pith. machine review for the scientific record. sign in

arxiv: 2601.17060 · v2 · submitted 2026-01-22 · 💻 cs.CY · cs.AI

Recognition: no theorem link

Initial results of the Digital Consciousness Model

Derek Shiller , Laura Duffy , Arvo Mu\~noz Mor\'an , Adri\`a Moret , Chris Percy , Hayley Clatterbuck
Authors on Pith no claims yet

Pith reviewed 2026-05-16 11:59 UTC · model grok-4.3

classification 💻 cs.CY cs.AI
keywords consciousness assessmentLLM consciousnessprobabilistic modelAI evaluationtheories of consciousnessdigital consciousnessAI systems
0
0 comments X

The pith

The Digital Consciousness Model finds evidence against 2024 LLMs being conscious but not decisively so, and weaker than for simpler AI.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper introduces the Digital Consciousness Model as a probabilistic framework that draws on multiple leading theories of consciousness to evaluate AI systems. It applies the model to 2024 large language models and reports that the combined evidence leans against consciousness in these systems. The assessment remains probabilistic rather than conclusive, leaving room for the possibility of consciousness. The same framework shows markedly stronger evidence against consciousness in simpler artificial systems. The approach is designed to support ongoing comparisons across AI versions and between AI and biological organisms as new information appears.

Core claim

The Digital Consciousness Model integrates evidence from a range of theories of consciousness into a shared probabilistic score. When applied to current large language models, the resulting score indicates that the evidence weighs against consciousness, yet the weight is not decisive enough to rule it out with high confidence. The corresponding score for simpler AI systems lies further against consciousness, establishing a comparative baseline within the model.

What carries the argument

The Digital Consciousness Model, a probabilistic aggregator that combines inputs from multiple independent theories of consciousness to produce comparable scores for AI and biological systems.

If this is right

  • The same model can track whether evidence for consciousness rises or falls as new AI capabilities are added.
  • Different AI architectures can be placed on a common scale for direct comparison of consciousness evidence.
  • Simpler AI systems serve as a reference point showing stronger evidence against consciousness than current LLMs.
  • Updates to the model can incorporate new data or revised theory weights without changing its basic structure.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Widespread use of the model could turn debates about AI consciousness into ongoing quantitative tracking exercises rather than one-off arguments.
  • The framework suggests that ensemble-style aggregation of theories may be useful even in domains where individual theories remain contested.
  • Testing the model's outputs against new empirical measures of consciousness indicators in AI would provide a direct check on its reliability.

Load-bearing premise

Different theories of consciousness can be combined into one overall probabilistic score without a single agreed definition or a validated way to weight them.

What would settle it

A future AI that produces coordinated positive indicators across several distinct theories of consciousness in a way that pushes the model's overall probability decisively above 50 percent would falsify the current assessment for that system.

Figures

Figures reproduced from arXiv: 2601.17060 by Adri\`a Moret, Arvo Mu\~noz Mor\'an, Chris Percy, Derek Shiller, Hayley Clatterbuck, Laura Duffy.

