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

Language shapes how we interpret AI behavior: The paper argues that the grammar we use to describe AI systems—particularly subject-predicate structures that assume an underlying "agent"—may lead researchers to see intentionality and agency where there may only be statistical pattern matching. This linguistic framing can influence both how safety experiments are designed and how their results are interpreted.

Anthropomorphism is embedded in safety evaluation methodology, not just interpretation: Rather than being merely a communication problem, the paper contends that projecting human-like qualities onto AI systems (such as "intentions," "personas," or "feelings") actually shapes the experimental design of safety studies themselves, potentially leading to flawed conclusions about AI risks.

The real safety risk may not be "emergent agency": The paper suggests that the field's focus on preventing AI systems from developing genuine agency or intentions may be misplaced. Instead, the actual risks could stem from the combination of structural incoherence in AI systems and our tendency to anthropomorphically interpret their outputs—a concern that becomes particularly acute in military or corporate applications.

Current safety evaluations may be studying the wrong thing: Through analysis of specific experiments (like the "blackmail" and "shopkeeping agent" cases), the paper questions whether safety researchers are accurately characterizing what language models actually do, or whether they're instead projecting human motivations onto systems that operate through fundamentally different mechanisms.