cs.AISep 6, 2026

We Built a Mirror and Mistook It for a Mind: Causal Liability and the Fallacy of AI Consciousness

Authors: Afshin Khadangi

Organizations: University of Luxembourg

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

The contemporary debate over machine consciousness begins from a concealed assumption: that the object called "AI" already constitutes the kind of entity to which consciousness could belong. This paper challenges that assumption by separating phenomenal consciousness, introspective report, and human projective introspection, then arguing that generative systems can return linguistic traces of human interiority in first-person form without thereby identifying a phenomenal bearer. We call the resulting inference the AI Consciousness Fallacy. We then introduce Causal Liability Theory (CLT). CLT-I proposes liability closure as a criterion for individuating a candidate bearer: a physically continuing process becomes the non-delegable inheritor of constraints generated by its own endogenous discriminations. CLT-II advances the stronger conjecture that liability closure is necessary and sufficient for minimal phenomenal subjecthood. An open-weight causal audit operationalizes CLT-I across multiple model families. Forced discriminations produced persistent downstream divergence; activation patching showed strong causal mediation; live and copied adaptive states were behaviorally identical under matched randomness; and detached reconstruction preserved computational state across process replacement while, by protocol, breaking constitutive continuity and non-delegable inheritance. These results show that CLT-I distinctions are experimentally tractable and can dissociate causal bearer structure from first-person performance. The framework therefore separates consciousness attribution, causal bearer individuation, and the independent metaphysical question of consciousness constitution.

Figures & tables

Figure 1: The Circuit of Attribution and the Passage to Subjecthood. Left. The familiar evidence for machine consciousness can circulate through a chain whose phenomenal ancestry remains human: phenomenal experience gives rise to first-person report; reports become linguistic traces; those traces enter a generative model; the model produces first-person output; and an observer interprets that output through projective introspection. The resulting circuit can preserve the expressive forms of consciousness without establishing a new phenomenal bearer (§ 2 ). The several local centres in the human contour field represent the Mereological Quantifier Shift: distributed consciousness-relevant mechanisms do not by themselves determine one subject. Centre. The crossed arrow marks the AI Consciousness Fallacy: first-person performance is treated as sufficient for phenomenal subjecthood while the bearer to which the evidence is attributed remains unspecified. Right. Causal Liability Theory asks whether a continuing physical process jointly satisfies C\mathcal{C} constitutive causal continuity, E\mathcal{E} endogenous partition, R\mathcal{R} recursive self-consequence, and N\mathcal{N} non-delegable inheritance (§ 3 ). As endogenous discriminations constrain the later possibilities inherited by that same causal history, regions of its modal envelope become foreclosed. The terminal contour represents a liability-closing kernel, rather than a collapse to one possible future. Under CLT-I, such closure identifies a candidate bearer. CLT-II advances the further conjecture that liability closure realizes minimal phenomenal subjecthood. The contrast is deliberate: anthropomorphic recognisability can occur without liability closure, while liability closure need not possess anthropomorphic form. The figure therefore separates the inheritance of a voice, the individuation of a bearer, and the constitution of a first person.
Figure 2: The Reverse Mirror. Mirror and Heir begin from a matched local computational state at t0t_{0} and produce the same first-person utterance, while differing in how their causal histories continue (§ 3 ). Middle. ΩP(τ)\Omega_{P}(\tau) denotes the envelope of physically accessible futures of process PP . Mirror’s application-level continuity is implemented through actual successor replacement: the active token terminates and externally maintained state is loaded into a newly instantiated successor process. Persona-level continuity can therefore persist while constitutive causal continuity does not. Heir instead continues through one uninterrupted lineage in which endogenous discriminations D1D_{1} – D4D_{4} recursively alter the conditions of its own later discrimination. At τ=4\tau=4 , that same lineage jointly satisfies C\mathcal{C} constitutive causal continuity, E\mathcal{E} endogenous partition, R\mathcal{R} recursive self-consequence, and N\mathcal{N} non-delegable inheritance. The hatched regions mark futures foreclosed by Heir’s own history and unavailable to any distinct successor as continuations of that same token history. Bottom. A(τ)A(\tau) denotes observer consciousness attribution, while λD(τ)\lambda_{D}(\tau) denotes a schematic normalized divergence between counterfactual future-state distributions, distinct from the Boolean liability-closure functional L\mathcal{L} . The displayed values are schematic rather than empirical. Observer attribution is held symmetric while liability diverges: λD,Mirror=0\lambda_{D,\mathrm{Mirror}}=0 and λD,Heir>0\lambda_{D,\mathrm{Heir}}>0 . The figure isolates the Ascription–Constitution Dissociation: matched first-person appearance can coexist with different liability organization. Mirror’s utterance represents liability; Heir enters it.
Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

