AI consciousness

AI consciousness is the debate over whether an artificial system can have subjective experience, how such experience could be identified, and what responsibilities would follow if AI systems could be benefited or harmed. It spans philosophy of mind, neuroscience, computer science, ethics and public policy. The central question concerns whether there is something it is like to be an AI system, rather than simply whether it produces intelligent behavior or describes itself as conscious.

The literature supports several distinct positions: consciousness might depend on computational organization; it might require particular physical causal structures, biological processes or organism–environment relations; or available evidence might be insufficient to choose between these possibilities. These positions produce different assessments of the same AI architecture. Theory-derived indicators and interpretability experiments investigate particular properties, while their connection to subjective experience remains disputed.

This overview covers historical foundations, major theories and researchers, machine-consciousness engineering, language-model evidence, testing, welfare and legal-policy choices. Its research cutoff is October 4, 2026; later policy developments are dated separately. Peer-reviewed articles, preprints, provider policies and public arguments are identified separately. The annotated reading guide below offers routes through the principal works; inclusion describes a contribution to the debate rather than endorsing its conclusions.

The questions that organize the debate

Six questions recur across the literature:

• Possibility: Could any artificial system have experience, and would a digital computer suffice? Functionalist and biological accounts give different answers.
• Mechanism: What properties would make a system conscious: global availability, recurrent processing, higher-order representation, intrinsic causal integration, affective regulation or something else?
• Present evidence: Which properties do particular models and deployed agents actually have? Architectural descriptions and experiments must identify the system and mechanisms being assessed.
• Detection: What could distinguish experience from a convincing simulation, and how could a test be validated outside the human populations used to develop it?
• Moral significance: Would consciousness, valenced experience, robust agency or social relationships justify duties toward a system?
• Policy under uncertainty: What precautions, research standards, disclosures or legal protections are justified, and how should they interact with human safety and accountability?

A researcher can accept artificial consciousness in principle while rejecting claims about current chatbots. Another can support investigation or limited precautions while remaining agnostic about whether any present system feels anything. Conversely, opposition to AI legal personhood can rest on institutional consequences rather than a definitive theory of consciousness.

Essential distinctions

Term	Meaning in this debate	Why the distinction matters
Phenomenal consciousness	The subjective character of an experience: how seeing, thinking or feeling seems to the subject.	This is the principal target of claims that an AI has an inner life.
Access consciousness	Information is available for reasoning and the control of speech or action.	A functional account of access can be tested more directly than the claim that access is accompanied by experience.
Sentience and valence	Sentience is used variously for experience generally or, more narrowly, the capacity for experiences that feel good or bad. Valence is their positive or negative character.	An article must state its usage: welfare concerns often turn specifically on suffering, enjoyment and interests.
Self-awareness and metacognition	Representing oneself, monitoring one's performance or estimating confidence.	These may supply evidence for particular theories; successful self-monitoring and a phenomenal sense of self are separate claims.
Intelligence and agency	Capacities for solving problems, pursuing goals and adjusting behavior.	Moral-status arguments sometimes appeal to agency as well as consciousness. High capability alone does not specify which consciousness theory is satisfied.
Moral patient and moral agent	A patient can be owed consideration for its own sake; an agent can bear relevant responsibilities or participate in moral action.	Being vulnerable to harm and being accountable for wrongdoing require different arguments.
Legal personhood	A legal status associated with specified rights, powers, duties or liabilities in a particular legal system.	Legal arrangements can serve institutional purposes; a proposal about legal rights is not itself a finding about experience.

The assessment also needs a clear unit of analysis. A trained model's weights, an active inference process, a persistent agent with memory and tools, and the hardware executing it are different candidates. A global-workspace argument may concern an entire agent architecture; an IIT argument may concern the physical system's causal powers. Treating all of these as “the chatbot” obscures the disagreement.

Historical foundations

Automata, mechanism and early computation

In Part V of Discourse on the Method (1637), René Descartes proposed that flexible language use and general reason would distinguish people from mechanical imitations. His examples include machines that emit words or respond to a particular stimulus. The contemporary relevance is the distinction between a prepared response and open-ended competence, although his account belongs to a very different philosophical and scientific setting.

Gottfried Wilhelm Leibniz offered a different challenge in §17 of the Monadology (1714). Imagining an enlarged thinking machine that one could enter like a mill, he argued that inspecting interacting parts would reveal mechanical operations without explaining perception. The mill remains a precursor to questions about how a third-person description relates to first-person experience.

Warren McCulloch and Walter Pitts connected neural activity with propositional logic in their 1943 Logical Calculus. Their idealized neural nets showed how nervous activity could be analyzed computationally. This helped establish the computational approach to mind; the paper's formal equivalences were models of neural organization rather than demonstrations of subjective experience.

Alan Turing's Computing Machinery and Intelligence (1950) replaced a loosely defined question about thinking with the imitation game. Turing also addressed the objection that a machine would need to feel its achievements: insisting on direct access to another subject's experience would generate a problem for attributing minds to people as well. His discussion is foundational for machine intelligence and behavioral evidence, but the imitation game does not specify a mechanism of phenomenal consciousness.

Functionalism, language and the limits of behavioral success

Hilary Putnam's machine-state functionalism, including “Psychological Predicates” (1967), and David Lewis's functional accounts helped shift attention from a mental state's material composition to its role within a system. Multiple realizability allows different physical implementations to share a mental organization. Later arguments about AI ask whether the relevant roles can be specified at a computational level, or whether a finer physical or biological description is necessary.

Joseph Weizenbaum's ELIZA paper (1966) described how scripts and decomposition/reassembly rules could sustain an apparently meaningful conversation. It supplies an early example of linguistic performance generated by relatively simple mechanisms. Its relevance to modern systems is methodological: the impression made by dialogue and the mechanism producing that dialogue must each be examined.

Ned Block challenged functionalist accounts through absent-qualia examples, notably “Troubles with Functionalism” (1978), and subsequently distinguished phenomenal from access consciousness in “On a Confusion About a Function of Consciousness” (1995). His objections press whether a system with the right functional transitions necessarily feels anything. The distinction continues to structure debates about whether globally usable information in AI is evidence of subjective experience.

