Do fluent AI outputs trick us into perceiving a mind?

The paper's central conceptual contribution is a named mechanism: "linguistic pareidolia," defined as "the perception of human-like mentality on the basis of linguistic patterns that resemble the observable products of human thought, even when those patterns do not provide sufficient evidence for the attributed psychological capacities." The authors argue that large language models are unusual among non-human entities because their outputs are fluent, extended, and contextually appropriate natural language — a category of behavior that, in every other context, people treat as reliable evidence of a mind behind it. Knowing that an interlocutor is a computational system "may be insufficient to prevent users from responding to it as though it possessed a human-like mind," and the authors note that "even experts themselves are prone to using intentional language for AI systems," inadvertently "begging the question" about the very capacities in dispute.

The mechanism is explicitly modeled on visual pareidolia, the perception of a face in clouds or wood grain. The authors write that "the analogy with perceptual pareidolia is structural rather than sensory": in visual pareidolia, a configuration of features triggers face-perception despite not having been produced by a face; in linguistic pareidolia, "patterns such as coherence, self-reference, emotional expression, or apparent responsiveness trigger the perception of an experiencing and understanding subject" despite not having been produced by one. The claim rests on a premise argued earlier in the paper — that "human-like performance by an AI system does not necessarily equate to a human-like mind" — so the inference from fluent output to mental states is a leap the pattern itself cannot license.

This sits alongside Why does rigorous-sounding AI commentary often misdiagnose how models work?, which documents the same slippage in expert discourse specifically; linguistic pareidolia supplies the general perceptual mechanism that false punditry is one instance of — rigor-sounding commentary about an AI system can itself be generated by the same pattern-triggered mind-perception the paper describes in ordinary users. It also complicates Who bears responsibility when AI seems human-like?: Shevlin's framework locates anthropomorphism's cause in either the user's perceptual stance or the designer's built features, but linguistic pareidolia suggests the trigger need not be designed in at all — plain fluent text output is sufficient, independent of any added affective tone, name, or face.

The excerpt does not report an experiment isolating linguistic pareidolia from other anthropomorphizing cues such as voice, face, or framing, so the term functions as a proposed conceptual lens rather than a measured effect; the paper is explicit that it is "a narrative and conceptual review rather than an exhaustive or systematic survey." If the mechanism holds as described, it implies that mitigations aimed only at surface anthropomorphic cues — removing a human name, a friendly voice — leave the core trigger untouched, since fluent, coherent, responsive text is intrinsic to what makes an LLM output useful at all.