. Is that progress? Probably. Is it easier? Almost certainly not.
A History of Hype
I love that the editorial brings up historical context. Remember when MOOCs (Massive Open Online Courses) were going to obliterate universities? Didn’t happen. They just became another tool. The move to online publishing didn’t kill journals; it exploded the scale of publishing. This pattern should make us deeply skeptical of any grand claim that AI will suddenly replace the entire scientific process. New tools change workflows and create new challenges—like the need for robust industrial computing to manage data—but they rarely erase the fundamental human components. Look, technology evolves, but human nature and the need for trusted curation don’t.
The Value of a Human Record
Ultimately, Science is making a bet on brand value through human oversight. In a world potentially flooded with AI-generated research sludge, a journal that can credibly say “humans carefully checked this” becomes more valuable, not less. Their stance is a quality control measure. No system is perfect, and humans make mistakes too—hence retractions. But a mistake made by a human is at least traceable to a fallible intelligence. An error hallucinated by a model is of a different, more insidious kind. By insisting on human accountability at every key step, they’re trying to future-proof the integrity of their pages. And in the long run, that’s probably what will help science stand the test of time.
