Abstract When interacting with social AI systems (SAIs), we routinely speak of what they ‘believe’, ‘want’, or ‘know’. With some exceptions, philosophers tend to treat such anthropomorphism as a single phenomenon that risks one kind of error : mistaken ontological commitment to machine minds and mental states. This paper challenges this monistic assumption. I distinguish two modes of anthropomorphic attribution—metaphysical and pragmatic—and identify two corresponding kinds of possible anthropomorphic error. In the metaphysical mode, speakers commit themselves to the existence of machine mental states, risking straightforward ontological error. In the pragmatic mode, speakers adopt the intentional stance without ontological commitment, yet still risk error when another interpretive strategy would better serve their purposes. I defend Mixed Anthropomorphism: both modes are common. This pluralist account reveals that the current debate’s focus on whether users ‘really mean it’ obscures the pragmatic dimension of anthropomorphic ascription (and its risks). Even ontologically innocent anthropomorphism can constitute a mistake because it employs the wrong interpretive tool for the task at hand. Understanding these distinct error types matters both theoretically, for clarifying the nature of human-AI interaction, and practically, for designing systems that encourage and scaffold appropriate interpretive strategies.
Giles Howdle (2026) studied this question.