Chain of Thought Monitorability: A New and Fragile Opportunity for AI Safety

Korbak, Tomek, Balesni, Mikita, Barnes, Elizabeth, Bengio, Yoshua, Benton, Joe, Bloom, Joseph, Chen, Mark, Cooney, Alan, Dafoe, Allan, Dragan, Anca, Emmons, Scott, Evans, Owain, Farhi, David, Greenblatt, Ryan, Hendrycks, Dan, Hobbhahn, Marius, Hubinger, Evan, Irving, Geoffrey, Jenner, Erik, Kokotajlo, Daniel, Krakovna, Victoria, Legg, Shane, Lindner, David, Luan, David, Mądry, Aleksander, Michael, Julian, Nanda, Neel, Orr, Dave, Pachocki, Jakub, Perez, Ethan, Phuong, Mary, Roger, Fabien, Saxe, Joshua, Shlegeris, Buck, Soto, Martín, Steinberger, Eric, Wang, Jasmine, Zaremba, Wojciech, Baker, Bowen, Shah, Rohin, Mikulik, Vlad

arXiv.org Machine Learning 

AI systems that "think" in human language offer a unique opportunity for AI safety: we can monitor their chains of thought (CoT) for the intent to misbehave. Like all other known AI oversight methods, CoT monitoring is imperfect and allows some misbehavior to go unnoticed. Nevertheless, it shows promise and we recommend further research into CoT monitorability and investment in CoT monitoring alongside existing safety methods. Because CoT monitorability may be fragile, we recommend that frontier model developers consider the impact of development decisions on CoT monitorability.

Duplicate Docs Excel Report

Title
None found

Similar Docs  Excel Report  more

TitleSimilaritySource
None found