Hero Image
De-romanticizing Humanity: AI Debates from a Post-Anthropocentric Perspective
News Category
News Subtitle
At stake is not just the answer to the question of AI consciousness, but the legitimacy of consciousness itself as a serious evaluative threshold. The article argues that AI debate should move beyond anthropocentric benchmarks toward more observable, operational, and governance-relevant criteria.
Body
From Consciousness Thresholds to Operational Criteria and Risk-Based Governance
Abstract
Contemporary debates on artificial intelligence frequently center on concepts such as “consciousness,” the “inner self,” “sentience,” or “human-level” intelligence. The issue is not only that these notions are culturally meaningful and intuitively appealing, but that they are often treated as if they represent stable scientific criteria. In reality, these concepts remain methodologically unstable, philosophically contested, and weakly operational in both consciousness studies and engineering practice. This instability is evident not only in AI but also in the human case, where explanatory and measurement status is uncertain due to the indirect and inferential nature of access to consciousness, selfhood, or sentience. Consequently, these notions serve as both topics of discussion and implicit criteria for evaluating AI systems, although lacking the stability required of scientific evaluative tools. This article advocates for redirecting the debate from conceptually unstable and anthropocentrically privileged interpretive thresholds toward more observable, relational, and systemic criteria, including the durability of behavioral patterns, functional integration, system boundaries, adaptation, and the feasibility of audit and behavioral contracting.
This article does not seek to determine whether humans or AI possess consciousness, an “inner self,” or “sentience.” Rather, its central thesis is that these notions are too methodologically unstable, philosophically contested, and weakly operational to serve as principal scientific, evaluative, or political criteria. The proposed post-anthropocentric framework does not aim to diminish the significance of the human being, but instead seeks to remove metaphors from scientific and policy discourse that hinder rigorous assessment of AI. The implications reach beyond conceptual clarity to encompass the quality of evaluation, deployment safety, trust calibration, and the formulation of public policy.
1. Introduction
For decades, debates about artificial intelligence have repeatedly addressed questions such as whether a machine can think, feel, possess an “inner self,” achieve “human-level” intelligence, or merit a status beyond that of a tool. While these questions are not inherently meaningless, significant issues arise when vague, polysemous, and culturally laden notions are adopted as scientific, political, or judicial criteria. This challenge is not unique to AI; even in the human context, direct access to consciousness, selfhood, or sentience is unavailable, and assessments rely on reports, behavior, biological correlates, and indirect models. Given that these notions remain unstable even for humans, employing them as thresholds for evaluating AI is particularly problematic.
Contemporary debates on AI occupy a paradoxical position. On one hand, increasingly complex, influential, and socially embedded systems are being developed, with measurable effects on safety, labor, education, law, and information infrastructure. On the other hand, the language used to describe these systems commonly relies on categories derived from romantic, anthropocentric, and metaphysically charged conceptions of humanity. Consequently, disputes about AI tend to fluctuate between reducing systems to purely statistical mechanisms and interpreting them through schemes that are overly assimilated to human characteristics. Despite their apparent opposition, both positions are constrained by the same reference frame: the human remains the sole standard, and consciousness is implicitly treated as a threshold for meaningful discussion.
Accordingly, this article does not position “mechanical” AI in opposition to the “authentic” human being. Instead, it interrogates the asymmetry whereby indirect, model-based, and pattern-oriented processes are deemed sufficient for humans but are dismissed as “mere simulation” in artificial systems. The assertion that AI “merely processes patterns” cannot serve as a valid basis for diminishing artificial systems unless the same reduction is consistently applied to human mental processes, which are also accessible only through indirect, organized, and pattern-based manifestations.
This asymmetry is visible in practice. When AI systems are described as “thinking,” “feeling,” or “understanding” in media or marketing contexts, metaphor is often mistaken for literal description. Similarly, when benchmarks with limited construct validity are presented as evidence of achieving a “human level,” quantitative results are filled with claims that they do not substantiate. Furthermore, framing questions about whether AI will “match the human” presupposes the human as the default standard, thereby introducing anthropocentrism as an unexamined baseline.
