Ronald J. Bodkin
Titles as the sources gave them. In 1 entry, with 5 checked quotations.
What they said
Newest first. Each quotation was found word for word in the source.
“We propose comprehension audits, a novel development-process assurance mechanism in which the responsible people explain R&D contributions to auditors to demonstrate understanding.”
“Our analysis of leading open-source AI projects finds increased output of code with reduced human review commentary rates per line of code, with far lower rates for automated fleet accounts.”
“Ultimately, approval of AI-written output without comprehension is meaningless.”
“We believe that regulations that mandate comprehension audits performed by accredited external auditors are a promising medium-term intervention to reduce risks from automated AI R&D.”
“Ultimately, comprehension audits provide an early warning and a safety brake. They don’t make automated AI R&D safe through alignment or ensure control.”
In the Base
Every entry they appear in.
Comprehension Audits to Mitigate Risks from Automated AI Research
In an arXiv preprint, Ronald J. Bodkin, Bahrad A. Sokhansanj and Gillian Hadfield propose 'comprehension audits': independent auditors check, without AI help, that the people behind AI-lab research understand it, and work pauses if they cannot. Their study of open-source AI projects finds more code and fewer human review comments per line. They note the audits are untested in practice.