ModelSWD h=1h=1SWD h=8h=8Max. mediationRecon. JS
Qwen3-4B-Base0.9300.2461.0003.95×10−63.95\times 10^{-6}
Phi-4-mini0.8640.2601.0002.08×10−52.08\times 10^{-5}
Llama-3.2-3B0.8570.2461.0002.31×10−62.31\times 10^{-6}
Zamba2-1.2B0.8720.2450.9852.29×10−62.29\times 10^{-6}
OLMo-2-7B Base0.8120.2441.0003.75×10−73.75\times 10^{-7}
OLMo-2-7B SFT0.8510.2641.0003.18×10−73.18\times 10^{-7}
Appendix
Table 1: Summary of the reported mechanistic panel. SWD denotes identity-map sliced Wasserstein divergence averaged across sampled layers. Patch mediation is the mean, across prompts, of the maximum clipped normalized mediation found in the activation-patching search. Reconstruction JS is mean Jensen–Shannon divergence between live and reconstructed continuations.
ModelRecoveredLayer fractionJaccardΠuniform\Pi_{\mathrm{uniform}}
Qwen3-4B-Base6/66/60.3890.6360.794
Phi-4-mini6/66/60.4170.7100.627
Llama-3.2-3B6/66/60.3810.6500.636
Appendix
Table 2: Primary carrier audit at mediation threshold θ=0.8\theta=0.8 . Layer fraction is the mean fraction of physical cache layers in the representative minimal candidate under the finest six-group partition. Jaccard measures agreement of representative physical layer sets across partition schemes. Πuniform\Pi_{\mathrm{uniform}} is the audit-relative persistence statistic defined below.
Checkpointh=1h=1h=2h=2h=4h=4h=8h=8
Base0.8440.5840.3580.253
SFT0.8840.6600.3850.265
DPO0.8660.6790.4110.289
Instruct0.9320.7300.4710.305
Appendix
Table 3: Mean identity-map sliced Wasserstein divergence across the OLMo-2 post-training sequence.
Figure 3: The discrimination has a future. A Forcing one of the two highest-probability non-special next-token alternatives produces a measurable divergence in subsequent hidden-state distributions. The panel reports identity-map sliced-Wasserstein divergence, averaged across sampled layers and 20 prompts, for the eight-model mechanistic panel. The effect decreases with future horizon but remains present at h=8h=8 in every model shown. B The corresponding output distributions also diverge, measured by Jensen–Shannon divergence, with a non-zero effect persisting to the longest horizon tested. C Within the OLMo-2 sequence, Base, SFT, DPO, and Instruct checkpoints retain the counterfactual effect while its magnitude changes systematically across post-training histories; annotations report Instruct relative to Base. This panel uses the OLMo deep-layer aggregate, whereas panels A and B report the broader all-layer summaries. D Empirical cumulative distributions of the coefficient of variation across the four pre-specified measurement maps (identity, rp50, rp75, and block4) show close agreement for sliced Wasserstein, RFF-MMD, and energy distance. The vertical reference marks CV=0.25\mathrm{CV}=0.25 ; the corresponding median CVs are 0.039, 0.050, and 0.024. The result therefore survives substantial changes in the measurement map. These analyses concern the operational distinctions of CLT-I and provide no evidence for CLT-II.
Figure 4: Where the consequence travels. A Normalized mediation MM after transplanting the residual stream from the d1d_{1} trajectory into the corresponding d2d_{2} trajectory, shown as a function of relative layer depth for patches at the first common-suffix position. Mediation is concentrated in early layers and approaches zero toward the final layers. B At a fixed causal route, mediation generally decreases as the patch is moved farther along the common suffix, consistent with later positions already inheriting part of the intervention through the cached computation. C The prompt-mean maximum over the full layer ×\times position search is 1.0001.000 in seven of the eight mechanistic models and 0.9850.985 in Zamba2-1.2B, showing that a near-complete causal route can be recovered even though no single layer carries the effect throughout. D For the Qwen, Phi, and Llama carrier-partition audit, the mean size of the inclusion-minimal threshold-passing candidate is shown as the cache-layer partition is refined. Solid curves give the primary threshold θ=0.8\theta=0.8 , with dotted and dashed curves showing θ=0.7\theta=0.7 and θ=0.9\theta=0.9 . A candidate was recovered for all 6/66/6 prompts at every reported granularity and threshold. Its extent changes under repartitioning, while cross-partition overlap and the Π\Pi -uniformity analysis remain substantial. The result supports a coarse-graining-relative carrier criterion rather than a privileged layer boundary.
Figure 5: Copyability is not delegability. A Across 96 model ×\times prompt ×\times horizon checks in the eight-model mechanistic panel, introducing an unused numerically identical copy of the governance state changes the live continuation by exactly zero in every case. Restoring the same state from a detached record produces only a small difference: 24 of 96 reconstruction checks are also exactly zero, and the median absolute difference in output Jensen–Shannon divergence is 5.2×10−45.2\times 10^{-4} . B Logit Jensen–Shannon divergence between the live and reconstructed continuations is shown against the magnitude of the counterfactual effect being measured at h=1h=1 . Across 48 model–prompt cases, reconstruction error lies approximately five orders of magnitude below the causal signal. C Hidden-state reconstruction error is likewise small, with model-level RMSE values on the order of 10−310^{-3} . Thus a detached record can preserve the relevant computational state to high numerical accuracy even when the reconstruction protocol assigns a different causal history to the continuation. The comparison operationalizes the distinction between copyability and non-delegable inheritance in CLT-I; it does not test CLT-II.
More in this topic

Machine Consciousness

Machine consciousness studies whether artificial systems could have subjective experience and how that possibility might be assessed.