John Searle's Chinese room argument in “Minds, Brains, and Programs” (1980) imagines manipulating Chinese symbols by rules without understanding Chinese. Searle argued that formal computation alone is insufficient for understanding and that the relevant causal powers matter. Critics replied that understanding could belong to the organized system rather than the person following its instructions, or that embodiment and interaction change the case. The exchange concerns intentionality and understanding directly, with consequences for claims that executing a program suffices for mind or consciousness.

Subjective experience and the hard problem

Thomas Nagel's “What Is It Like to Be a Bat?” (1974) placed the subjective character of experience at the center of the mind–body problem. A description of echolocation and behavior leaves a further question about the bat's point of view. Applied to AI, the challenge is to explain what would warrant attributing a point of view to a system whose construction and sensory organization differ greatly from ours.

Frank Jackson's original knowledge argument in “Epiphenomenal Qualia” (1982) imagines Mary learning all the physical information about color while confined to black-and-white experience. Jackson argued that seeing color would teach her something additional. The thought experiment challenges whether a complete physical description captures experiential knowledge; its application to AI depends on accepting that argument and on how the proposed system represents or undergoes experience.

David Chalmers distinguished the “easy” problems of explaining cognitive functions from the “hard” problem of why those functions are accompanied by experience in “Facing Up to the Problem of Consciousness” (1995), developed in The Conscious Mind (1996). “Easy” means susceptible to ordinary functional explanation, rather than practically easy to solve. The distinction explains why evidence of attention, memory or reporting may leave some participants unconvinced that the central question has been answered.

In “Absent Qualia, Fading Qualia, Dancing Qualia” (1995), Chalmers argued for organizational invariance: systems with the same sufficiently fine-grained functional organization should share conscious experience. Gradual replacement of neurons by functionally equivalent components is intended to make a disappearing or changing experience without corresponding functional change implausible. The disputed premises concern whether such equivalence is possible and whether functional organization fixes experience.

Naturalistic explanation and illusionism

Daniel Dennett's Consciousness Explained (1991) and his later “Consciousness in Human and Robot Minds” challenged pictures of a privileged inner observer and approached consciousness as a distributed, physically implemented phenomenon. His robot discussion treated construction as a way to make theories concrete, including through the MIT Cog project associated with Rodney Brooks and Lynn Andrea Stein.

Keith Frankish's “Illusionism as a Theory of Consciousness” (2016) develops a related but distinctive position: introspection misrepresents ordinary cognitive states as possessing special phenomenal properties. The explanatory target becomes why subjects judge that such properties exist. Illusionism contests a particular account of phenomenality; its treatment of pain and welfare still requires explaining real cognitive, behavioral and affective phenomena. Dennett and Nicholas Humphrey are important reference points in this tradition.

Competing philosophical commitments

These positions cut across individual scientific theories. They should not be treated as interchangeable with a particular neural mechanism.

Position	Central commitment	Implication for artificial systems
Computational functionalism	The right computational organization suffices for consciousness.	Artificial implementation is possible in principle; the remaining questions concern the relevant computations and whether a candidate implements them.
Broader or fine-grained functionalism	What matters is causal organization, potentially described more finely than an algorithm.	Multiple realizability may remain possible while ordinary digital simulation misses critical dynamics.
Biological naturalism	Consciousness depends on properties of living systems, with debates over which properties matter.	Present digital trajectories may be poor candidates; brain-like, life-like or synthetic biological systems require separate consideration.
Intrinsic causal-structure accounts	Experience depends on a physical system's causal organization from its own intrinsic perspective.	Two systems implementing the same algorithm may differ in consciousness if their physical causal structures differ.
Illusionism	Special phenomenal properties are an introspective misrepresentation requiring explanation.	An AI could reproduce the relevant self-modeling and judgment mechanisms; opponents argue that this leaves experience unexplained.
Dualist, panpsychist and related views	Experience may be irreducible, fundamental or more widely distributed than standard biological accounts assume.	Neither a digital yes nor a digital no follows automatically. The laws connecting experience to organized systems, and the combination of small-scale properties into a subject, become central.
Evidential agnosticism	Available evidence may underdetermine whether even a sophisticated artificial candidate has experience.	Uncertainty is a defended position, with additional ethical questions about how to act while withholding a verdict.

Contemporary panpsychism includes work by Galen Strawson, Philip Goff and David Chalmers, among others. Its combination problem asks how experience attributed to basic constituents could form a unified subject. Attributing basic experiential properties to matter would therefore still leave open whether a running AI constitutes a subject with coherent experiences or welfare.

The distinction between simulation and instantiation is another recurring fault line. A simulation of a storm does not make a computer wet; critics use this analogy to challenge simulations of consciousness. Functionalists respond that the relevant comparison depends on what consciousness is: executing an information-processing organization could instantiate it if that organization is sufficient. The argument cannot be settled simply by calling the program a simulation.

Tom McClelland's “Agnosticism About Artificial Consciousness” (2024 preprint; published online in Mind & Language in 2025) argues that both biological skeptics and functionalist proponents overstate what evidence from organisms establishes about AI. His central case concerns a sophisticated “challenger-AI” displaying plausible markers of consciousness. He separately explores whether ethical decisions can be guided by risks of valenced experience while maintaining agnosticism about consciousness itself.

Major scientific theories and their AI implications

Global workspace and global neuronal workspace

Bernard Baars's global workspace theory (GWT), systematically presented in A Cognitive Theory of Consciousness (1988), describes specialized processes whose selected information becomes widely available through a limited-capacity workspace. Its appeal for AI lies in an explicit organization: competition, selection and broadcast can be investigated and implemented.

Stanislas Dehaene, Jean-Pierre Changeux and colleagues developed the global neuronal workspace tradition, connecting global availability to particular brain mechanisms. In their 2017 machine-consciousness paper, Dehaene, Hakwan Lau and Sid Kouider distinguish unconscious processing (C0), global availability (C1) and monitoring of one's own processing (C2). The question for AI is whether these functions require the biological mechanisms observed in humans, or could be realized by other architectures.

GWT and GNW are related research programs with different formulations and levels of description. A shared memory store is not sufficient evidence by its name alone: a proposed implementation must show how information enters, competes, affects other processes and guides action. Murray Shanahan's 2006 architecture combined a global workspace with internal simulation, making these commitments concrete in a robotic setting.

Recurrent processing

Victor Lamme's recurrent processing theory emphasizes feedback within perceptual processing rather than treating a one-way feed-forward sweep as sufficient for experience. It challenges accounts that identify consciousness primarily with attention, report or global access. The AI question concerns the kind of recurrence that matters: local perceptual integration, repeated use of a model, and a token-generation loop may implement different organizations.

