SafetyAccountabilityTransparencycs.CL
Introducing v0.5 of the AI Safety Benchmark from MLCommons
Bertie Vidgen, Adarsh Agrawal, Ahmed M. Ahmed, Victor Akinwande, Namir Al-Nuaimi, Najla Alfaraj, Elie Alhajjar, Lora Aroyo, Trupti Bavalatti, Max Bartolo, Borhane Blili-Hamelin, Kurt Bollacker, Rishi Bomassani, Marisa Ferrara Boston, Siméon Campos, Kal Chakra, Canyu Chen, Cody Coleman, Zacharie Delpierre Coudert, Leon Derczynski, Debojyoti Dutta, Ian Eisenberg, James Ezick, Heather Frase, Brian Fuller, Ram Gandikota, Agasthya Gangavarapu, Ananya Gangavarapu, James Gealy, Rajat Ghosh, James Goel, Usman Gohar, Sujata Goswami, Scott A. Hale, Wiebke Hutiri, Joseph Marvin Imperial, Surgan Jandial, Nick Judd, Felix Juefei-Xu, Foutse Khomh, Bhavya Kailkhura, Hannah Rose Kirk, Kevin Klyman, Chris Knotz, Michael Kuchnik, Shachi H. Kumar, Srijan Kumar, Chris Lengerich, Bo Li, Zeyi Liao, Eileen Peters Long, Victor Lu, Sarah Luger, Yifan Mai, Priyanka Mary Mammen, Kelvin Manyeki, Sean McGregor, Virendra Mehta, Shafee Mohammed, Emanuel Moss, Lama Nachman, Dinesh Jinenhally Naganna, Amin Nikanjam, Besmira Nushi, Luis Oala, Iftach Orr, Alicia Parrish, Cigdem Patlak, William Pietri, Forough Poursabzi-Sangdeh, Eleonora Presani, Fabrizio Puletti, Paul Röttger, Saurav Sahay, Tim Santos, Nino Scherrer, Alice Schoenauer Sebag, Patrick Schramowski, Abolfazl Shahbazi, Vin Sharma, Xudong Shen, Vamsi Sistla, Leonard Tang, Davide Testuggine, Vithursan Thangarasa, Elizabeth Anne Watkins, Rebecca Weiss, Chris Welty, Tyler Wilbers, Adina Williams, Carole-Jean Wu, Poonam Yadav, Xianjun Yang, Yi Zeng, Wenhui Zhang, Fedor Zhdanov, Jiacheng Zhu, Percy Liang, Peter Mattson, Joaquin Vanschoren
April 18, 2024
arXiv.org
80 citations
3 influential
Key Takeaways
-
The benchmark is still a work-in-progress: Version 0.5 explicitly should not be used to assess AI safety, with the team planning to release a more reliable version 1.0 by the end of 2024. This reflects the early stage of developing standardized safety testing methods for AI systems.
-
A comprehensive taxonomy of AI hazards has been created: The benchmark identifies 13 distinct categories of potential harms that chat-based AI systems could cause, though only 7 categories have tests included in v0.5, indicating that safety assessment is multifaceted and requires systematic categorization.
-
Testing covers different user types and scenarios: The benchmark assesses how AI systems respond to typical users, malicious users attempting to misuse the system, and vulnerable users who may be at higher risk of harm—recognizing that safety risks vary depending on who is using the AI.
-
Large-scale testing infrastructure has been developed: The benchmark includes 43,090 test items (prompts) created using templates, demonstrating that meaningful AI safety assessment requires extensive, systematic testing rather than ad-hoc evaluation.
-
Transparency about limitations is prioritized: The researchers have "fully documented the limitations, flaws, and challenges" of v0.5, emphasizing that responsible AI development requires honest acknowledgment of what current safety benchmarks can and cannot reliably measure.