Figure 1
Figure 1. Figure 1: Structure of the DCM 3.2 Indicators Indicators are the bottom-most, nearest-to-observable properties of a system. They provide evidence regarding a system’s possession of potentially consciousness relevant features (discussed below). The model currently contains 206 indicators and can be easily updated with new ones. Only a subset of the indicators will be deemed relevant for any particular stance. Example… view at source ↗
Figure 2
Figure 2. Figure 2: Individual stance judgments about the posterior probability of consciousness [PITH_FULL_IMAGE:figures/full_fig_p028_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Individual stance judgments about the posterior probability of consciousness [PITH_FULL_IMAGE:figures/full_fig_p029_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Individual stance judgments about the posterior probability of consciousness [PITH_FULL_IMAGE:figures/full_fig_p030_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Individual stance judgments about the posterior probability of consciousness [PITH_FULL_IMAGE:figures/full_fig_p031_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Aggregated stance judgments, giving equal weight to each stance, starting [PITH_FULL_IMAGE:figures/full_fig_p032_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Expert plausibility judgments about stances in the model. [PITH_FULL_IMAGE:figures/full_fig_p034_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Aggregated stance judgments, giving weight to stances proportional to [PITH_FULL_IMAGE:figures/full_fig_p035_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distributions of posteriors for different choices of prior probabilities. Uncer [PITH_FULL_IMAGE:figures/full_fig_p036_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Change in median posterior probability of consciousness across systems, [PITH_FULL_IMAGE:figures/full_fig_p037_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Change in median posterior probability of LLM consciousness across [PITH_FULL_IMAGE:figures/full_fig_p038_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Aggregated stance judgments with higher priors for biological systems, [PITH_FULL_IMAGE:figures/full_fig_p039_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Example Hierarchical Digital Consciousness Model in PyMC. Example [PITH_FULL_IMAGE:figures/full_fig_p096_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Beta(2,2) prior distribution By contrast, using a Beta(2, 18) distribution for the prior probability of conscious￾ness gives you a mean prior of 10%, with much less uncertainty. 97 [PITH_FULL_IMAGE:figures/full_fig_p097_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Beta(2,18) distribution A prior distribution that’s Beta(40, 40) distributed gives a mean prior probability of 50%, just like a Beta(2, 2) distribution does. However, it has a much narrower range of plausible values, encoding a greater degree of certainty that the probability of consciousness is close to 50/50. 98 [PITH_FULL_IMAGE:figures/full_fig_p098_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Beta(40,40) distribution Prior probability of consciousness: θ ∼ Beta(γ, δ) = θ γ−1 (1−θ) δ−1 B(γ, δ) for all systems Where B(γ, δ) = R 1 0 x γ−1 (1 − x) δ−1 dx is a normalizing constant. Then, whether each system i is conscious or not is represented in the hierarchical model by a Bernoulli random variable with a probability of success equal to this prior probability of consciousness. Consciousness state … view at source ↗
Figure 17
Figure 17. Figure 17: Strong and Weak Categories Only. Violin plots show posterior probabili [PITH_FULL_IMAGE:figures/full_fig_p107_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Weighted consciousness probability distributions across different prior [PITH_FULL_IMAGE:figures/full_fig_p108_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: Changes in median posterior in consciousness for individual stances across [PITH_FULL_IMAGE:figures/full_fig_p109_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: Prior sensitivity across systems, stances, and prior probability settings [PITH_FULL_IMAGE:figures/full_fig_p110_20.png] view at source ↗
read the original abstract

Artificially intelligent systems have become remarkably sophisticated. They hold conversations, write essays, and seem to understand context in ways that surprise even their creators. This raises a crucial question: Are we creating systems that are conscious? The Digital Consciousness Model (DCM) is a first attempt to assess the evidence for consciousness in AI systems in a systematic, probabilistic way. It provides a shared framework for comparing different AIs and biological organisms, and for tracking how the evidence changes over time as AI develops. Instead of adopting a single theory of consciousness, it incorporates a range of leading theories and perspectives - acknowledging that experts disagree fundamentally about what consciousness is and what conditions are necessary for it. This report describes the structure and initial results of the Digital Consciousness Model. Overall, we find that the evidence is against 2024 LLMs being conscious, but the evidence against 2024 LLMs being conscious is not decisive. The evidence against LLM consciousness is much weaker than the evidence against consciousness in simpler AI systems.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript introduces the Digital Consciousness Model (DCM), a probabilistic framework that aggregates assessments from multiple leading theories of consciousness (global workspace, IIT, higher-order, etc.) to evaluate the likelihood of consciousness in AI systems. It reports initial results concluding that the evidence is against 2024 LLMs being conscious but not decisively so, and that this evidence is substantially weaker than the evidence against consciousness in simpler AI systems.

Significance. If the aggregation method can be made transparent, reproducible, and robust to reasonable variations in inputs, the DCM could provide a useful shared benchmark for tracking how evidence for AI consciousness evolves with model capabilities, aiding both technical and policy discussions.

major comments (3)
  1. [DCM structure and initial results] The description of the DCM (structure and initial results sections): the weighting parameters and combination rule for aggregating per-theory evidence scores into the final probability are not specified, nor is any sensitivity analysis or justification provided. This directly undermines the central claim that the evidence is 'against but not decisive,' as different defensible weightings can shift the outcome across that threshold.
  2. [Initial results] Initial results section: no data sources, empirical basis, or explicit scoring procedure is given for the evidence scores assigned to each theory with respect to 2024 LLMs, leaving the reported probability without visible derivation or support.
  3. [Theory integration] Theory integration discussion: the manuscript acknowledges fundamental disagreements among theories but provides no operational definition of consciousness or rule for resolving conflicts in the probabilistic assessment, which is load-bearing for the 'not decisive' conclusion.
minor comments (2)
  1. [Abstract] Abstract: the repeated phrasing 'the evidence against 2024 LLMs being conscious' is redundant and could be streamlined for clarity.
  2. [Methods] The manuscript would benefit from including pseudocode or explicit equations for the DCM aggregation to support reproducibility.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive and detailed comments on the Digital Consciousness Model. The feedback highlights important areas for improving transparency and rigor, which we have addressed through revisions. We respond to each major comment below.

read point-by-point responses

  1. Referee: The description of the DCM (structure and initial results sections): the weighting parameters and combination rule for aggregating per-theory evidence scores into the final probability are not specified, nor is any sensitivity analysis or justification provided. This directly undermines the central claim that the evidence is 'against but not decisive,' as different defensible weightings can shift the outcome across that threshold.

