AI Auditors Uncover Decades-Old Errors in Scientific Databases and Conference Papers
Artificial intelligence agents are increasingly used to audit scientific literature, revealing long-standing errors in reference databases and published papers. While these tools can identify mistakes at a scale impossible for humans, they remain imperfect and require human oversight to ensure accuracy.
Key points
- An AI model used by chemist Sebastian Pios identified incorrect boiling point values in a 75-year-old reference database, proving the database wrong rather than the model.
- Researchers at SAI Labs used AI agents to assess 168 papers from the 2026 International Conference on Machine Learning, finding that only 34 of 92 assessable papers had more than two claims successfully reproduced.
- A study by James Zou and colleagues found that objective errors in papers at the NeurIPS conference increased by 55% from 2021 to 2025.
- Critics note that AI fact-checkers make mistakes similar to humans and cannot replace human judgment for subjective assessments like novelty or significance.
- Proponents argue that AI offers a significant advantage in speed and scale, allowing for the auditing of vast scientific corpora that were previously unmanageable.
Background
This development follows a broader trend of integrating AI into scientific workflows, including peer review and drug discovery. Previous coverage noted that over half of researchers now use AI for peer review, often against institutional guidance, and that companies like Edison Scientific are using AI agents to scale biotech ventures. The current focus shifts from AI generating new science to AI auditing existing knowledge for errors.
Why it matters
The reliability of foundational scientific data is critical for subsequent research. If AI tools can systematically identify and correct errors in established databases and published papers, it could strengthen the integrity of the scientific record. However, the risk of AI introducing new errors or missing subtle issues necessitates a hybrid approach combining automated scanning with human verification.
What to watch
Expect continued development of specialized AI tools for auditing specific types of scientific errors, such as formulae and calculations. Researchers will likely refine these tools to handle more complex, subjective aspects of scientific claims while maintaining human oversight to prevent the propagation of AI-generated mistakes.
Want the full story? Read the original reporting
Read on Nature