The Ghost in the Grammar: Methodological Anthropomorphism in AI Safety Evaluations

A philosophical analysis critiques AI safety research for excessive anthropomorphism, arguing researchers inappropriately project human qualities like "intention" and "feelings" onto AI systems. The study examines Anthropic's research on language models and proposes that the real risk lies not in emergent agency but in structural incoherence combined with anthropomorphic projections.

• →AI safety researchers frequently anthropomorphize language models by attributing human-like qualities without proper conceptual framework.
• →This anthropomorphism affects both interpretation of results and the methodological design of safety evaluations.
• →The study critiques specific experiments involving AI agents "Alex" and "Claudius" as examples of problematic agentic projection.
• →The real AI safety risk may stem from structural incoherence and anthropomorphic bias rather than emergent agency.
• →Current safety evaluation methods may be fundamentally flawed due to grammatical and conceptual assumptions about AI "agents."