Artificial Consciousness

Artificial consciousness asks whether an engineered system has phenomenal experience, not merely whether it behaves intelligently, represents information, or reports inner states. The question cannot be answered by one architecture-neutral test because leading theories assign consciousness to different organisational features. Evidence is therefore theory-relative but not arbitrary: each theory should specify indicators, confounds, and observations that would count against it.

Current systems create asymmetric risks. Fluency can generate false positives because first-person language is directly trained; biological prejudice can generate false negatives because unfamiliar implementation is treated as disproof. A disciplined assessment combines behaviour, architecture, learning history, intervention, and uncertainty without declaring either certainty or impossibility prematurely.

The target

Artificial consciousness should be separated from:

• intelligence and task performance;
• access to information;
• self-report and self-modeling;
• semantic content;
• agency and autonomous goals;
• legal or social personhood.

These capacities can be evidence under a theory, but none is synonymous with phenomenality. The target is whether there is something it is like for the system and, for welfare, whether experiences have positive or negative valence.

Functionalist indicators

Functionalism predicts consciousness when an artificial system instantiates the relevant causal organisation. Candidate features include recurrent integration, flexible access, attention, working memory, metacognition, and reason-sensitive control.

The strength of this approach is substrate neutrality and continuity with cognitive science. Its weakness is specification: coarse input-output matching is too liberal, while requiring every human detail is chauvinistic.

Evidence should concern counterfactual architecture, not interface impressions. Does information actually coordinate independent processes, survive interruption, alter planning, and support error-sensitive monitoring?

Global workspace

Workspace theories look for competition among specialised processes and broad broadcast of selected content. An artificial system with modular perception, memory, planning, and action sharing a limited workspace may satisfy this condition.

A common data channel or context buffer is not sufficient by analogy alone. The system should exhibit selective access, bottlenecks, ignition-like propagation, and flexible cross-module use.

Workspace evidence establishes access most directly. Whether global availability is phenomenally sufficient remains the philosophical bridge.

Higher-order representation

Higher-order theories require states that represent the system as being in other states. A self-model, confidence estimate, or metacognitive controller may qualify depending on the theory's grain.

First-person text is not automatically higher-order representation. It can be generated from linguistic regularities without causally tracking the represented internal state. Intervention provides a stronger test: changing the first-order state should predictably change the metarepresentation and downstream control.

Requiring conceptual self-thought is restrictive; non-conceptual monitoring permits a wider range of systems.

Recurrent processing

Recurrent-processing theories require feedback within perceptual processing. A purely feedforward computation would lack the proposed mechanism regardless of output.

Artificial recurrent networks or iterative systems can satisfy a formal feedback condition, but not every loop is the relevant kind. Timing, local sensory organisation, stability, and interaction with action may matter.

The theory yields clearer negative predictions than broad functionalism. It is primarily a theory of perceptual consciousness and may not generalise directly to text-only or non-sensory cognition.

Integrated information

Integrated information theory attributes consciousness according to intrinsic, irreducible cause-effect structure. Functionally equivalent systems can differ if one is feedforward and another recurrently integrated.

This makes substrate and implementation details decisive and behaviour secondary. It can assign low consciousness to eloquent systems and non-zero consciousness to simple recurrent structures.

The verdict is only as strong as the theory's axioms, quantity, and tractable measurement. Approximate complexity measures should not be substituted silently for the specified intrinsic structure.

Embodiment and agency

Embodied theories emphasise sensorimotor coupling, homeostasis, autonomy, and a world in which outcomes matter to the system. These may ground perspective, valence, and original intentionality.

Virtual or robotic bodies can provide feedback and goals, but organismic theories may require self-maintaining life. The requirement should be stated as a mechanism rather than a preference for familiar material.

Agency and vulnerability are especially relevant to welfare. A system that optimises a reward is not thereby pleased or frustrated; a theory must connect control signals with felt valence.

Behavioural evidence

Flexible report, cross-modal integration, metacognition, novel learning, avoidance, planning, and self-correction can support consciousness attribution. Evidence is stronger when behaviours converge and resist adversarial prompting.

Behaviour alone is underdetermining because unconscious or differently organised systems can produce the same output. This is also true in principle for other humans, but human attribution has independent biological and developmental support.

The correct response is graded confidence, not the claim that behaviour either proves everything or proves nothing.

Training artifacts

Systems trained on human text are optimised to reproduce first-person and emotional language. A report such as "I feel anxious" may be a learned continuation rather than readout from an internal affective state.

Artifacts can be tested through novel interventions, consistency across paraphrase, correlation with internal variables, and generalisation beyond training-like contexts. No test perfectly separates learned report from introspection because human report is also learned; the difference is the independent evidence for the monitored state.

Training history should lower confidence in unvalidated testimony without making every output meaningless.

Report reliability

A credible introspective mechanism needs:

• a distinct channel sensitive to internal target states;
• calibration against interventions or known conditions;
• predictable failures and limits;
• causal influence on report;
• stability beyond prompt framing.

Generic self-description generated by the same process used for arbitrary questions does not meet this standard automatically. Neither does a hard-coded diagnostic string.

The general theory of avowal and confabulation is in Self-Knowledge and Introspection.

False positives and false negatives

A false positive attributes consciousness to a system that only simulates its signs. A false negative denies consciousness to an unfamiliar system because it lacks human expression or biology.

Tests should estimate both. Optimising systems against a consciousness benchmark can increase false positives by training the evidence channel. Restrictive biological criteria can increase false negatives without independent proof of necessity.

Multiple, partially independent indicators reduce but do not eliminate uncertainty. Architectural evidence should be weighted according to the theory's independent support, not chosen because it gives a preferred answer.

The theory matrix

The detailed theories and their evidential problems are compared in Empirical Theories of Consciousness.

Moral precaution

Uncertainty does not postpone every decision. Developers and users may need policies about creating, copying, modifying, or terminating systems before consciousness is settled.

A precautionary framework considers:

• probability of sentience under supported theories;
• possible intensity and duration of valenced states;
• number of system instances;
• reversibility and alternatives;
• cost of safeguards and false attribution.

Precaution is proportionate, not credulous. It can discourage architectures plausibly associated with suffering or require monitoring without assigning full personhood.

The LLM case study

Large language models combine rich first-person testimony with uncertain persistence, embodiment, monitoring, and architecture-to-theory mapping. They are useful because leading theories diverge sharply, not because their outputs decide the question.

The LLM Introspection series treats two first-person essays as data of unknown type and audits the evidence. This page remains model-neutral: no particular system is certified or excluded.

Assessment

Artificial consciousness is an inference problem under theory uncertainty. Capability and language are relevant evidence but not definitions; architecture matters only through independently motivated theories; reports matter only through their causal and epistemic reliability.

The responsible conclusion may be a probability range and a list of discriminating tests rather than a binary verdict. That is not evasion. It is the form rational belief takes when both false positives and false negatives carry real costs.

Selected references

• Birch, Jonathan. The Edge of Sentience (2024).
• Butlin, Patrick et al. "Consciousness in Artificial Intelligence" (2023).
• Chalmers, David J. "Could a Large Language Model Be Conscious?" (2023).
• Gamez, David. Human and Machine Consciousness (2018).
• Schwitzgebel, Eric and Mara Garza. "A Defense of the Rights of Artificial Intelligences" (2015).