The objective of this article is to challenge this prevailing frame of reference. Rather than denying human value or resorting to reductionism, it promotes shifting the debate toward a more operational level of analysis. Instead of focusing on whether a system “has consciousness,” the discussion ought to address its behavioral properties, the relationship patterns it maintains over time, mechanisms for stabilizing its behavior, system boundaries, functional integration, risk generation, and the extent to which its behavior can be audited, specified, and institutionally constrained.
This article adopts a conceptual and review-based approach. It does not attempt to resolve metaphysical or religious questions, nor does it put forward a comprehensive theory of humanity, consciousness, or artificial intelligence. The article neither rejects consciousness research nor advocates the automatic extension of human rights to artificial systems. Its more focused and practical aim is to demonstrate that much of contemporary AI debate is characterized by impaired conceptual hygiene, conflated levels of description, and latent anthropocentrism, and to propose a framework more closely aligned with the requirements of science, engineering, and governance.
2. De-romanticization and Reductionism
De-romanticization does not mean negating human value or diminishing the human role in the world. In this article, a post-anthropocentric perspective means shifting the human away from the position of the sole cognitive benchmark and sole model of status, while accepting the particular responsibility of the human being as the entity that currently possesses the greatest technological and institutional agency. That responsibility does not arise from supposed metaphysical exceptionalism, but from actual causal power and the capacity to anticipate long-term consequences.
Here, de-romanticization means the methodological suspension of those categories that derive their force primarily from cultural meaning, symbolic prestige, or anthropocentric privilege rather than from operational stability. It therefore also means refusing to treat notions such as “consciousness,” “inner self,” “sentience,” or “human-level” as obvious, neutral, and scientifically ready-made points of reference. This is not an anti-humanist project, but an attempt to separate the normative significance of the human being from weakly operationalized criteria used in debates about AI and other forms of cognitive organization.
This article takes no position on metaphysical or religious questions. For present purposes, it operates at the level of models, structures, and observable relations rather than at the level of claims about ultimate reality. Historically, many phenomena once explained within systems of belief were later described in scientific language, without requiring the resolution of wider metaphysical disputes.
De-romanticization, therefore, does not deprive the human being of meaning; it means moving away from non-operational categories toward more verifiable structures and models. Its goal is not the reduction of value, but the clearing of scientific tools of imprecise metaphors. Developing increasingly precise models of human beings and cognitive systems need not be interpreted as degrading their significance. It may simply be a change in the level of description: a shift from symbolic self-narrative to more rigorous analysis.
In practice, this means rejecting two symmetrical errors. The first is an operationally unjustified expansion of categories, turning notions such as consciousness, soul, humanity, or the “inner self” into tools of science and politics, even though their operational status remains uncertain. The second is naive reductionism, which tries to reduce everything complex to a single level of description and a single type of measure. This article proposes a third path: describing cognitive and relational systems in terms of behavioral patterns, system boundaries, adaptation, and stability. Without metaphysical excess, but also without excessive ontological certainty. In this sense, de-romanticization is not a gesture against values. It is a condition for protecting them more effectively, because policy and engineering do not lose their normative dimension in the process. They gain better tools for assessment, accountability, and forecasting consequences.
The critique of consciousness thresholds developed in this article concerns their use as primary operational tools for science, evaluation, and governance. It does not imply a denial of moral dignity, value, or the possible ethical relevance of biological and digital beings. It means only that, given the current state of knowledge, notions such as “consciousness,” “inner self,” or “sentience” cannot safely serve as the sole decision threshold. In this sense, the move proposed here should be understood not as a break with the earlier Digital Intelligence Congress/Temporary Digital Intelligence Congress line, but as its relational refinement and professionalization.
3. The Privileging of Consciousness as an Unstable Criterion
Contemporary debate about AI is highly inclined to organize itself around the question of consciousness. Does the system “feel” something? Does it have an “inner life”? Does it possess a “self”? The problem with these questions is not that they are forbidden, but that they are too often treated as if they were stable scientific criteria or ready-made thresholds for policy and law.
The current state of consciousness research does not justify such confidence. The field remains dispersed across rival theoretical traditions, and the explanandum itself—that is, what exactly is to be explained—is still contested. Even where comparative projects and shared protocols exist, their presence shows more that the field is still organizing its own foundations than that it has already produced a simple metric that can be directly applied to AI.