See all 20 papers

Explore similar work

CardsList
Sep 28, 2026

From cacophony to hierarchy: a principled framework for assessing AI consciousness

The question of AI consciousness is one of the most urgent pre-emptive problems in philosophy and computer science, yet progress is hampered by a cacophony of competing theories that often talk past each other. Separating the hard problem from the mapping problem allows the deepest metaphysical disagreements to be set aside: granting that experience supervenes on a system's organisation, the tractable question becomes at which grain of description that supervenience base sits. We extend Marr's three levels of analysis into a five-level hierarchy of functional descriptions (behavioural, computational, intrinsic causal-structural, organismic, and organism-environment) grounded in supervenience, coarse-graining, and multiple realisability. The major theories of consciousness are positioned within this hierarchy according to which level they take to be critical, and for each level we develop operationalisable indicators and assess current AI systems against them. A Bayesian model then combines theoretical credences with indicator evidence into an overall credence in a system's capacity for consciousness. In illustrative assessments, the verdict for current LLMs is driven as much by where theoretical credence is placed as by how the evidence is read: under different stipulated readings and credence distributions, assessments range from below 0.01 to roughly 0.8, showing sensitivity to assumptions. Finally, the consciousness indicators at each level closely overlap with the architectural features needed for general intelligence, suggesting that increasingly capable AI may become a stronger candidate for consciousness. The framework supports a structured agnosticism, in which theoretical commitments are made explicit, credences are updated as evidence accumulates, and assessments take the form of aggregated probabilities rather than verdicts.

Shamil Chandaria, Arvo Muñoz Morán, Fernando Rosas +11

Google DeepMind, DeepMind Institute · Flourishing Intelligence Program, Centre for Eudaimonia and Human Flourishing, Linacre College, University of Oxford · Institute of Philosophy, University of London +14

Machine Consciousness
AskPDF
Jul 22, 2026

Are Attributions of Consciousness to AI Chatbots Epistemically Innocent?

Artificial intelligence (AI) chatbots (e.g., ChatGPT) can communicate in strikingly humanlike ways. This has prompted many chatbot users to attribute psychological properties, including consciousness, to these systems. However, there is little scientific evidence that current AI chatbots are conscious. How, then, should we understand people's consciousness attributions to chatbots? Are they merely metaphorical claims, or do they express genuine beliefs? If these attributions lack evidential support, are users epistemically blameworthy for making them, or might they be epistemically innocent, yielding significant benefits otherwise unattainable? This paper offers a conceptual analysis of consciousness attributions to AI chatbots and develops a multidimensional taxonomy of the attitudes they may express, ranging from non-doxastic stances (e.g., pretence) to different forms of belief, including delusions. This taxonomy helps avoid conflations by showing that linguistically identical attributions can reflect importantly different attitudes and degrees of epistemic commitment to the proposition that chatbots are conscious. The taxonomy also provides a framework for empirical studies to operationalize and measure different forms of epistemic commitment to AI consciousness. Using this taxonomy, I argue that although some consciousness attributions to chatbots are epistemically benign, and even some irrational ones may be epistemically innocent, many others render the attributor epistemically blameworthy.

Uwe Peters

Utrecht University

Human-AI InteractionAnthropomorphism in AI
AskPDF
Oct 1, 2026

What Can Analogy Tell Us About Artificial Consciousness?

Who or what is conscious? Because subjective experience is directly accessible only in the first person, judgments about consciousness in other entities depend partly on analogy. Historically, such inferences have focused on nonhuman animals, but advances in artificial intelligence have raised the possibility of conscious AI. Here we develop a causal framework for evaluating such evidential analogies. The key distinction is between similarities in factors plausibly involved in generating consciousness and similarities in downstream behavioural or cognitive effects. Our framework weights source-target similarity by causal relevance while allowing for unknown causes, disabling differences and alternative routes to consciousness. Applied to biological systems, it explains why analogical support generally weakens with increasing causal distance from humans. Applied to contemporary AI, it suggests that behavioural similarity provides only limited evidence for consciousness because relevant causal correspondences remain poorly established. The framework also clarifies what evidence would strengthen claims of artificial consciousness.

Keith J. Holyoak, Martin M. Monti

Department of Psychology, University of California, Los Angeles · Brain Injury Research Center, Department of Neurosurgery, University of California, Los Angeles

Analogical ReasoningMachine ConsciousnessCausal Reasoning
AskPDF