Based on two years of research and patterning, while not dedicated strictly to accountability, I've found a lot of repeatable signals that for me, show these systems are designed with maximum plausible deniability as a primary feature. While it's difficult to prove anything with blackbox systems, I think sufficient repeatable examples show a non-emergent, deliberate defense against accountability as a high-priority design objective in frontier models.
Based on two years of research and patterning, while not dedicated strictly to accountability, I've found a lot of repeatable signals that for me, show these systems are designed with maximum plausible deniability as a primary feature. While it's difficult to prove anything with blackbox systems, I think sufficient repeatable examples show a non-emergent, deliberate defense against accountability as a high-priority design objective in frontier models.