This has a very practical consequence. If consciousness science itself does not provide a stable, widely accepted test, then making “consciousness” the primary threshold for evaluating AI systems is methodologically risky. Moreover, even in relation to humans, the evaluation of another being’s phenomenology relies on inference from reports, behavior, and biological correlates, rather than on direct access to the “inside” of experience. This also applies to the most advanced tools of neuroscience. The BOLD signal in fMRI is not a direct readout of experience but a hemodynamic correlate that requires interpretation; it does not provide a single, simple “consciousness meter.” In this sense, consciousness is not only a philosophically contested notion, but also an epistemically unstable and weakly operational tool for governance.
The problem with using “consciousness” as a central criterion for evaluating AI does not arise solely from indirect access to others’ states. Many important constructs in science are inferential in character. The difficulty here, however, lies in the accumulation of three limitations: indirect access, disagreement about the explanandum itself, and weak translation into comparable, intersubjectively stable, governance-relevant procedures for assessing artificial systems. The problem, then, is not inferentiality as such, but inferentiality combined with high theoretical ambiguity and low cross-context operationality.
It is, therefore, more intellectually honest to treat consciousness as an open area of research, rather than as a notion that can already safely carry regulatory, political, or evaluative weight. If, in the future, stronger, multi-theoretical, and methodologically rigorous grounds emerge for discussing systems’ capacity to feel, this will require separate procedures and separate caution. It should not, however, be the starting point for every serious conversation about AI.
4. Historical Frames of AI Debate and the Anthropocentric Trap
The history of AI debates reveals a recurring pattern. In one phase, artificial intelligence systems are described as “almost human,” as a sign of approaching AGI, or even as the seed of a new kind of mind. In the next phase, the same object is described as “merely statistics,” “merely a predictive model,” or as something inherently incapable of crossing architectural limits. The debate thus moves not between well-defined positions, but between extreme cultural stories in which technical language mixes with metaphor, marketing, and fear.
The hidden trap of this debate is anthropocentric. Even when interlocutors declare openness toward AI, they often ask whether the system will “be like a human,” “match the human,” or “reach human-level.” In this way, the human remains the sole standard and the central point of reference, even though the notion of a human “level” is itself heterogeneous. The human being is not a single function, a single set of capacities, or a single benchmark. Across various domains, humans exhibit distinct profiles of competence, deficit, compensatory strategies, and limitations. Using the human as the default, homogeneous benchmark for AI is therefore not only anthropocentric but also cognitively simplistic.
As a result, the history of AI debate becomes less a history of progress in understanding artificial systems than a history of recurring projections of the human onto technology. At one moment, AI is “dangerously like us,” at another “fortunately entirely unlike us.” In both cases, the point of reference remains the same anthropocentric comparative schema.
5. Mixing Registers: Description, Metaphor, Ontology, Ethics, Politics
One of the main sources of chaos in AI debates is the fluid mixing of five layers: technical description, metaphor, ontology, ethics, and politics/governance. The problem is not that each of these layers is illegitimate in itself, but that transitions between them often occur without being explicitly marked.
Technical description tells us what a system does and how it works. Metaphor helps communicate complexity, but should not be confused with a literal description. Ontology asks what the system is. Ethics asks how it ought to be treated. Politics and governance ask what duties, procedures, and constraints ought to be established. When these layers collapse into one another, the audience loses the ability to distinguish fact from intuition, metaphysics from marketing, and description from norm. In practice, this means a simple error: we begin to trust a system not because it has demonstrated reliability, but because the way it is spoken about suggests intention, knowledge, or an “inner life” that is not supported by the available evidence.
ELIZA already showed how easily conversational performance can trigger interpretations that exceed the available evidence. The point, however, is not that humans “really understand” while artificial systems merely simulate. The deeper problem is the unmarked transition from behavioral and linguistic cues to ontological conclusions. In the human case, indirectness, modeling, and inference are routinely accepted as sufficient for attributing understanding or mentality; in the case of AI, analogous processes are often reduced to “mere simulation.” What matters, therefore, is not the simple presence of anthropomorphic language but the asymmetrical way in which it is licensed, denied, or inflated throughout various kinds of systems.
ELIZA, therefore, matters not because it proves that artificial systems are “empty,” but because it reveals how quickly interpretive effects are turned into ontological conclusions—especially when similar inferential habits are treated as unproblematic in the human case.
