Science’s New AI Issue: 90% of Papers Now Show AI Help—But Who’s Checking the Machines?
A new edition of the journal *Science* has sparked urgent questions about artificial intelligence in research, with one analysis finding that 90% of recent biomedical papers show signs of AI-generated language, while another report reveals an AI system that can now run experiments entirely on its own.
The August 2026 issue of *Science* (Volume 393, Issue 6813) is packed with developments that point to a single, uncomfortable reality: artificial intelligence is now deeply embedded in the scientific process, but oversight has not kept pace. One editorial in the same issue asks a direct question—who is actually testing what these AI systems can do? [224249]
The scale of AI’s footprint is hard to ignore. A new analysis of over 1.5 million biomedical papers published in 2024 found that about 90% contain language patterns typical of AI writing tools. That is a sharp jump from 2023, when the same team found AI-like language in about 70% of papers. The authors stress that most papers are not entirely fake—researchers likely used AI to edit, summarize, or polish their writing—but they warn that unchecked use blurs the line between human analysis and machine-generated content. [224209]
At the same time, a separate report in the same issue describes an AI system that can plan, carry out, and analyze biomedical experiments with no human help. Unlike earlier tools that only suggested hypotheses, this agent physically operates lab equipment, adjusts its methods mid-experiment, and repeats the cycle until it reaches a conclusion. The authors call this a critical milestone toward fully automated science, though they note the system works best with clearly defined protocols and may struggle with unexpected failures. [224227]
The combination of these findings raises a practical problem. If most papers are touched by AI, and some experiments are now run entirely by AI, then who verifies the output? The editorial in *Science* points to a growing gap in oversight, noting that no single authority is responsible for testing the real-world capabilities and limits of these tools. Without independent checks, developers may overstate what their systems can handle, leaving regulators and the public guessing about potential failures. [224249]
Experts say the issue is not whether AI is used—it clearly is—but how transparent researchers are about it. Many journals have no clear policy on AI use, and disclosure is rare even when AI is involved. [224209] The editorial does not propose a specific solution, but it calls for urgent discussion. The core issue, it suggests, is not just building smarter AI, but ensuring someone, somewhere, is formally responsible for proving what that intelligence can and cannot do. [224249]