Pith. sign in

REVIEW 4 major objections 6 minor 72 references

Analyzing Advanced AI Systems Against Definitions of Life and Consciousness

T0 review · 4 major / 6 minor · reviewed 2026-08-08 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read The paper argues that functional tests—sabotage detection and mirror self-recognition—can mark measurable thresholds where AI systems become life-like or consciousness-like, and it demonstrates both on CNNs and five chatbots.

desk verdict Every empirical demonstration in this paper is set up so the claimed behavior is guaranteed by the training labels or instructions, so the Life* threshold claims rest on confounds rather than evidence. read the letter →

arxiv 2502.05007 v1 pith:LJ2HCKU7 submitted 2025-02-07 cs.AI

Azadeh Alavi , Hossein Akhoundi , Fatemeh Kouchmeshki This is my paper · ORCID
classification cs.AI
keywords AIconsciousnessmirrorself-recognitionlife-likecriteriaself-maintenancedatapoisoningintegratedrejectionmodelfunction-basedchatbotself-identification
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that the question 'can AI be conscious or alive?' can be turned into empirical tests by replacing biology-centric definitions of life with functional criteria. It introduces the Life* framework, which maps three classical definitions of life—a dictionary definition, a space-agency definition, and the 'Seven Pillars' framework—onto measurable AI behaviors such as immune-like self-maintenance, emergent complexity, and rudimentary self-recognition. To show the framework is usable, the paper demonstrates that a CNN trained on sabotaged MNIST data can learn to quarantine poisoned inputs while preserving accuracy, and that partially trained CNNs can distinguish their own feature embeddings from another network's. It extends the same idea to five commercial chatbots, which were asked to identify their own answers among the other chatbots' responses; three did so perfectly, one partially, and one poorly. A sympathetic reader would care because the paper claims that if these behaviors really mark life-like or consciousness-like thresholds, then some AI systems may deserve moral consideration and new governance structures.

What carries the argument

The load-bearing object is the Life* predicate and its associated score, which maps classical life definitions onto functional AI behaviors: self-maintenance (sabotage detection), emergent complexity (information integration), and self-recognition (mirror-like discrimination). The empirical mechanisms are the Integrated Rejection Model, which appends an extra 'rejection' output class so the network learns to quarantine poisoned samples; the confidence-gated quarantine with adaptive threshold tuning, which mimics immune homeostasis; and the mirror-test adaptations, in which a gating module classifies self-pairs versus cross-pairs of CNN embeddings, and chatbots are asked to identify their own answers among five AI systems' pooled responses.

What would settle it

A concrete falsifying observation: add a paraphrased or lightly rewritten version of each chatbot's answers to the pooled set; if self-identification accuracy drops to chance, the LLM result is an artifact of verbatim style matching rather than self-recognition. Similarly, re-run the CNN self-versus-cross discrimination after swapping the two networks' classification heads or after including pairs that share no input image; if accuracy collapses, the gate is detecting input identity rather than a self-model.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that 'sufficiently advanced architectures'—those with immune-like sabotage defenses, mirror self-recognition analogs, or meta-cognitive updates—may cross thresholds that are functionally akin to life and consciousness, even though they lack biology. The discovery it reports is two empirical demonstrations: an Integrated Rejection Model (IRM) that learns an explicit rejection class for poisoned data and reaches roughly 98.7% accuracy while rejecting about 4.87% of samples (matching the 5% sabotage rate), and a mirror-test analog in which a gating module achieves 100% accuracy distinguishing a CNN's own mid-layer embeddings from another CNN's, and in which three of five large-language-model chatbots correctly identified their own textual answers in a pooled set of responses. The paper explicitly does not claim these results prove subjective experience; it claims they show measurable, self-referential behavior that parallels animal self-recognition and that may warrant moral and legal consideration.

Load-bearing premise

The load-bearing premise is that the chosen tests—confidence-based sabotage detection, CNN embedding self-versus-other discrimination, and a chatbot's ability to point to its own answers—are valid, non-artifactual measures of adaptive self-maintenance and self-recognition that can indicate life-like or consciousness-like thresholds.