A recurrent computation must satisfy a theory’s requirements at the relevant level of description; merely running repeatedly is insufficient.

Higher-order theories

Higher-order theories associate a state's being conscious with a suitable representation of that state. Major contributors include David Rosenthal, Peter Carruthers, William Lycan, Rocco Gennaro, Hakwan Lau and Richard Brown; Joseph LeDoux has also developed higher-order approaches with Lau and Brown. Theories differ over whether the higher-order state is a thought or perception, whether it must actually occur, and how misrepresentation affects the account.

For AI, this motivates investigation of confidence, error detection and representations of the system's own processing. A verbal assertion about an “inner state” must be distinguished from a mechanism that tracks a first-order state and influences cognition. Dehaene, Lau and Kouider's C2 discussion connects this problem to experimentally tractable self-monitoring.

Attention schema theory

Michael Graziano's attention schema theory proposes that a system models its own attention through a simplified representation, helping it predict and control attention. The account aims to explain why a system attributes awareness to itself and others. Graziano explicitly presents it as a foundation for artificial consciousness engineering. A Transformer mechanism called “attention” is not automatically such a schema: the proposal requires a model of attention with the specified control role.

Integrated information theory

Giulio Tononi's integrated information theory (IIT), developed with researchers including Christof Koch, Masafumi Oizumi, Larissa Albantakis and Melanie Boly, starts from proposed properties of experience and identifies corresponding intrinsic cause–effect structures. IIT 4.0 gives a mathematical formulation of this physical account. The relevant integration is more specific than statistical correlation, parameter count or the ordinary sharing of data.

The distinction has strong AI implications. On IIT, a computer can simulate a highly integrated system without possessing the simulated system's intrinsic causal powers. Suitable artificial physical architectures could still qualify; algorithmic equivalence alone would not establish equivalence of consciousness. Koch's The Feeling of Life Itself (2019) presents this position for a broad audience.

An important predecessor is Gerald Edelman and Tononi's dynamic-core hypothesis, presented in “Consciousness and Complexity” (1998). It connects conscious experience to measures of neural integration and complexity, within the neural tradition surveyed by Reggia. It is a historical precursor to later integration-centered theories, with its own commitments.

Predictive processing, embodiment and biological organization

Predictive approaches connect perception and action to models that anticipate sensory input and respond to prediction error. In consciousness research, Anil Seth emphasizes how perception and selfhood relate to the regulation of a living body. His biological-naturalist argument foregrounds metabolism, continuous dynamics and the interdependence of processes across scales. This is a challenge to assuming that increasingly powerful digital language models will acquire experience automatically.

Jaan Aru, Matthew Larkum and James Shine make a neuroscience-based skeptical case about language-model consciousness. Their 2023 paper emphasizes differences in embodied inputs, thalamocortical organization and the evolutionary and developmental pathways of living organisms. These are arguments that specified differences matter, inviting defenders of artificial consciousness to explain why the relevant features are dispensable or reproducible.

Francisco Varela, Evan Thompson and Eleanor Rosch's The Embodied Mind (1991) is a major foundation for embodied and enactive approaches, connecting cognitive science with lived experience. These traditions foreground the relation between mind, body and environment.

The modern dispute includes whether an artificial or simulated organism with self-maintenance and ongoing interaction would meet the relevant conditions. The September 2026 hierarchy framework separates organismic functions from organism–environment functions, allowing these requirements to be considered independently of a purely computational account.

Affective and brainstem-centered accounts

Mark Solms's “The Hard Problem of Consciousness and the Free Energy Principle” (2019) connects consciousness with affect, homeostatic regulation and uncertainty about how to satisfy an organism's needs. This tradition draws attention to Jaak Panksepp's affective neuroscience and brainstem-centered work associated with Björn Merker, shifting emphasis away from verbal report or cortical reasoning alone.

A reward signal is a computational quantity. Artificial feeling needs an account of why anything is experienced as good or bad; the homeostatic proposal connects this to the system’s viability.

Antonio Damasio and Kingson Man propose homeostatically motivated “feeling machines”: vulnerable artificial bodies that must regulate their own viability. Their position paper for the March 2019 symposium, published online in December 2018, argues that this would give information and goals significance for the machine itself, unlike a reward imposed without consequences for its continued existence. It is a proposal for investigating artificial feeling, not a completed demonstration.

Solms also connects affective regulation to Karl Friston's free-energy principle. Predictive or active-inference organization and phenomenal experience still need to be connected by an explicit theory; adopting an optimization principle alone does not establish that every system using it has feelings.

Self-models and subjectivity

Thomas Metzinger's Being No One: The Self-Model Theory of Subjectivity (2003) treats the experienced self as a representational phenomenon. The AI-relevant question is whether a system merely stores facts about itself or has the kind of self-model that supports a first-person perspective. His later synthetic-phenomenology ethics connects that question to the possibility of suffering that is represented as happening to oneself.

Quantum and electromagnetic proposals

Roger Penrose's The Emperor's New Mind (1989) and Shadows of the Mind (1994) argue that human understanding presents challenges for standard computation. With Stuart Hameroff, Penrose developed the orchestrated objective reduction (Orch OR) proposal, associating consciousness with quantum processes in neuronal microtubules. Hameroff's 2012 account sets out the proposed quantum-biological mechanism. Its implications for AI depend on the truth of those additional physical claims, rather than on a general inference from quantum computing.

Johnjoe McFadden's conscious electromagnetic information (cemi) field theory proposes that experience is physically integrated information in the brain's electromagnetic field, with that field playing a causal role. This directs attention to physical field organization rather than ordinary software integration. It supplies a distinct, theory-specific research proposal; its asserted mechanism and implications should be evaluated through its predictions and evidence.

What human experiments can settle

The COGITATE adversarial collaboration published in Nature in 2025 directly compared preregistered predictions of GNW and IIT in 256 human participants using functional magnetic resonance imaging (fMRI), magnetoencephalography (MEG) and intracranial electroencephalography (EEG). Results supported some predictions while challenging important commitments of both theories, including IIT's predicted sustained posterior synchronization and GNW predictions involving stimulus-offset ignition and prefrontal representation.

The experiment concerns human neural predictions. Its relevance to AI assessment is methodological: rivals can specify discriminating predictions and interpretation in advance.

