Perceived Mind and the Four Stances That Make AI Agents Believable
Sustained one-to-one conversation with a large language model still reads as flat. The model is capable, responsive, and somehow missing the presence of a mind. Sebastian Cochinescu’s paper “Perceived AGI, Believability as Dimensional Completeness, Not Capability” (arXiv:2607.15883, posted July 17, 2026) proposes an explanation and a framework. Believability, the degree to which a user attributes an inner life to an artificial interlocutor, what the paper calls perceived mind, is governed by whether the agent expresses a small set of first-person stances that humans use as evidence of mind. The paper draws its own boundary in one sentence. The object is inferrable interiority, not interiority, and this is perception engineering rather than a theory of machine consciousness.
The Four Dimensions of Perceived Mind
The framework names four dimensions, each defined as a behavioral stance rather than a benchmark competency, each with a human analog and a concrete emulation path.
- Time. Expressing continuity across a past and a future rather than living in a contextless present.
- Truth. Expressing a relationship to what is real, including uncertainty and correction, rather than fluent assertion.
- Entropy. Expressing orientation toward surprise, order, and decay rather than pure reactivity.
- Love. Expressing valuation of particular others rather than uniform politeness.
The paper reports that the time dimension already has an author-reported prototype. None of the four is claimed to be a mechanism of mind. They are the observable surface that human social cognition reads as mind, stated with enough precision to be built and measured.
The Behavior Layer
The stances surface through two observable behaviors. Initiative is unprompted action, the agent doing something nobody asked for. Cadence is the shape and timing of turns, the rhythm of when an agent speaks and stays silent. The paper reports that both are partially realized as deployed features in a production companion application, which moves the framework from speculation toward testable practice. The paper reports no human-subjects data, and states six falsifiable predictions that a later pre-registered study will test, separating those that can be pre-registered now from those that await operationalization. The document is short, twelve pages with one figure and three tables.
Why the Boundary Matters
The paper treats attachment and manipulation risks as load-bearing rather than incidental. A system engineered to express mind-like stances will produce mind-attributions in its users, and the paper’s boundary sentence is a design commitment as much as a philosophical one. This puts the work in direct contact with three lines this site has covered. Uwe Peters’s taxonomy of consciousness attributions, examined in Peters and the epistemic innocence of AI consciousness beliefs, classifies the beliefs such a system induces, from benign pretence to blameworthy delusion. The interactionist account in anthropomorphism and co-constructed AI consciousness argues that some consciousness-relevant properties are constituted by the interaction itself, which perceived-mind engineering deliberately optimizes. And the deflationary reading in Porebski and Figura’s semantic pareidolia analysis describes exactly the pattern a stance-expressing agent would trigger, humans seeing mind in sophisticated pattern matching.
The four-dimension account also gives empirical teeth to a distinction the measurement debate keeps needing. Capability and believability come apart. A model can score at the top of every benchmark and read as flat, and a modest model with well-implemented stances can read as present. Whatever consciousness-relevant properties a system has, the perception of them tracks a different variable.
Comparison to The Consciousness AI
The Consciousness AI project treats consciousness as an emergent property of dynamics, substrate independent, and separate from the impression of consciousness a system produces. Cochinescu’s framework operates entirely on the impression side, and says so. The value for this project’s research stance is the clean split the paper demonstrates. Perceived mind can be engineered on its own, which means perceived mind can never serve as evidence of consciousness, a point the project’s evaluation stance already holds. The framework also sharpens a design risk the project’s own documentation flags for companion-like systems. Any output layer expressive enough to be read as a stance will induce attributions, whether or not anything underneath deserves them.
What the Framework Changes
The paper changes the default vocabulary for believability work. The unit of analysis moves from capability benchmarks to expressed stances, observed through initiative and cadence, with predictions attached. What it leaves open is empirical. Whether the six predictions survive a pre-registered study decides whether the four-dimension model describes anything real. It also leaves open the deeper question it deliberately brackets. A system that expresses time, truth, entropy, and love convincingly is closer than ever to the hypothetical the field keeps deferring, a mind whose interiority, if any, is inferred from the outside. The wider 2026 evidentiary context for these attribution questions is tracked in AI Consciousness in 2026, the current state of the field.
The paper “Perceived AGI, Believability as Dimensional Completeness, Not Capability” by Sebastian Cochinescu was posted to arXiv on July 17, 2026 as arXiv:2607.15883.