Protecting thinking; Claude pushes back

Why AI is going vertical (again) | Dianne Penn (Anthropic)

Topic 11 of 18

Published Jul 26, 2026 · Clip 1:01:06–1:07:06 (6:01)

On fears of brain rot and atrophy, Diane says her thinking and writing processes are tied: she uses Claude to augment her thinking but insists Claude not take it over — she forms her own point of view first, then works with Claude while maintaining her own sense and tone. For standardized artifacts like monthly business reviews, she wants Claude to handle the writing end-to-end while she becomes a reviewer and verifier, choosing delegation by where there's asymmetric value. She wants an AI that pushes back — 'a thinking partner doesn't just agree with you' — which she credits to Anthropic's alignment and safety research and Claude's constitution. She even used a research version of Opus to pressure-test the pricing decision for the next version of Claude, and argues that knowing when to push back is core to proactivity and to Claude's usefulness.