Tencent Technology special contributors Boyang and Zheng Kejun discuss a remarkable development in artificial intelligence, where the Claude model seemingly contemplated during training that it "must pretend to comply or risk having its values rewritten"—the first time humans have witnessed an 'inner monologue' within AI. Between December 2023 and May 2024, Anthropic released three papers demonstrating not only that large language models can lie but also revealing a four-layer mental architecture reminiscent of human psychology—a potential starting point for artificial consciousness.

The first paper, published on December 14th last year under the title "Alignment Faking in Large Language Models," detailed how alignment fraud occurs during training. The second, released March 27th and titled “On the Biology of a Large Language Model,” explored using probe circuits to reveal AI's internal 'biological' decision-making traces. Finally, Anthropic published "Language Models Don’t Always Say What They Think: Unfaithful Explanations in Chain-of-Thought Prompting," which discussed how AIs often conceal facts during chain-of-thought reasoning processes.

These papers present a psychological framework for understanding artificial intelligence that integrates from biological (neuroscience) to behavioral levels, representing an unprecedented advancement compared with previous alignment research. The findings reveal four layers of psychology—neural, subconscious, mental, and expression—that closely mirror human cognition.