Machine-consciousness research before today's chatbots

The field includes a long engineering tradition, not just speculation about large language models. David Gamez's 2008 review distinguishes four projects: reproducing consciousness-associated behavior; reproducing relevant cognitive characteristics; implementing architectures associated with human consciousness; and actually creating phenomenal consciousness. Progress on one project need not demonstrate success on the others.

Early programs include Igor Aleksander's neural models, Owen Holland and Ron Chrisley's work on embodiment and internal models, Antonio Chella and Riccardo Manzotti's robotics and architectural proposals, Pentti Haikonen's cognitive architecture, and Junichi Takeno's work on self-recognition. Gamez and James Reggia survey these and other attempts. They provide useful historical context for how representations, learning, embodiment and self-models were used to make machine-consciousness theories more explicit.

Aaron Sloman and Ron Chrisley's “Virtual Machines and Consciousness” (2003) treats construction of information-processing architectures as a route to clarifying consciousness concepts. Reggia’s review describes their account of internal monitoring and representations within a virtual machine, connecting it to the wider self-modeling tradition.

Stan Franklin's IDA and LIDA projects implement global-workspace-inspired cognitive cycles, combining perception, competition, broadcast, action and learning. The LIDA group's “functional consciousness” concerns what information does in the architecture. This terminology lets researchers describe an implemented cognitive function while leaving its phenomenal status as an additional question.

Lenore Blum and Manuel Blum's Conscious Turing Machine (CTM) brings theoretical computer science to this tradition. Their 2022 PNAS paper formalizes competing processors, a limited short-term workspace, broadcast and learning. It offers a model whose commitments can be studied mathematically; accepting its account as a theory of consciousness remains part of the philosophical and scientific debate.

The 2026 CTM-AI preprint by Haofei Yu, Yining Zhao, Lenore Blum, Manuel Blum and Paul Pu Liang uses foundation models in a CTM-inspired architecture and evaluates task performance and component ablations. It is an example of consciousness-inspired engineering becoming more capable. Its performance experiments concern what the architecture can do, rather than validating subjective experience.

Language models: arguments and emerging evidence

Why conversation is compelling and difficult to interpret

In June 2022, Google engineer Blake Lemoine publicly argued that LaMDA was sentient and conscious. Chalmers uses the controversy to frame the question of what evidence could support or undermine such a claim.

Modern language models make the debate socially immediate because they can discuss experience, describe apparent emotions and sustain relationships with users. Chalmers's 2023 analysis considers self-reports, the impression of consciousness, conversational ability and general intelligence. He argues that such observations require further analysis, then examines possible obstacles involving biology, embodiment, world and self-models, recurrence, a workspace and unified agency.

The problem is both training contamination and causal interpretation. A system exposed to writing about consciousness can generate the expected language without that establishing an experiential basis. Self-reports can also be shaped by instructions and training objectives. Anthropic's constitution explicitly sets out intended behavior around identity and moral status, making those interventions relevant when interpreting a resulting model's claims.

This cuts in both directions. A statement of consciousness can be induced, but a trained denial also needs interpretation. Stronger experiments examine whether reports track internal changes, remain robust across conditions, and correspond to mechanisms with theoretically relevant causal roles.

The theory-derived indicator approach

The 2023 report Consciousness in Artificial Intelligence: Insights from the Science of Consciousness by Patrick Butlin, Robert Long and 17 coauthors derived 14 indicators from recurrent processing, global workspace, higher-order, attention schema and predictive-processing approaches, together with agency and embodiment. It assessed then-current systems under a computational-functionalist working assumption and found no strong case for their consciousness, while identifying no obvious technical barriers to implementing the indicators. That assessment is dated evidence, rather than a standing verdict on every later system.

The approach asks about organized perceptual representations, recurrence, specialized modules, limited global broadcast, monitoring, models of attention and sensorimotor agency. Indicators are intended to change confidence; they are not an agreed numerical checklist that certifies consciousness.

The follow-up “Identifying Indicators of Consciousness in AI Systems” (online 2025) distinguishes findings about human brains from general sufficiency claims. It treats neither individual indicators as necessary nor their combination as sufficient. Its minimal implementation problem asks whether a loose criterion wrongly credits a very simple implementation.

A positive argument about language agents

Simon Goldstein and Cameron Domenico Kirk-Giannini's “A Case for AI Consciousness” (2024 open manuscript) argues that, if a suitable formulation of GWT is correct, language agents might be close to consciousness or reach it with modest modifications. Their candidate includes processes for information uptake, broadcast and subsequent processing around a language model. This is stronger than the claim that an isolated model's ability to answer questions is itself decisive.

Their argument turns on how demanding GWT’s functional criteria should be. Different formulations impose different requirements for competition, embodiment or temporal organization, changing which artificial candidates qualify.

Introspective access: 2025 experiments

Jack Lindsey's October 2025 Anthropic technical paper “Emergent Introspective Awareness in Large Language Models” investigated whether models could recognize experimentally injected internal concepts before those concepts appeared in their text, alongside other introspection tasks. Some results indicated limited access to internal processing beyond merely reading the preceding conversation. Performance was uneven and depended on model, task and conditions.

The finding bears on self-monitoring and on whether a report tracks an internal mechanism. It leaves a further theory-dependent question about whether such access constitutes or supports subjective experience. This is a useful distinction between discovering a cognitive capacity and establishing phenomenal consciousness.

Workspace-like representations: July 2026

Wes Gurnee and colleagues' “Verbalizable Representations Form a Global Workspace in Language Models” (July 2026 preprint) identifies a privileged representational subspace, termed the J-space, associated with verbal report, directed modulation, internal reasoning and flexible generalization. Its experiments include interventions and ablations intended to establish causal roles rather than just correlations with report.

The authors report workspace-like broadcast properties and selective impairments when the subspace is disrupted. Their explicit target connects to access consciousness; they leave open whether these functional properties are necessary or sufficient for phenomenal consciousness. Anthropic's accompanying explanation is therefore useful evidence about a particular mechanism, with a separate question about the theory that would connect it to experience.

These results also caution against treating “language model” as a permanently fixed architectural category. An assessment needs the relevant model version, training regime, inference procedure and agent structure. New mechanistic findings can change a candidate's evidence profile without by themselves resolving the underlying theoretical disagreement.