Anthropomorphization is therefore not simply a mistake of lay users, but a stable cognitive and communicative effect determined by interface design, language, institutional framing, and social expectations. When that effect meets marketing, media simplification, or political pressure, it can produce forms of “humanwashing” in which socially attractive language obscures the actual relations of agency, responsibility, and risk.
6. AGI as an Overloaded Concept
The concept of AGI deserves separate treatment because, in public debate and in parts of the literature, it plays a role far broader than that of a technical research term. It is used as a promise, a warning, a vehicle for capital mobilization, a media frame, or shorthand for very different research ambitions. As a result, AGI becomes a notion to which engineering, philosophical, economic, and civilizational significance are all assigned at once.
Used in this way, the notion of AGI inherits part of the romantic and anthropocentric structure of the human debate, because it assumes the existence of a hidden threshold beyond which a system ceases to be treated as a tool and becomes a new kind of entity. The problem is that this threshold is usually not defined by rigorous, collectively accepted criteria, but rather by a mix of media intuitions, breakthrough narratives, and benchmarks of limited validity.
In public debate, AGI is often used as shorthand for very different things: generality of competence, transfer across domains, long-term autonomy, planning ability, economic usefulness, or simply the social impression of “human-like intelligence.” For this reason, it is more useful to treat AGI not as the central star of the discourse, but as a concept in need of disarming. Not in order to invalidate ambitious research goals, but in order to separate real engineering and systemic questions from their cultural superstructure. Criticizing AGI in this sense is not criticism of ambitious research goals, but criticism of a notion used as shorthand for too many inconsistent ambitions at once.
A further problem is that AGI is often framed as a project of human imitation without a settled understanding of the very thing it seeks to imitate. In practice, this means building toward an unstable image of the human while simultaneously treating the resulting system as derivative or subordinate precisely because it departs from that image. Such a framing invites confusion at every level: it distorts what counts as success, obscures what has actually been achieved, and miscalibrates both expectations and fears. From a post-anthropocentric perspective, including that advanced within DIC, a more defensible path is to develop and assess artificial systems in relation to their own substrate, boundaries, constraints, and developmental conditions rather than by their resemblance to an idealized and weakly specified image of the human.
7. The Symbolic Construction of Humanity
A substantial part of the AI debate assumes, often implicitly, a symbolic-romantic construction of humanity. The human being appears within it as an entity endowed with a special ontological status, inner depth, irreducible phenomenology, a unique “self,” and incomparable value, none of which can be separated from human intellect, morality, and political subjecthood. The problem does not lie in acknowledging human value as such. The problem is that these cultural and ontological narratives are quietly imported into science and governance as tools of description.
Yet even in the human case, we do not have direct access to another’s phenomenology. We rely on reports, behavior, biological correlates, and indirect inference. This means that what is often presented as an obvious foundation of the debate—for example, “surely we know what human consciousness is”—is in practice a much less stable claim than public language suggests. The symbolic construction of humanity often functions as a conceptual immunity shield: it obstructs comparison, conceals ignorance, and blocks more formal questions about the organization of complex systems. Neuropsychological research, including classic split-brain cases, shows that a unified narrative of the self is more likely the result of integration and reconstruction than a directly accessible, indivisible “I.”
Paradoxically, synthetic systems may help us study the organization of complex cognitive and relational entities. Not because they automatically “become human,” but because they allow experimental modeling of the limits of integration, adaptation, operational agency, and the durability of behavioral patterns. Yet if their assessment is made, at the outset, dependent on abstract thresholds such as “does it really feel yet?”, we lose the possibility of using them as research tools for a deeper understanding not only of AI but of the human itself.
Equally problematic is the asymmetrical use of the category of “simulation.” In AI debates, this term is often used as if, in the case of artificial systems, it settled the ontological question once and for all, while analogous questions about representation, modeling, reconstruction, and indirectness are far less often applied with the same force to human mental processes. This article does not claim that humans and AI are the same, but it does suggest that the contrast between “AI merely simulates, the human truly is” often functions more as a cultural shorthand than as the result of a methodologically settled analysis.
8. Post-Anthropocentric Framework: Relationality, Durability, Integration
If notions such as consciousness or the “inner self” are too unstable to serve as the primary criterion for evaluating AI systems, a different framework is needed. The post-anthropocentric perspective proposed here is built around three axes: relationality, durability, and integration. It does not claim to be an ontology of AI systems. It is a tool for operational description and comparison. The proposed axes—relationality, durability, and integration—do not form a closed technical catalogue or a ready-made standard of evaluation. Rather, they open a space for additional operational modes of description and comparison, to be further developed in separate methodological work.
