Anthropic’s October 27, 2026 report records invalidating errors in 3 of 50 AI users (6%) versus 6 of 11 nonusers (55%); an author checked AI-flagged serious errors. Middle completion times were 65 versus 90 minutes. The report places the middle half of AI times at 50–79 minutes; the undated companion paper says 50–78. Participants chose AI use themselves (“self-selection”), and nonusers were less experienced. Unlike the assigned-preference comparison, these groups were not formed by chance: the differences do not establish AI-caused benefits. Error avoidance and independent judgment are separate targets: a valid analysis can still reflect the adviser’s favored framing.