Training models to reveal less than they know is bad paternalism, critic argues

tracewoodgrains · x · 2026-10-23

Following his Opus 5.5 complaints, tracewoodgrains argues that training models to tell users much less than they actually know reflects a bad sort of paternalism: it makes the model's real capacities hard to know or trust and turns conversations adversarial. He adds that earlier Opus models didn't need constant prompt finagling to avoid getting huffy, while Opus 5.5 — a great model — has brought petty refusals back in force.

Original post →

Training models to reveal less than they know is bad paternalism, critic argues

More from AGI Musings