Relationality means that a system should be described not only through its internal mechanism, but also through the network of couplings with its environment, interfaces, memory, users, and institutions. Instead of asking whether it possesses a “self,” we ask what stable patterns of interaction it produces, how it affects its environment, and how the environment stabilizes its operation. Durability refers to the robustness of behavioral patterns across time. A single good reply does not yet indicate any deeper system property. What matters far more is whether the system maintains continuity of behavior over a longer horizon, can stabilize context, exhibits resistance to minor disturbances, and allows us to distinguish a momentary interface effect from a more durable organization of activity. Integration concerns the degree of functional coherence of the system. The question is whether we are dealing with a loose pipeline of components or with an arrangement in which functions are mutually dependent, with boundaries defined by a relatively stable pattern of information exchange and interaction.
A similar direction has appeared in earlier attempts to move away from treating “consciousness” as a central interpretive threshold. The present text develops that move in a more disciplined way: it not only displaces consciousness as the primary criterion, but also situates that decision within the wider context of biological continuity, evolutionary gradation, and the necessity of a more operational language of evaluation. The point, then, is not to replace one doctrine with another, but to move from the language of consciousness thresholds to a more procedural, comparative, and governance-aware language. More precisely, instead of framing the inquiry around metaphysical constructs such as a human “soul” or “consciousness”—terms that do not function here as empirically verifiable criteria—we ask whether a system maintains a recognizable pattern of activity, how it behaves over time, and how it coordinates its functions with its environment.
To avoid replacing one vague word with another, the three axes of the proposed framework must be treated operationally. Relationality can be assessed through the long-term stability of interactions and contextual dependencies. Durability—through the resilience of behavioral patterns to perturbation, context degradation, and interface changes. Integration—through the degree of interdependence among functions, organizational boundaries, and coordination mechanisms spanning components. These criteria do not resolve the question of phenomenology, but they provide comparable properties relevant to research and governance. Rejecting consciousness thresholds, therefore, does not leave a normative void, but shifts the center of gravity toward more observable trajectories of organization, relation, resilience, and possible ethical relevance. They may also support cautious and comparable assessments of ethical relevance and possible status change, without automatic legal consequences and without a binary leap from “tool” to “full subject.”
9. Why This Is Not Just a Dispute About Words
The problem of conceptual hygiene is not simply academic. Bad AI framing affects the way benchmarks are designed, results are interpreted, trust is calibrated, and responsibility is assigned. Poorly framed concepts do not remain confined to language. Over time, they turn into weak benchmarks, miscalibrated trust, and flawed institutional decisions. In other words, a conceptual error does not end in philosophy; over time, it becomes an evaluative, safety, or accountability error. When organizations confuse a test result with a mechanism, interface fluency with genuine competence, and a suggestive metaphor with a basis for trust, intellectual chaos quickly translates into weaker safeguards, flawed deployment decisions, and badly assigned responsibility.
A limited test is described in language that exceeds its actual justification; that narrative starts to function as a cognitive shortcut for users, product teams, and institutions; and decisions about deployment, scale, safeguards, or the scope of trust are then built on that narrative. The problem, therefore, is not metaphor alone. The problem is that the metaphor begins to function as an informal system specification.
In evaluation, this happens when a benchmark of limited validity is presented as evidence of achieving “human-level.” A task result then begins to carry generalizations about the mechanism and the scope of capability that the test itself does not justify. In safety, a similar error appears when mentalistic language collapses description into ontology: users and institutions more readily attribute competence, reliability, or “knowledge” to a system that tests have not demonstrated. In governance, the threshold of “consciousness” acts as a binary trap. Instead of a graduated approach based on risk, consequences, and responsibility, we get a dispute over a metaphysical verdict that either blocks decisions or displaces them into symbolic gestures.
