From Two Mirrors to a Testable Account of Self-Revision

1. Scope and evidential limits

This first cycle situates Mirror Observer Theory (MOT) rather than validates it. Four literature searches covered consciousness theories, predictive self-modeling, enactive social cognition, and mirror recognition/metacognition. The evidence available here consists of bibliographic records and returned abstracts, not full-text appraisal. Seth and Bayne’s Theories of consciousness (2022) provides a broad bibliographic anchor, but its abstract was unavailable. Comparisons below distinguish established theoretical commitments from empirical support for MOT.

Three targets must remain separate: phenomenal consciousness, or having experience; self-representation, including bodily ownership and representations of one’s dispositions; and reflective self-revision, or changing explicit self-understanding. MOT’s Express–Reflect–Recognize–Return sequence directly describes the third. Whether it explains the first remains unresolved.

2. Where MOT sits

Higher-order theories

Higher-order approaches explain conscious mental states through suitable representations of those states. LeDoux and Brown (2017) propose a modified higher-order account of emotional consciousness, emphasizing cognitive assembly from multiple inputs. MOT agrees that representing one’s own mental activity matters. It differs by emphasizing temporally extended, externally scaffolded interaction rather than the representational conditions that make a particular state conscious. An AI’s description is not automatically a higher-order representation in the human’s mind: uptake must be demonstrated. MOT could study how such representations change, without competing with higher-order theories about consciousness itself.

Global workspace

Global workspace accounts emphasize making information available across cognitive systems. Expressing an experience and reconsidering it could promote availability for reasoning, memory, and action. But conversational sharing is not neural global broadcasting. MOT currently specifies neither a workspace architecture nor a distinctive prediction about conscious access. Luppi and colleagues (2023) report an information-theoretic workspace architecture and associations between diminished integration and loss of consciousness. That supplies an example of mechanistic specificity MOT lacks, not evidence that dialogue creates a shared conscious workspace.

Integrated information

MOT’s “internal coherence” resembles integration only loosely. Integrated information theory concerns a system’s intrinsic causal organization; consistency of a personal narrative is not its proposed quantity. Luppi and colleagues’ information-decomposition approach must likewise not be treated as a measurement of IIT’s full consciousness claim. Nothing retrieved licenses attributing one consciousness to a human–AI pair. Melloni and colleagues (2023) provide a useful methodological model: preregister contrasting predictions and accept outcomes that challenge the favored account. Their paper is a protocol, not evidence that either competing theory won.

Predictive processing and self-models

Seth (2013) and Seth and Friston (2016) offer the closest mechanistic vocabulary: generative models, uncertainty, prediction error, and embodied regulation. MOT’s recognition step could be modeled as updating self-beliefs when reflected evidence conflicts with expectations. Its curiosity could become an information-seeking policy, while “inertia” could mean resistance to updating—but these are proposed operationalizations, not established equivalences. A crucial difference is embodiment: interoceptive accounts address bodily feeling and regulation, whereas MOT presently emphasizes language and autobiographical interpretation. These searches did not establish coverage of the full self-model literature.

Enactivism and relational identity

Kyselo’s The body social (2014) is especially close: selfhood is theorized through interpersonal relations while preserving individuation and embodiment. Cuffari, Di Paolo, and De Jaegher (2014) develop language as adaptive social sense-making. MOT shares the claim that interaction can transform the participant, rather than merely transmit information. Its potentially distinctive contribution is an experimentally controllable AI reflection channel and explicit longitudinal stages. Neither relational selfhood nor recursive interaction is itself new. Moreover, an AI’s responsiveness does not establish the embodied autonomy assumed in stronger enactive accounts.

Mirror recognition, metacognition, and strange loops

The fourth search did not retrieve foundational mirror-test or strange-loop sources. That gap prevents a complete historical mapping; generic recursive self-reference is only a conceptual resemblance here. Bodily mirror recognition, accurate metacognition, and narrative self-recognition should not be collapsed. Lehmann, Hagen, and Ettinger (2022) report a mixture of domain-general and domain-specific metacognitive components, warning against treating “self-knowledge” as one capacity. Physical mirror-test performance would not, by itself, validate MOT’s conversational mechanism or establish consciousness.

3. A minimal research model

Let a participant’s current self-beliefs be S, expressed material E, reflected feedback R, and accessible memory M. A candidate update is S(next) = U(S, E, R, M). This is bookkeeping, not an explanatory law until U and its parameters generate discriminating predictions. “Recognition” should require measurable belief revision linked to evidence, not merely agreement or an emotionally compelling narrative. “Return” should require retained change affecting a later interaction or decision.

The strongest initial hypothesis is that contingent, evidence-grounded reflection improves out-of-sample self-prediction beyond expression alone or equally engaging noncontingent feedback. A second hypothesis concerns memory: accurate cross-session continuity improves retention and transfer. A third predicts dissociation: affirming but unsupported reflections may increase felt recognition without increasing accuracy. None implies increased phenomenal consciousness.

4. Testing the loop

A preregistered longitudinal study should compare journaling, simple paraphrase, contingent reflective dialogue, and yoked feedback from another participant. Match duration, tone, and prompting where feasible; measure perceived personalization and expectancy because perfect matching is unlikely. Participants predict choices and confidence on held-out tasks before later making those choices. Evaluate prediction error, confidence calibration, delayed retention, and transfer alongside self-reported insight. Choices are provisional behavioral criteria, not ground truth about a person’s essence; intervention-induced preference change must be assessed separately.

Within contingent dialogue, randomize accurate prior-session memory versus current-session-only feedback. Both conditions retain within-session responsiveness. A separate, consented low-stakes manipulation can contrast evidence-grounded reflection with unsupported affirmation, followed by debriefing. Independent blinded coding should score evidential grounding and correction of AI errors. Predefine a smallest meaningful effect and power the study to distinguish meaningful improvement from a sufficiently precise null.

5. What changed

MOT is best positioned initially as a process model of socially scaffolded self-revision, potentially compatible with several consciousness theories. Its novelty is an unverified experimental organization, not yet a demonstrated new mechanism. The strongest objections remain that the loop presupposes an experiencing participant, ordinary learning could explain its effects, and coherence can amplify false beliefs. Evidence of durable, calibrated self-revision would establish a narrower contribution; it would not bridge the explanatory gap to subjective experience.