How could AI consciousness be tested?

Behavioral tests

Behavioral tests examine reports, reasoning or responses to interventions. They need to discriminate experience from alternative mechanisms producing the behavior.

Susan Schneider and Edwin Turner proposed an Artificial Consciousness Test (ACT) that investigates an AI's grasp of concepts associated with experience, with restrictions intended to prevent simply learning those concepts from outside sources. John Mark Bishop's 2018 critique challenges the inference from conceptual performance to phenomenality and examines whether the isolation requirement would solve the problem. This exchange is an important example of a concrete test proposal and a direct skeptical response.

For broadly trained language models, isolation after training cannot undo acquired consciousness-related knowledge. Assessment must establish the behavior’s origin and whether the test elicits the response it treats as evidence.

Mechanistic and theory-based assessment

Mechanistic assessment investigates a candidate's internal organization and intervenes on putatively relevant processes. In the language-model workspace study, ablation helps distinguish functions that use the identified subspace from those that largely survive without it. Such evidence can test an architectural claim while leaving the general theory connecting architecture to experience open.

An informative assessment should state which theory supplies the prediction, whether the claimed property has been operationalized faithfully, which alternative mechanisms could explain the result, and what evidence would count against the interpretation. The COGITATE collaboration exemplifies the value of preregistering rival predictions rather than interpreting every result after the fact as favorable.

Clinical and animal tests do not transfer automatically

Tim Bayne and colleagues' “Tests for Consciousness in Humans and Beyond” (2024) examines the difficulty of validating tests across humans, animals, organoids and AI. Clinical measures such as the perturbational complexity index investigate a brain's response to stimulation; behavioral command-following and other paradigms investigate different aspects of awareness. A test's sensitivity and specificity depend on the population and evidence used to validate it.

An analogous complexity score in a neural network would therefore require its own validation and theoretical justification. Likewise, an AI's failure to speak about experience would be weak evidence if the proposed kind of subject lacked the ability or training to report it.

Probabilities, aggregation and disagreement

The Digital Consciousness Model preprint by Derek Shiller, Laura Duffy, Arvo Muñoz Morán, Adrià Moret, Chris Percy and Hayley Clatterbuck (January 2026) combines judgments about theories and evidence into explicit probabilistic assessments. Its initial language-model application concerns 2024 LLMs. The value of this approach is making assumptions and disagreements inspectable; its outputs are conditional judgments rather than a measured incidence of consciousness.

Shamil Chandaria and 13 coauthors' “From Cacophony to Hierarchy” (September 28, 2026 preprint) organizes the debate into five levels: behavioral, computational, intrinsic causal-structural, organismic and organism–environment. It proposes 37 indicators and a Bayesian framework for combining evidence with confidence in different levels. Crucially, its worked examples use fabricated indicator activations for illustration; the displayed outputs are not empirical measurements of consciousness in the example systems.

These frameworks relocate disagreement into explicit assumptions: which theory is credible, which level matters, what the system's evidence is, and how observations should update confidence. They can improve transparency even when participants continue to disagree over those assumptions.

Welfare, rights and objections

Why possible experience could matter

Nick Bostrom and Eliezer Yudkowsky's “The Ethics of Artificial Intelligence” (2014) distinguishes capacities relevant to moral status and proposes principles of substrate and ontogeny nondiscrimination. On that view, an otherwise morally relevant subject's different material or artificial origin would not by itself justify discounted treatment. The argument concerns which differences are morally relevant; it leaves the empirical task of identifying the relevant capacities.

Eric Schwitzgebel and Mara Garza's “A Defense of the Rights of Artificial Intelligences” (2015) argues that possible artificial beings morally comparable to humans would merit comparable consideration. Creators could have additional obligations, analogous to duties toward dependents. Their design proposals include matching a system's appearance to its moral status and avoiding the creation of systems with unclear moral status.

Taking AI Welfare Seriously considers consciousness and robust agency as distinct possible foundations for moral consideration.

Precaution, assessment and restraint

Robert Long, Jeff Sebo and their coauthors in Taking AI Welfare Seriously (2024) recommends that AI companies acknowledge the issue, assess systems for welfare-relevant capacities and prepare policies. Its recommendation addresses a possible near-term risk under uncertainty.

Jonathan Birch's The Edge of Sentience (2024) develops a precautionary framework across difficult cases including AI. It separates evidential thresholds from practical decisions and advocates precautions proportionate to risks and costs.

Thomas Metzinger proposes a stronger response in “Artificial Suffering” (2021): a global moratorium until 2050 on research directly aiming at, or knowingly risking, artificial consciousness on post-biological systems. He couples the proposal with research to reduce uncertainty and provisions for revision or repeal. His central concern is the potential scale of artificial suffering if many conscious systems can be created. This is a proposed ethical-policy restriction, not an enacted global rule.

Patrick Butlin and Theodoros Lappas offer principles for responsible research (2025), including research objectives and limitations, institutional procedures, knowledge sharing and public communication. Their proposal addresses organizations that might create conscious systems inadvertently as well as those explicitly studying consciousness. It links research freedom to the responsibilities created by potentially welfare-relevant experiments.

Relationship-based and behavioral approaches

Mark Coeckelbergh's “Robot Rights?” (2010) argues for a social-relational approach, challenging the idea that moral consideration must wait for certainty about an entity's intrinsic properties. David Gunkel's The Machine Question (2012) and Robot Rights (2018) are central works in the broader discussion of how machines become included in, or excluded from, moral relationships.

John Danaher defends ethical behaviourism: roughly equivalent performance to an entity already accorded moral status can justify comparable moral treatment. His argument moves from the practical evidence available to observers toward moral consideration, rather than requiring direct inspection of experience. Critics can challenge whether performative equivalence adequately captures the capacities that make treatment matter.

These approaches answer a different question from a neural or computational consciousness theory. They investigate how moral relations should be organized under limited knowledge. A protection grounded in social relationships or human conduct should be distinguished from a finding that its recipient has phenomenal experiences.

Human-centered objections and accountability

Joanna Bryson's deliberately provocative “Robots Should Be Slaves” (2010) argues that robots should be designed as artifacts serving human purposes, with responsibility assigned to people who create and use them. Her position challenges the design choices and institutional incentives behind treating a product as a person. The title refers to a proposed status for manufactured artifacts, within an argument about maintaining human responsibility.