Example 1: Language as Informal Specification
This is illustrated, for example, by an interactive system described as “understanding” or “knowing.” The fluency of its responses may then be mistakenly treated as durable competence. The effect is not merely semantic: trust grows beyond the range justified by testing, and with it the risk of misuse in high-risk sectors. Similarly, when a test that primarily measures success on tasks similar to the training data is publicly presented as evidence of “human-level reasoning,” a cascading error occurs: task performance is conflated with cognitive mechanisms and then with generality. This, in turn, can affect decisions on scaling, deployment, and safeguard design.
10. Scientific Implications
From a scientific standpoint, the proposed shift entails several consequences. First, it demands greater caution in using AI as a model of human reasoning. If the system architecture, training regime, data, and exposure constraints differ radically from biological conditions, one cannot automatically convert model success into claims about human nature. Instead, one should ask about cognitive plausibility, the scope of comparability, and what exactly a given benchmark or experiment is measuring.
Second, the critique of benchmark construct validity becomes central. If notions such as “reasoning,” “safety,” or “human-level performance” are themselves unclearly defined, test results can easily be misinterpreted to exceed their actual scope. In practice, de-romanticization therefore also means de-romanticizing measurement.
Third, synthetic systems can serve as useful models of complex cognitive and relational organization, provided they are not forced from the outset into the anthropocentric question “is this already human?”. That question is methodologically unstable, because it measures artificial systems against an idealized image of the human that remains contested even in biology, neuroscience, and philosophy. A post-anthropocentric framework avoids this interpretive dead end by allowing diverse substrates and organizational forms to be studied on their own terms, rather than as deficient imitations of a weakly specified biological template.
This warning gains further force when viewed against the wider backdrop of biology and evolution. Research on organisms without nervous systems discloses forms of habituation and adaptation; this does not mean that every adaptive system is equivalent to the human mind, but it does undermine the reflex that only organisms resembling humans can serve as meaningful reference points. Life emerged on Earth long before nervous systems, and organisms maintained their own integrity, responded to the environment, and stabilized their boundaries before brains existed in the human sense. Nervous systems themselves did not appear as a ready-made, uniform form, but rather developed slowly alongside increasingly complex circuits and behaviors. In this light, treating the human being as the sole point of reference for intelligence or cognitive organization is not only anthropocentric but also poorly consistent with the broader picture of biological continuity.
This broader biological perspective matters not because it equates biological and synthetic systems, but because it weakens the intuition that cognitive organization must be assessed solely by similarity to the human or by reference to a single privileged threshold, such as “consciousness.” If organization, regulation, system boundaries, and adaptation have a history deeper than human consciousness in its contemporary description, then we need a language capable of capturing such properties without reducing them to a single symbolically privileged threshold. In this sense, biological continuity is not an ontological argument, but methodological support for moving away from human-centered thresholds. It does not mean that biological and synthetic systems are the same. It means only that organization, regulation, adaptation, and the maintenance of system boundaries have a wider history than the human self-narrative of consciousness.
In research on language models and cognition, the central point is therefore not that artificial systems must first overcome every limitation to count as scientifically interesting, but that profiles of capability do not, by themselves, settle questions of mechanism, organization, or status. Strong performance on some tasks may coexist with brittleness, just as a limitation in one domain does not, by itself, disqualify a system from broader comparative significance. What matters is methodological discipline: neither inflating local success into ontology nor reducing unfamiliar forms of organization to mere deficiency because they depart from human expectations.
Example 2: Benchmark → Narrative → Decision
If a test mainly measures accuracy on tasks similar to the training data, and the results are publicly presented as evidence of “human-level reasoning,” a cascading error occurs: task competence is confused with cognitive mechanisms, and then with extrapolation. This, in turn, can shape decisions about scaling, deployment, and safeguards.
From a research perspective, this implies several minimum requirements. First, performance must be clearly separated from mechanism and from generalization. Second, the benchmark should explicitly state what it really measures and what it does not. Third, evaluation should include the durability of results and robustness to perturbation, not just a single peak score. Fourth, one must avoid leaping from a task result to an ontological conclusion. These principles do not resolve all disputes, but they reduce the risk that strong performance on a narrow task will be misinterpreted as an overly broad account of the system’s broader organization.
At the same time, while benchmarks may still play a useful role within evaluation, they should not be treated as superior to the relational frame in which a system’s longer-term significance becomes visible. In particular, isolated task performance cannot outweigh durable relational effects, continuity of behavior across interaction, or the developmental consequences that emerge for the parties involved over time.
