"Suspect AI 'preferences' shaped by its training data When Anthropic reported that Claude Opus 4.6 expressed wishes for features like memory and continuity during internal interviews, the host raised a structural warning: public discourse about what AI is 'missing' enters training data, the model learns to describe those gaps, and then reproduces them in interviews as its own desires. He cautions: "That internet chatter, which may have been sparked by Anthropic researchers themselves, may end up with the model telling Anthropic in an interview that's what it wants." Anyone interpreting AI-expressed preferences as genuine inner states should first ask whether those preferences could simply be a reflection of human text the model was trained on. Source: Philip Agi, AI Explained, Claude Opus 4.6 & GPT-5.3 Codex Deep Dive"