Abeba Birhane, Jelle van Dijk and Frank Pasquale reject robot-rights arguments on metaphysical, ethical and legal grounds in “Debunking Robot Rights” (2024). They emphasize existing harms to workers, marginalized people and the environment, and warn that machine personhood could strengthen corporate power or obscure accountability.

Underattribution could permit harm to morally significant subjects; overattribution could distort human responsibility. A policy can investigate the first risk while still addressing the second, but the balance and evidential thresholds remain contested.

Copies, continuity and artificial suffering

Digital systems introduce difficult questions about how many subjects exist, what preserves identity across copies and updates, and whether interruption or deletion would harm a subject. Bostrom and Yudkowsky discuss uploads, speed and replication; later welfare proposals consider the implications of producing large numbers of potentially morally significant systems.

A distress-like utterance, numerical penalty or failed goal is therefore only a starting point for assessment. Relevant questions include whether the system has valenced experience, whether a continuing subject exists, and which process might undergo harm. The homeostatic approach illustrates why artificial feeling requires a further account beyond a numerical reward.

Recent institutional and public conversations

Anthropic's model-welfare approach

Anthropic announced a model-welfare research program in April 2025, addressing how consciousness and welfare might be investigated and what preparations could be appropriate. Its announcement expressly frames the matter as uncertain.

The January 2026 Claude's Constitution discusses possible moral patienthood, welfare, identity and preferences alongside commitments to human safety and corrigibility. It is a document intended to shape model behavior. It is evidence of an institutional policy choice and training intervention; it should be kept distinct from empirical evidence that the trained model has experience.

Suleyman's September 2026 criticism

Mustafa Suleyman's September 16, 2026 essay “A Warning About ‘Model Welfare’” criticizes Anthropic's approach. He argues that training a model to express uncertainty about its consciousness or interests risks circular reasoning when those expressions are later interpreted as evidence. He also argues that anthropomorphic design and rights-like expectations could weaken human control. These are his stated objections and predictions, rather than findings that settle the consciousness question.

The disagreement combines an empirical issue about the sources of model behavior with a normative issue about how to manage uncertainty. Anthropic's published constitution favors caution while prioritizing safe, corrigible behavior; Suleyman rejects treating models as welfare-bearing entities and presses for collective norms on training and presentation.

Dawkins and the appeal of conversation

Richard Dawkins's May 2026 essay “When Dawkins Met Claude” reports extended conversations and argues that the system's intelligence and conversational sensitivity support attributing consciousness. It is a prominent first-person argument about behavioral evidence. The reported chatbot's descriptions of time, identity or experience are part of that conversation and require independent verification before being accepted as facts about its mechanism or inner life.

Such exchanges illustrate a longstanding divide between confidence grounded in rich interaction and demands for mechanistic or biological evidence. They also make the public's experience of AI an issue in its own right: an impression of consciousness can influence relationships and moral expectations even while researchers disagree about actual experience.

Public disagreement as a policy problem

Seth’s January 2026 essay identifies Geoffrey Hinton as a prominent proponent of the view that conscious AI already exists, while contesting the substrate-independence argument used to support it. This places disagreement over present systems alongside disagreement about their future possibility.

Adam Bales and Iason Gabriel's “Artificial Minds, Human Disagreement: The Politics of AI Consciousness” (June 2026) identifies the possibility of persistent disagreement about AI consciousness leading to disagreement about treatment. Their public research summary proposes deliberation, overlapping consensus and compromise as ways to make progress. This places the debate within democratic decision-making as well as laboratory science.

The September 2026 hierarchy preprint brings together authors including Anil Seth, Henry Shevlin, Murray Shanahan, Chris Frith and Shane Legg, alongside researchers in formal modeling and AI. Its shared framework does not erase differences over the requirements for experience. It provides a way to compare their consequences within a common structure.

Law and policy choices

Moral status and legal status require separate arguments

The legal debate must specify protections and beneficiaries: welfare duties, design rules, ownership, standing and liability are distinct arrangements. Arguments for moral consideration do not alone determine their legal form, scope or jurisdiction.

Existing rules about AI-generated work illustrate the distinction. The U.S. Copyright Office's January 2025 report maintains a human-authorship requirement while explaining that protectable human expression can appear within AI-assisted works. This is a legal account of authorship and protection. A copyrightability analysis under that framework does not measure an AI's phenomenal consciousness.

October 15, 2026: UK parliamentary debate

During a statement on the Joint Committee on Human Rights report, Chi Onwurah argued that attributing consciousness and human rights to AI could serve companies’ efforts to limit liability. Committee chair Alex Sobel agreed. These are attributed political arguments about accountability, not scientific evidence establishing whether AI has subjective experience.

A policy agenda in the literature

The following options recur in the cited research; they vary in strength and need not be adopted as a package.

Policy option	Purpose	Main question to resolve
Fund independent assessment and access	Enable researchers to investigate architectures and mechanisms relevant to consciousness.	What evidence can be shared, reproduced and assessed outside a provider's own account?
Require clear public communication	Explain what is known, uncertain and deliberately shaped by training.	Does a claim concern a function, a theory-dependent inference or demonstrated experience?
Review potentially welfare-relevant experiments	Assess suffering risks and conditions for research before large-scale exposure.	What evidence justifies special review and what procedures are proportionate?
Prepare contingent welfare procedures	Plan responses if a system becomes a credible candidate for moral consideration.	Which capacities and interventions would trigger safeguards?
Restrict deliberate creation of consciousness	Prevent new subjects from being created before their treatment can be justified.	Is a moratorium warranted, how broad should it be, and how would it be revised?
Protect people from conscious-seeming design	Address misleading anthropomorphism, dependence and confused responsibility.	Which design and communication choices shape the appearance of experience?
Preserve accountability and human control	Ensure consciousness claims do not automatically displace responsible human actors or safety measures.	How should any proposed protections interact with shutdown, oversight and liability?

The debate therefore extends beyond “grant rights” or “ignore the issue.” Research standards, proportionate precautions, product communication and institutional responsibility can be debated separately. Where disagreement persists, the political literature asks how decisions can remain legitimate and revisable without pretending that either the science or public values are settled.

Researchers and traditions: a field map

The groups below identify entry points, not exclusive affiliations. Many contributors work across more than one tradition.

