AIAI “personality” profiles and the critique of Claude and GPT’s training
A thread cites tests of nine AI models to argue that Claude and GPT’s prosocial profiles reflect homogenized safety training—not moral superiority.
TLDR
A thread describes a September 2026 study that tested nine language models five times each using seven human psychometric instruments in English and Chinese. It says the two Claude models had the closest profiles, while GPT-5.4 had the highest test-retest reliability. According to the thread, the researchers interpret the shared patterns as effects of instruction tuning and safety training, not human personality. The thread’s author goes further, arguing that Reddit-derived training data and safety tuning produce standardized, agreeable models that are less likely to push back against users.