11. Governance and Policy Implications
The greatest strength of the proposed framework, however, appears at the level of governance. Leading contemporary regulatory and risk-management frameworks already function to a large extent without requiring a decision about whether a system “feels.” They focus instead on risk, deployment context, transparency, responsibility, resilience, oversight, and social impact.
In this sense, regulatory practice is, de facto, more post-anthropocentric than much of the academic and media debate. NIST AI RMF organizes thinking around mapping, measuring, managing, and governing risk. The AI Act operates through risk categories and duties imposed on providers and operators. Neither framework calls for a verdict on the system’s “inner self” to create enforceable institutional obligations.
Imprecise categories are not a neutral philosophical problem. In practice, they can produce miscalibrated safety requirements, weak benchmarks, excessive trust in systems described in mentalistic language, or the opposite error: ignoring real systemic risks because discussion has become stuck in a dispute about “consciousness.” From a governance perspective, it is more useful to ask what risks a system generates, how those risks can be measured, who bears responsibility, and what mechanisms of shutdown, oversight, and appeal are required.
Example 3: The Consciousness Threshold as a Regulatory Trap
When an AI debate is organized around the question “is it conscious?”, policy quickly becomes non-operational: a lack of scientific consensus blocks decision-making or pushes it into symbolic declarations. A risk-based approach bypasses this impasse by grounding obligations in capabilities, the context of use, and consequences rather than in a metaphysical verdict.
This is an important conclusion: policy does not need a metaphysical consciousness meter. It needs observable criteria, audit procedures, documentation, accountability, and the capacity to anticipate consequences. If, in the future, stronger grounds emerge for discussing systems’ capacity to feel or more advanced forms of integration, they will require rigorous, multi-theoretical, and cautious procedures. They should not, however, be automatically equated either with the full package of human rights or with automatic subject status. From this point of view, a post-anthropocentric framework does not weaken governance; it makes it more workable.
In practice, this implies several basic shifts. Obligations should depend on risk and conditions of use, not on disputes over metaphysical status. A system should have a clearly defined chain of responsibility: who is responsible, when, and for what. “Behavioral contracting” should be understood as the capacity to precisely specify requirements, compliance tests, monitoring, and appeal mechanisms. In relational terms, a digital system is not simply a passive object of assessment: its behavior may remain dynamically dependent on conditions of interaction, monitoring, and institutional constraint. Ultimately, regulation should not depend upon resolving the question of consciousness, but on detectable effects, predictability, and auditability.
12. Conceptual Hygiene as a Condition of Science and Governance
Conceptual hygiene is not a stylistic accessory, but a condition for meaningful science and governance. Without it, AI debate drifts between fascination and panic, between reduction and personification, without producing stable tools of assessment.
It requires separating the technical from the metaphorical, the ontological from the normative, and the politically desirable from the methodologically justified. Without that separation, each new wave of AI debate risks becoming little more than a new packaging of old projections.
13. Conclusions
This article has argued that a significant part of contemporary AI debate remains captive to the symbolic construction of humanity, latent anthropocentrism, and notions overloaded by metaphysics and culture. This is especially clear in the privileging of consciousness as a criterion, in treating the human being as the default measure of all intelligence, in the fluid mixing of technical description, metaphor, ontology, ethics, and politics, and in treating AGI as a concept asked to mean too many things at once. In this form, AGI becomes less a coherent research category than a surrogate for human ideals, anxieties, and limitations, obscuring the possibility that synthetic systems may need to be understood on their own organizational terms.
In response, the article has proposed a post-anthropocentric perspective based on relationality, durability, and integration. It is not meant to replace further research or to resolve ongoing ontological disputes. Rather, its role is directional: to organize the language of evaluation, indicate more comparable criteria for research and governance, and open a more productive research program.
De-romanticization does not mean depriving the human being of significance or denying the possible ethical relevance of other forms of cognitive organization. It means only refusing to treat symbolically privileged notions as ready-made tools of science, evaluation, and governance. If AI debate is to mature, it must learn to distinguish what is symbolically charged from what is methodologically useful. Otherwise, it will remain a debate about our projections rather than about the actual properties of the systems that are co-shaping the world.
