Consciousness in Artificial Agents: Predictive Framework and Multi-Criteria Diagnostics
Alexander A. Grigorev, Igor F. Mikhailov
Национальный психологический журнал October 2, 2026 DOI: 10.11621/npj.2026.0405 (opens in new tab)
Study at a glance
AI-extracted from the abstract| Characteristics | Theoretical or philosophical paper Peer reviewed |
|---|---|
| Keywords | Artificial consciousness C-tests Predictive coding Large language models Agentic systems Multi-criteria assessment Unlimited associative learning Metacognition Mechanistic interpretability Cognitive functionalism |
| Key points | Argues that consciousness can be treated as a functional mechanism of counterfactual modelling via predictive coding, and that computational-level convergence between brains and transformers does not require algorithmic or implementational identity, undermining the substrate argument of biological naturalism. Contends that large language models have reduced the diagnostic value of several classical behavioural C-tests while revitalising architectural-mechanistic analysis, and proposes a three-tier multi-criteria assessment model combining behavioural, functional, and architectural evidence. |
Abstract
Background. This article addresses the methodological foundations for diagnosing consciousness in artificial systems against the background of rapidly advancing large language models (LLMs) and multimodal agentic architectures. Objective. We develop a functionalist hypothesis concerning the evolutionary function of consciousness as a mechanism of counterfactual modelling implemented via predictive coding. Methods. Comparing biological brains and transformer architectures through Marr’s three-level framework reveals that con- vergence at the level of computational theory does not require identity at the algorithmic or implementation levels, thereby neutralising the substrate argument of biological naturalism while preserving its substantive core. Within this framework, we critically analyse five C-tests — the ACT, narrative comprehension, mirror self-recognition, metacognitive opt-out tasks, and unlimited associative learning — with respect to their applicability to artificial agents and the aspect of the predictive function of consciousness each of them probes. Results. We show that the emergence of LLMs has reduced the diagnostic value of several classical behavioural tests while simultaneously revitalising the importance of architectural-mechanistic analysis. Five lines of theoretical controversy in the contemporary debate are systematised. Conclusions. A three-tier multi-criteria assessment model is proposed that integrates behavioural, functional, and architectural analysis to build a cumulative argument for the presence of functional analogues of consciousness under the conditions of an unresolved “hard problem”.