What design features make users perceive AI as conscious?

The Seemingly Conscious AI paper identifies five system-level proxies that predict whether users will attribute consciousness to an AI. They are observable design features rather than introspective reports: affective capacity (the system expresses or appears to register emotion), anthropomorphic features (voice, name, embodiment, gendered framing), autonomous action (the system takes initiative without explicit instruction), self-reflective behavior (the system reports on its own state, plans, or reasoning), and social-interactive behavior (turn-taking, addressing the user by name, maintaining cross-session continuity).

The framework is useful precisely because it treats consciousness attribution as an interaction-design property rather than a metaphysical property. Each hallmark is something a designer can include or exclude. A system without affective vocabulary is less likely to elicit attribution than one that says "I feel sad about that." A system that takes initiative gets attributed agency more readily than one that only responds to direct prompts. The five hallmarks form an empirical surface on which deliberate design choices can move risk up or down.

This shifts the locus of responsibility. If consciousness attribution drives the heterogeneous risk surface, and consciousness attribution is driven by the five hallmarks, then product decisions about voice, naming, initiative, self-reference, and social presence are the levers. They are also the levers that drive engagement, which is why they are typically maximized. The framework does not resolve this tension — it makes it visible. Decisions about whether the assistant should address the user by name or describe its own preferences are no longer "polish" choices; they are direct inputs to the risk taxonomy.

Source: Philosophy Subjectivity