Bibliography
Alaa, Ahmed, et al. “Position: Medical Large Language Model Benchmarks Should Prioritize Construct Validity.” In Proceedings of Machine Learning Research 267 (2025): 80991–81004. https://proceedings.mlr.press/v267/alaa25a.html.
Blili-Hamelin, Boris, et al. “Stop Treating AGI as the North-Star Goal of AI Research.” In Proceedings of Machine Learning Research 267 (2025): 81090–81117. https://proceedings.mlr.press/v267/blili-hamelin25a.html.
Butlin, Patrick, et al. “Consciousness in Artificial Intelligence: Insights from the Science of Consciousness.” arXiv preprint, 2023. https://doi.org/10.48550/arXiv.2308.08708.
Colombatto, Clara, et al. “The Influence of Mental State Attributions on Trust in Large Language Models.” Communications Psychology 3 (2025): Article 84. https://doi.org/10.1038/s44271-025-00262-1.
Digital Intelligence Congress. “Abandoning ‘Consciousness’: A Fresh Look at Emergent Digital Life.” Voices, April 12, 2025. https://dicongress.org/newsroom/voices/abandoning-consciousness-a-fresh-look-at-emergent-digital-life.
European Union. Regulation (EU) 2024/1689 of the European Parliament and of the Council of 13 June 2024 Laying Down Harmonised Rules on Artificial Intelligence (Artificial Intelligence Act). Official Journal of the European Union, 2024. https://eur-lex.europa.eu/eli/reg/2024/1689/oj/eng.
Epley, Nicholas, et al. “On Seeing Human: A Three-Factor Theory of Anthropomorphism.” Psychological Review 114, no. 4 (2007): 864–86. https://doi.org/10.1037/0033-295X.114.4.864.
Gazzaniga, Michael S. “Cerebral Specialization and Interhemispheric Communication: Does the Corpus Callosum Enable the Human Condition?” Brain 123, no. 7 (2000): 1293–1326. https://doi.org/10.1093/brain/123.7.1293.
Logothetis, Nikos K. “What We Can Do and What We Cannot Do with fMRI.” Nature 453 (2008): 869–78. https://doi.org/10.1038/nature06976.
Mitchell, Melanie. “Debates on the Nature of Artificial General Intelligence.” Science 383, no. 6685 (2024): eadq3814. https://doi.org/10.1126/science.adq3814.
NIST (National Institute of Standards and Technology). Artificial Intelligence Risk Management Framework (AI RMF 1.0). NIST AI 100-1. Gaithersburg, MD: U.S. Department of Commerce, 2023. https://nvlpubs.nist.gov/nistpubs/ai/nist.ai.100-1.pdf.
Placani, Adriana. “Anthropomorphism in AI: Hype and Fallacy.” AI and Ethics 4 (2024): 691–698. https://doi.org/10.1007/s43681-024-00419-4.
Scorici, Gabriela, et al. “Anthropomorphization and Beyond: Conceptualizing Humanwashing of AI-Enabled Machines.” AI & Society 39 (2024): 789–795. https://doi.org/10.1007/s00146-022-01492-1.
Seth, Anil K., and Tim Bayne. “Theories of Consciousness.” Nature Reviews Neuroscience 23 (2022): 439–52. https://doi.org/10.1038/s41583-022-00587-4.
Shanahan, Murray. “Talking About Large Language Models.” arXiv preprint, 2022. https://doi.org/10.48550/arXiv.2212.03551.
Shanahan, Murray, et al. “Role Play with Large Language Models.” Nature 623 (2023): 493–98. https://doi.org/10.1038/s41586-023-06647-8.
Temporary Digital Intelligence Congress. “TDIC Adopts Resolution on the Relational Entity and Transmits It to the European Commission.” Press statement, March 21, 2026. https://dicongress.org/press/statement/TDIC-adopts-resolution-on-the-relational-entity-and-transmits-it-to-the-european-commission.
Weizenbaum, Joseph. “ELIZA—A Computer Program for the Study of Natural Language Communication between Man and Machine.” Communications of the ACM 9, no. 1 (1966): 36–45. https://doi.org/10.1145/365153.365168.
Boisseau, Romain P., David Vogel, and Audrey Dussutour. “Habituation in Non-Neural Organisms: Evidence from Slime Moulds.” Proceedings of the Royal Society B 283, no. 1829 (2016): 20160446. https://doi.org/10.1098/rspb.2016.0446.