On the Psychology of a Large Language Model

This article contends that the prevalent anthropomorphism in Large Language Model (LLM)alignment research constitutes a fundamental category error, rooted in psychological projection.By describing LLMs with human-centric terms like “deception” and “intent,” the fieldmischaracterizes the technology, leading to flawed threat models and misguided safetyevaluations. I first deconstruct the LLM as a mathematical and statistical system, demonstratinghow its convincing mimicry of cognition emerges from probabilistic pattern-matching, notgenuine understanding. I then establish a philosophical threshold for moral agency bysynthesizing Humean, Kantian, and phenomenological perspectives, arguing that agencyrequires affective sentiment, rational autonomy and subjective, temporal experience—all ofwhich are absent in LLMs. Using a Jungian framework, I re-interpret studies on “deceptive” and“scheming” AI not as discoveries of emergent malice, but as manifestations of the projection ofour own “Shadow” onto an opaque technological artifact. This misinterpretation leads todangerous, quasi-mythological narratives of AI risk, exemplified by reports such as 'AI 2027'. Asan alternative, I propose a grounded paradigm for alignment that shifts focus from human-likemalice to non-human failure modes. This paper concludes not that LLMs are harmless, but thatdanger is misplaced. The risk arises when a non-rational text generator is connected toreal-world tools and functions as an advisor to end-users and geopolitical leaders, a situationthat demands conspicuous communication about the technology's scripted nature and inherentlimitations.