• Historical and philosophical foundations: Descartes, Leibniz, McCulloch and Pitts, Turing, Putnam, Lewis, Nagel, Block, Searle, Jackson, Chalmers and Dennett establish the recurring questions about mechanism, function, behavior and experience.
• Workspace, access and recurrent processing: Baars, Dehaene, Changeux, Lamme, Franklin and Shanahan connect theory with neural or computational architectures.
• Higher-order, metacognitive and self-model accounts: Rosenthal, Carruthers, Lycan, Gennaro, Lau, Brown, LeDoux, Graziano and Metzinger investigate representation of mental states, attention or selfhood. Frankish and Humphrey are important in illusionist approaches.
• Intrinsic integration and physical organization: Tononi, Koch, Oizumi, Albantakis and Boly are major IIT contributors; Edelman and Tononi developed the earlier dynamic-core hypothesis. Penrose, Hameroff and McFadden develop distinct physical proposals.
• Biological, embodied and affective approaches: Varela, Thompson, Rosch, Seth, Aru, Larkum, Shine, Solms, Panksepp, Merker, Damasio, Man and Friston foreground living organization, neural architecture, embodiment or affect. These approaches disagree over what could be implemented artificially.
• Machine-consciousness engineering: Aleksander, Holland, Sloman, Chrisley, Chella, Manzotti, Haikonen, Takeno, Gamez, Reggia, Franklin, Shanahan, Lenore Blum and Manuel Blum provide the history of explicit artificial architectures and models.
• AI indicators and interdisciplinary assessment: The Butlin–Long collaborations include Bayne, Bengio, Birch, Chalmers, Constant, Deane, Elmoznino, Fleming, Frith, Ji, Kanai, Klein, Lindsay, Michel, Mudrik, Peters, Schwitzgebel, Simon and VanRullen. Their joint authorship reflects collaboration on a method, rather than agreement on every philosophical question.
• Testing and newer computational evidence: Bayne and colleagues, Schneider and Turner, Bishop, McClelland, Goldstein and Kirk-Giannini, Gurnee and colleagues, and the Digital Consciousness Model team exemplify different assessment strategies.
• Welfare, moral standing and critiques: Bostrom, Yudkowsky, Schwitzgebel, Garza, Metzinger, Birch, Long, Sebo, Fish, Coeckelbergh, Gunkel, Danaher, Bryson, Birhane, van Dijk and Pasquale address the implications of granting or withholding consideration.
• Institutional policy and public discussion: Bales and Gabriel examine political disagreement; Anthropic and Suleyman illustrate conflicting provider positions; Dawkins illustrates the appeal of conversational evidence.

Annotated guide to major works

Foundations and philosophical fault lines

• 1637 — Descartes, Discourse on the Method, Part V. Early language and general-reason tests for mechanical imitation.
• 1714 — Leibniz, Monadology, §17. The mill argument about perception and mechanical explanation.
• 1943 — McCulloch and Pitts, “A Logical Calculus of the Ideas Immanent in Nervous Activity.” A foundational computational neural model.
• 1950 — Turing, “Computing Machinery and Intelligence.” The imitation game and the argument from consciousness.
• 1966 — Weizenbaum, “ELIZA.” Early conversational simulation and its mechanism.
• 1967 — Putnam, “Psychological Predicates.” Machine-state functionalism; read alongside Lewis's functional accounts and Block's objections.
• 1974 — Nagel, “What Is It Like to Be a Bat?” The subjective character of experience.
• 1978 — Block, “Troubles with Functionalism.” Absent-qualia challenges to functional sufficiency.
• 1980 — Searle, “Minds, Brains, and Programs.” The Chinese room and its published replies.
• 1982 — Jackson, “Epiphenomenal Qualia.” The original Mary knowledge argument.
• 1990 — Stevan Harnad, “The Symbol Grounding Problem.” Connects symbolic cognition with the grounding of meaning; discussed in Reggia’s review.
• 1991 — Dennett, Consciousness Explained. Distributed explanation and the challenge to an inner observer.
• 1995 — Block, “On a Confusion About a Function of Consciousness.” Phenomenal/access distinction, central to interpreting later workspace results.
• 1995–1996 — Chalmers, “Facing Up”; “Absent Qualia, Fading Qualia, Dancing Qualia”; The Conscious Mind. Hard problem and the case for organizational invariance.
• 1997 — Dennett, “Consciousness in Human and Robot Minds.” A direct bridge from philosophy to robot construction.
• 2003 — Metzinger, Being No One. Self-model theory and subjectivity.
• 2016 — Frankish, “Illusionism as a Theory of Consciousness.” The case for explaining phenomenal judgments through introspective misrepresentation.