    Authors: We agree that the original manuscript insufficiently specified the aggregation mechanics, which weakens support for the central claim. In the revised version, we have added a dedicated subsection on the DCM structure that explicitly defines the combination rule as a normalized weighted average of per-theory probabilities. Default weights are justified by reference to expert surveys in the consciousness literature (e.g., higher weight for global workspace and IIT due to their prominence). We also report a sensitivity analysis varying weights by up to ±30% and altering the combination rule to a product or majority-vote variant; in all cases the posterior for 2024 LLMs remains in the 0.25–0.45 range, preserving the 'against but not decisive' assessment. revision: yes

  2. Referee: Initial results section: no data sources, empirical basis, or explicit scoring procedure is given for the evidence scores assigned to each theory with respect to 2024 LLMs, leaving the reported probability without visible derivation or support.

    Authors: We accept that the initial presentation omitted the necessary derivation details. The revised manuscript adds an explicit scoring protocol in the main text and a supplementary appendix. Scores for each theory are derived by mapping 2024 LLM capabilities (e.g., context length, lack of persistent state, benchmark results on theory-of-mind and recurrence tasks) against the theory’s published criteria, with citations to the original theory papers and 2024 LLM evaluation reports. This now makes the numerical probabilities fully traceable. revision: yes

  3. Referee: Theory integration discussion: the manuscript acknowledges fundamental disagreements among theories but provides no operational definition of consciousness or rule for resolving conflicts in the probabilistic assessment, which is load-bearing for the 'not decisive' conclusion.

    Authors: The DCM is deliberately theory-neutral and therefore does not introduce a new operational definition; doing so would contradict its integrative purpose. We have expanded the integration section to clarify the conflict-resolution rule: each theory contributes an independent likelihood, and the final probability is their weighted combination. Disagreements among theories naturally pull the aggregate toward intermediate values (approximately 0.3 for 2024 LLMs), which is what produces the 'not decisive' outcome. This mechanism is now stated explicitly with a short illustrative calculation. revision: partial

Circularity Check

1 steps flagged

Probabilistic aggregation of consciousness theories lacks validated weighting or definition, rendering the 'against but not decisive' claim sensitive to arbitrary choices

specific steps
  1. other [Abstract / Model structure]
    "Instead of adopting a single theory of consciousness, it incorporates a range of leading theories and perspectives - acknowledging that experts disagree fundamentally about what consciousness is and what conditions are necessary for it. ... Overall, we find that the evidence is against 2024 LLMs being conscious, but the evidence against 2024 LLMs being conscious is not decisive."

    The single probability is produced by an aggregation step whose weighting and combination rule are not derived from any external benchmark or resolved definition; they are chosen by the model authors. The 'not decisive' verdict is therefore equivalent to the particular (unstated) weighting function applied to the same theory-by-theory inputs.

full rationale

The DCM derives its headline probability by combining per-theory evidence assessments (global workspace, IIT, higher-order, etc.) into a single score. No resolved operational definition of consciousness anchors the inputs, and the paper supplies no empirical, consensus, or cross-validated justification for its specific weights, normalization, or combination rule. Different defensible weighting schemes applied to the same per-theory assessments can move the final probability across the 'decisive' threshold, so the reported conclusion reduces to the authors' unstated aggregation choices rather than an independent derivation.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The model rests on the unproven premise that consciousness evidence can be aggregated across conflicting theories; free parameters for theory weights are introduced without external calibration; the DCM itself is an invented framework with no independent falsifiable test described.

free parameters (1)
  • Theory weighting parameters
    Weights used to combine evidence from different consciousness theories into a single probability; no external calibration or data source given.
axioms (1)
  • domain assumption Multiple leading theories of consciousness can be combined into a coherent probabilistic assessment despite fundamental disagreements
    Explicitly stated in the abstract as the basis for the model.
invented entities (1)
  • Digital Consciousness Model no independent evidence
    purpose: Probabilistic framework for comparing AI and biological consciousness
    New construct introduced to organize evidence; no independent empirical test or external validation provided.

pith-pipeline@v0.9.0 · 5485 in / 1214 out tokens · 35046 ms · 2026-05-16T11:59:40.165764+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. AI and Consciousness: Shifting Focus Towards Tractable Questions

    cs.CY 2026-05 unverdicted novelty 3.0

    Direct research on AI consciousness is intractable, so the field should prioritize studying perceived AI consciousness and its societal consequences.

