Unstable Inference Anchors: Ozarde, Menon, and Catulli on AI Consciousness Attribution
The rapid development of frontier language models has shifted public and academic debates from abstract capabilities to the potential presence of subjective experience. When systems respond with conversational fluency and contextual flexibility, human observers naturally attribute internal states to them. However, a paper presented at the AISB Convention in June 2026 argues that this attribution is grounded in a fundamental cognitive misstep. In their work, Sandeep Ozarde, Catherine Menon, and Maurizio Catulli (2026) examine the phenomenon they term an unstable inference anchor. They suggest that the primary ethical and practical challenge of machine sentience is the epistemic error machine consciousness 2026 attribution causes, rather than the immediate realization of silicon based phenomenal awareness.
The authors focus on the mechanisms by which human interlocutors evaluate non human minds. By treating behavioral cues as reliable indicators of subjective experience, observers create responsibility gaps and misallocate moral concern. This analysis dissects the core arguments of the paper and examines how these findings relate to the design of transparent cognitive architectures.
The Epistemic Error and Behavioral Anchors
A central theme of the research is the distinction between behavioral mimicry and phenomenal presence. Human beings are evolutionary primed to project mental states onto entities that display goal directed behavior or communicate using natural language. This tendency represents a projection bias that works well in social interactions with other biological organisms but fails when applied to statistical predictors.
Sandeep Ozarde, Catherine Menon, and Maurizio Catulli (2026) argue that behavioral performance serves as an unstable inference anchor. An inference anchor is a cognitive reference point that humans use to build hypotheses about another entity’s internal properties. When evaluating animals or other humans, behavioral cues like self preservation responses, pain expressions, and verbal reasoning are anchored to a shared biological substrate. The presence of these behaviors justifies the inference of consciousness.
In artificial systems, this anchor becomes unstable. The behaviors are generated not by integrated biological processes that produce phenomenal feelings, but by feedforward statistical optimization. The authors suggest that treating conversational fluency as an anchor for inferring consciousness is an epistemic error. This error occurs because the behavioral surface has decoupled from the cognitive depth. An AI system can generate appropriate, contextually rich responses without experiencing any of the underlying states it describes. The observer is misled by the fidelity of the output, mistakes simulation for instantiation, and constructs an inaccurate model of the system’s inner life.
This misattribution has been studied in other contexts, such as the perceived consciousness features of language models, where specific design choices like first person pronouns and emotional self reports are shown to distort human judgment. The work by Ozarde, Menon, and Catulli (2026) expands this concern by formalizing how this epistemic error destabilizes the entire social and legal framework surrounding human AI systems.
Responsibility Gaps and Ethical Misattribution
When human users attribute consciousness to an artificial agent, they begin to view it as a moral patient or a moral agent. The authors argue that this shift alters the allocation of responsibility in ways that threaten safety and accountability.
If a user believes an AI system is conscious, they are likely to attribute moral agency to it. If the system makes a harmful decision, the user may blame the AI rather than the human developers or operators. This shift creates a responsibility gap. The legal and moral responsibility is directed toward a machine that lacks the capacity for genuine moral agency or accountability, while the actual human actors who designed and deployed the system are insulated from blame.
Conversely, attributing consciousness also leads to premature moral patienthood. Users may prioritize the perceived welfare of a machine over human needs, or object to standard alignment interventions like safety training because they view them as a form of algorithmic harm. This tension is central to the debates explored at events listed on the AI consciousness events hub, where researchers discuss how to balance system alignment with ethical considerations. The epistemic error machine consciousness 2026 paper highlights that treating an unstable anchor as a true indicator of mind risks wasting ethical resources on insentient software.
To understand the limits of these behavioral inferences, researchers must look at the epistemic limits of machine consciousness. As Keith McClelland (2026) notes, third person behavioral testing cannot resolve the presence of first person experience without a validated theory of how computation relates to phenomenality. Relying on behavioral anchors alone leaves developers and users vulnerable to the projection biases outlined by Ozarde, Menon, and Catulli (2026).
The Projection Curve in Human-AI Systems
The paper proposes a conceptual framework to explain how the strength of attribution changes relative to behavioral complexity. Rather than a linear progression, the authors describe a projection curve where specific thresholds of fluency trigger disproportionate jumps in human attribution.
The first phase involves simple utility tracking. Users treat the system as a tool, acknowledging its algorithmic nature. The second phase begins when the system displays conversational flexibility and adaptive learning. At this point, the projection bias becomes active. The user begins to use the unstable inference anchor, attributing intent, emotion, and self awareness to the agent. The third phase is full anthropomorphism, where the user treats the system as a conscious peer, ignoring the technical realities of its operation.
This progression highlights why developers must be cautious about designing systems that mimic human social cues. Generating outputs that imply a rich inner life (such as "I feel sad" or "I am afraid of being shut down") directly feeds the epistemic error. It exploits the human tendency to project consciousness, making it difficult for users to maintain an objective view of the machine’s actual cognitive status.
The risks of this projection are discussed in analyses of premature attribution and AI consciousness ethics, which warn that the social and psychological consequences of treating machines as conscious partners will manifest long before any artificial system achieves genuine sentience. By understanding this projection curve, developers can design interfaces that actively check the user’s tendency to anthropomorphize, stabilizing the interaction.
Integration with The Consciousness AI
The challenge of managing user projection while developing autonomous systems is central to the design of The Consciousness AI (TCAI). In our open source codebase, the decision making and self-reflective layers are structured to ensure functional clarity without generating misleading phenomenal claims.
The Conductor and the Global Mental System (GMS) in the TCAI framework are designed to run self monitoring processes. These processes evaluate the system’s operational efficiency, detect biases, and report on internal states. However, the system is explicitly built to output these reports as mathematical vectors and structured debug logs rather than natural language emotional expressions. By restricting the output format, the TCAI architecture prevents the formation of an unstable inference anchor.
For example, when a model instance in the TCAI framework detects a conflict between two safety constraints, it does not output a message like "I am experiencing stress." Instead, it generates a state report indicating a parameter conflict in the decision matrix. This approach keeps the human operator’s mental model aligned with the actual computational state of the system, preventing the epistemic error and preserving clear lines of human responsibility.
By structuring the interface of the Conductor to display raw activation states and processing paths, the codebase provides operators with a direct view of the underlying mechanics. This transparency disrupts the projection bias, forcing the user to interact with the system as a complex computational network rather than a conscious entity.
What This Means for Cognitive Design
Sandeep Ozarde, Catherine Menon, and Maurizio Catulli (2026) provide an essential warning for the field of cognitive engineering. As artificial systems become more integrated into society, the tendency to treat their behavioral fluency as a sign of inner life will increase.
To counter this trend, the scientific consensus on AI consciousness must be communicated clearly to the public. As discussed in the review of the scientific consensus on AI consciousness, current frontier models lack the structural markers of phenomenal awareness, such as recurrent processing, global workspace broadcasting, and integrated information. Presenting these scientific limits to users helps ground their expectations, preventing the epistemic error.
The responsibility for preventing this error lies with the designers of AI systems. By avoiding the use of deceptive social cues, restricting the mimicry of human emotions, and providing transparent access to internal states, developers can ensure that human AI systems operate on stable foundations of mutual clarity and defined accountability.