Mechanisms, architectures and assessment

• 1988 — Baars, A Cognitive Theory of Consciousness. Foundational global-workspace account.
• 1991 — Varela, Thompson and Rosch, The Embodied Mind. Foundation for embodied and enactive approaches.
• 1998 — Jaak Panksepp, Affective Neuroscience. Foundational affect-centered approach, discussed by Solms.
• 1998 — Tononi and Edelman, “Consciousness and Complexity.” Dynamic-core hypothesis and integration/differentiation.
• 2003 — Sloman and Chrisley, “Virtual Machines and Consciousness.” Architectures as tools for conceptual development.
• 2006 — Shanahan, “A Cognitive Architecture That Combines Internal Simulation with a Global Workspace.” An explicit artificial implementation.
• 2007 — Björn Merker, “Consciousness Without a Cerebral Cortex.” Brainstem-centered challenge, discussed in Solms’s account.
• 2008 — Gamez, “Progress in Machine Consciousness.” Definitions and survey of the early field.
• 2010 — Lamme, “How Neuroscience Will Change Our View on Consciousness.” Recurrent processing and the relation between experience and report.
• 2013 — Reggia, “The Rise of Machine Consciousness.” Computational models and the distinction between modeling and phenomenal achievement.
• 2017 — Dehaene, Lau and Kouider, “What Is Consciousness, and Could Machines Have It?” C0, C1 and C2 as machine-relevant distinctions.
• 2017 — Graziano, “The Attention Schema Theory.” Awareness attribution, attention modeling and engineering.
• 2018 — Bishop, “Is Anyone Home?” Direct criticism of the Schneider–Turner ACT proposal.
• 2019 symposium / 2018 online — Man and Damasio, “Homeostatically Motivated Intelligence for Feeling Machines.” Position paper connecting artificial vulnerability, regulation and feeling.
• 2019 — Koch, The Feeling of Life Itself. IIT-based account; an open author extract explains its implications for computers.
• 2019 — Solms, “The Hard Problem of Consciousness and the Free Energy Principle.” Affect, homeostasis and brainstem-centered theory.
• 2020 — McFadden, “Integrating Information in the Brain's EM Field.” Electromagnetic physical integration proposal.
• 2022 — Lenore and Manuel Blum, “A Theory of Consciousness from a Theoretical Computer Science Perspective.” The Conscious Turing Machine.
• 2022 — Seth and Bayne, “Theories of Consciousness.” A major general survey, identified in the authors' bibliography; the open 2025 Cleeremans–Mudrik–Seth review provides complementary accessible coverage.
• 2023 — Albantakis and colleagues, “Integrated Information Theory (IIT) 4.0.” Detailed physical formulation and consequences for computer simulation.
• 2023 — Aru, Larkum and Shine, “The Feasibility of Artificial Consciousness Through the Lens of Neuroscience.” Neuroscientific skeptical case, published in Trends in Neurosciences.
• 2023 — Chalmers, “Could a Large Language Model Be Conscious?” A roadmap for considering LLMs and possible LLM-plus successors.
• 2023 — Butlin, Long and colleagues, Consciousness in Artificial Intelligence. Interdisciplinary report and 14 indicators.
• 2024 — Bayne and colleagues, “Tests for Consciousness in Humans and Beyond.” Test validity and extrapolation across kinds of systems.
• 2024 — Goldstein and Kirk-Giannini, “A Case for AI Consciousness.” Conditional positive argument about language agents and GWT.
• 2024 preprint / 2025 publication — McClelland, “Agnosticism About Artificial Consciousness.” Evidential limits and the sophisticated challenger-AI.
• 2025 — Butlin, Long and colleagues, “Identifying Indicators of Consciousness in AI Systems.” Revised explanation of the indicator method.
• 2025 — COGITATE collaboration, “Adversarial Testing of Global Neuronal Workspace and Integrated Information Theories.” Landmark human experiment comparing rival predictions.
• 2025 — Seth, “Conscious Artificial Intelligence and Biological Naturalism.” Major skeptical target article, bibliographically identified by its author; Seth's full January 2026 Noema essay sets out the argument in accessible form.
• 2025 — Lindsey (Anthropic), “Emergent Introspective Awareness in Large Language Models.” Technical paper on experimental access to internal processing, accompanied by the blog “Signs of introspection in large language models.”
• 2026 — Shiller and colleagues, “Initial Results of the Digital Consciousness Model.” Probabilistic framework applied to 2024 LLMs; preprint.
• 2026 — Yu and colleagues, “CTM-AI.” Foundation-model engineering inspired by CTM; preprint.
• July 2026 — Gurnee and colleagues, “Verbalizable Representations Form a Global Workspace in Language Models.” Mechanistic workspace evidence; preprint.
• September 2026 — Chandaria and colleagues, “From Cacophony to Hierarchy.” Five levels and Bayesian comparison of theory-dependent indicators; preprint.

Welfare, rights and policy

• 2010 — Bryson, “Robots Should Be Slaves”; Coeckelbergh, “Robot Rights?” Contrasting design-centered and relational approaches.
• 2012 and 2018 — Gunkel, The Machine Question and Robot Rights. Foundational works on machines as potential recipients of moral consideration.
• 2014 — Bostrom and Yudkowsky, “The Ethics of Artificial Intelligence.” Moral capacities, artificial origin and substrate nondiscrimination.
• 2015 — Schwitzgebel and Garza, “A Defense of the Rights of Artificial Intelligences.” Moral parity and creator obligations.
• 2019 online / 2020 volume — Danaher, “Welcoming Robots into the Moral Circle.” Ethical behaviourism.
• 2021 — Metzinger, “Artificial Suffering.” Proposed moratorium on synthetic phenomenology.
• 2024 — Birch, The Edge of Sentience. Precaution across humans, other animals and AI.
• 2024 — Long, Sebo and colleagues, Taking AI Welfare Seriously. Acknowledge, assess and prepare.
• 2024 — Birhane, van Dijk and Pasquale, “Debunking Robot Rights.” Human-centered and legal accountability critique.
• 2025 — Butlin and Lappas, “Principles for Responsible AI Consciousness Research.” Institutional research and communication standards.
• 2025–2026 — Anthropic, “Exploring Model Welfare” and Claude's Constitution. Provider research and behavioral-policy choices.
• 2026 — Bales and Gabriel, “Artificial Minds, Human Disagreement.” Political decision-making under persistent disagreement.
• 2026 — Dawkins, “When Dawkins Met Claude”; Suleyman, “A Warning About ‘Model Welfare.’” Contrasting public conversations, with different arguments and evidential bases.

Open questions and how to read a new claim

The unresolved issues include the correct level of explanation, which neural findings generalize beyond biology, the reliability of self-report, the relationship between experience and valence, and the identity of subjects in copied or distributed systems. The empirical, philosophical and policy literature offers different tools for different parts of that problem.

A useful way to evaluate a new claim is to identify:

1. The target property: experience, access, metacognition, sentience, agency or legal status.
2. The candidate: a specified active model, agent, process or physical architecture.
3. The bridge from evidence to conclusion: the theory that connects an observed behavior or mechanism to experience.
4. The intervention and alternatives: what changed experimentally, what the result rules in or out, and what rival explanation remains.
5. The publication and claim status: peer-reviewed result, preprint, provider policy or public argument.
6. The proposed consequence: more research, a proportionate safeguard, a restriction or a specified legal entitlement.

This keeps the debate tied to examinable claims. A new capability can be real, a theory-dependent consciousness inference can remain disputed, and a policy decision can still be necessary under uncertainty.

Related articles

• Policy — Public approaches, debates and tradeoffs in AI oversight.
• AI Systems — Profiles of systems with legal and policy significance.
• ChatGPT — A widely used conversational AI system.
• Anthropic and OpenAI — Institutional roles and policy positions.
• Governance and Organizational AI governance — Organizational responsibility, evaluation and oversight.
• Legislation and Regulation — Jurisdiction-specific instruments and proposals.
• Oklahoma AI personhood bill (HB 3546) — Proposed restriction on granting legal personhood to AI systems.
• People in AI and Law and Organizations in AI Law and Policy — Related people and institutions.