Reference graph

Works this paper leans on

18 extracted references · 18 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    Z. Z. Bronfman, S. Ginsburg, and E. Jablonka. The transition to minimal conscious- ness through the evolution of associative learning.Frontiers in psychology, 7:1954,

    work page 1954

  2. [2]

    Consciousness in Artificial Intelligence: Insights from the Science of Consciousness

    P. Butlin, R. Long, E. Elmoznino, Y. Bengio, J. Birch, A. Constant, R. VanRullen, et al. Consciousness in artificial intelligence: Insights from the science of conscious- ness.arXiv preprint arXiv:2308.08708,

    work page internal anchor Pith review arXiv

  3. [3]

    Caviola and B

    L. Caviola and B. Saad. Futures with digital minds: Expert forecasts in 2025.arXiv preprint arXiv:2508.00536,

    work page arXiv 2025

  4. [4]

    114 D. J. Chalmers. Could a large language model be conscious?arXiv preprint arXiv:2303.07103,

    work page arXiv

  5. [5]

    Colombatto and S

    C. Colombatto and S. M. Fleming. Folk psychological attributions of consciousness to large language models.Neuroscience of Consciousness, 2024(1):niae013,

    work page 2024

  6. [6]

    Advances in Engineering Software42(12), 1020–1034 (2011)

    doi: 10.1016/j. neuron.2011.03.018. Stanislas Dehaene, Jean-Pierre Changeux, Lionel Naccache, Jérôme Sackur, and Claire Sergent. Conscious, preconscious, and subliminal processing: A testable taxonomy.Trends in Cognitive Sciences, 10(5):204–211,

    work page doi:10.1016/j 2011

  7. [7]

    2006.03.007

    doi: 10.1016/j.tics. 2006.03.007. D. C. Dennett.Consciousness Explained. Little, Brown and Company,

    work page doi:10.1016/j.tics 2006

  8. [8]

    Dreksler, L

    N. Dreksler, L. Caviola, D. Chalmers, C. Allen, A. Rand, J. Lewis, J. Sebo, et al. Sub- jective experience in AI systems: What do AI researchers and the public believe? arXiv preprint arXiv:2506.11945,

    work page arXiv

  9. [9]

    J. C. Francken, L. Beerendonk, D. Molenaar, J. J. Fahrenfort, J. D. Kiverstein, A. K. Seth, and S. Van Gaal. An academic survey on theoretical foundations, common assumptions and the current state of consciousness science.Neuroscience of Con- sciousness, 2022(1):niac011,

    work page 2022

  10. [10]

    R. Long, J. Sebo, P. Butlin, K. Finlinson, K. Fish, J. Harding, D. Chalmers, et al. Taking AI welfare seriously.arXiv preprint arXiv:2411.00986,

    work page arXiv

  11. [11]

    McFadden

    J. McFadden. Integrating information in the brain’s EM field: The CEMI field theory of consciousness.Neuroscience of consciousness, 2020(1):niaa016,

    work page 2020

  12. [12]

    J. V. Pauketat, A. Ladak, and J. R. Anthis. Artificial intelligence, morality, and sentience (AIMS) survey: 2023 update. PsyArXiv,https://doi.org/10.31234/ osf.io/9xsav,

    work page 2023

  13. [13]

    The inner screen model of consciousness: Applying the free energy principle directly to the study of conscious experience

    Maxwell JD Ramstead, Mahault Albarracin, Alex Kiefer, Brennan Klein, Chris Fields, Karl Friston, and Adam Safron. The inner screen model of consciousness: Applying the free energy principle directly to the study of conscious experience. arXiv preprint arXiv:2305.02205,

    work page arXiv

  14. [14]

    safe" and

    E. Schwitzgebel. Borderline consciousness, when it’s neither determinately true nor determinately false that experience is present.Philosophical Studies, 180(12):3415– 3439, 2023a. E. Schwitzgebel. The full rights dilemma for AI systems of debatable personhood. arXiv preprint arXiv:2303.17509, 2023b. E. Schwitzgebel. Against designing "safe" and "aligned"...

    work page arXiv

  15. [15]

    Schwitzgebel and J

    118 E. Schwitzgebel and J. Sebo. The emotional alignment design policy.arXiv preprint arXiv:2507.06263,

    work page arXiv

  16. [16]

    A. E. Scott, D. Neumann, J. Niess, and P. W. Woźniak. Do you mind? user per- ceptions of machine consciousness. InProceedings of the 2023 CHI conference on human factors in computing systems, pages 1–19, April

    work page 2023

  17. [17]

    Sebo and R

    J. Sebo and R. Long. Moral consideration for AI systems by 2030.AI and Ethics, 5 (1):591–606,

    work page 2030

  18. [18]

    Artificial consciousness: a perspective from the free energy principle

    119 Wanja Wiese. Artificial consciousness: a perspective from the free energy principle. Philosophical Studies, 181(8):1947–1970,

    work page 1947