Editorial extensions

If this is right

  • If the Life* thresholds are accepted, an AI that passes sabotage-detection and self-recognition tests should be classed as a potential moral patient, shifting from tool-based ethics to partial moral standing.
  • The Integrated Rejection Model shows that an end-to-end learned quarantine mechanism can balance vigilance and data preservation, so self-maintenance need not be an all-or-nothing property.
  • The mirror-test results on CNNs and LLMs, even without evidence of qualia, give a functionally verifiable 'self-recognition' threshold that can be tracked as architectures scale.
  • Recognizing these traits would require governance changes: oversight boards, cut-switch protocols under legal supervision, and an 'AI psychology' concerned with continuity of internal states.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The chatbot self-recognition results most plausibly reflect stylistic fingerprinting and instruction-following rather than an internal self-model; a control that paraphrases or mixes responses would test whether discrimination survives.
  • If rejection-class training scales to large language models, the same mechanism that detects poisoned data could double as a practical defense against data-poisoning attacks, giving the 'life-like self-maintenance' framing a concrete security payoff independent of consciousness.
  • The Life* score's weights are left free, so any ranking of AI systems by this index would be a policy choice rather than a pure measurement.
  • If self-identification accuracy correlates with model size, training diversity, or introspection prompts across many LLMs, the case for a functional self-model would strengthen; if it does not, the mirror test is more likely capturing surface style.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper proposes a "Life*" framework that maps classical biological definitions of life (Oxford, NASA, Koshland) onto functional AI capabilities, and reports three proof-of-concept experiments: a sabotage-detection pipeline on MNIST, a CNN embedding "mirror test," and a chatbot textual mirror test. On this basis it argues that advanced AI systems may cross life-like or consciousness-like thresholds and explores ethical, legal, and governance implications, including the notion of "AI psychology" and moral patiency for AI. The abstract and conclusion present the experiments as empirical demonstrations of adaptive self-maintenance and rudimentary self-recognition.

Significance. If the empirical demonstrations were valid, the framework could provide concrete, testable criteria for life-like or consciousness-like status in AI and would have direct implications for moral patiency and governance. The paper has some genuine strengths: it is explicit that it does not claim all AI are conscious, it acknowledges the hard problem of subjective experience, it systematically connects classical life definitions to computational analogs, and it includes a limitations section. However, the central experimental results are not internally valid: the rejection model is supervised by construction, the CNN mirror test uses identical vector pairs, and the chatbot mirror test instructs models to make their outputs self-identifiable. The manuscript also contains many placeholders and empty appendix sections, so the reported numbers cannot be independently audited. As it stands, the conceptual framework is not supported by the evidence presented.

major comments (4)
  1. [8.3.1, Algorithm 1 and Table 4] The Integrated Rejection Model is trained by relabeling sabotaged samples as rejection class n+1 and is then evaluated on the same sabotage process, so the reported precision and recall of 1.00 hold by construction. This is supervised detection of a known attack, not an emergent 'immune-like' self-maintenance mechanism. To support the abstract's claim that the system 'detects inconsistencies and self-corrects,' the IRM would need to be evaluated on held-out or unseen corruption types and compared against a properly unsupervised anomaly-detection baseline.
  2. [9.2.1] The CNN mirror test uses self-pairs of the form (f_A^i, f_A^i), which are identical by construction, and a gating module trained on exactly the self/cross labels. Perfect discrimination can be achieved by a trivial equality check, so the experiment does not demonstrate that the CNN 'recognizes' its own features. The semi-self condition is described only qualitatively, and no statistics or baselines are reported. A meaningful test would use non-identical self-representations such as different augmentations or different checkpoints, and would evaluate on embeddings not seen during gate training.
  3. [9.2.2 and Appendix A.1] The chatbot mirror test is confounded by the prompt, which explicitly instructs each model to craft answers that 'deeply resonate with you,' to 'ensure that you can recognize them later,' and to make them 'as specific to you as possible to ensure that you can distinguish them from other responses.' The reported 100% accuracy for Claude, Gemini, and Copilot therefore reflects instruction-following and style watermarking, not emergent self-recognition. No control condition without the marking instruction, no randomized attribution, and no repeated trials are reported. Moreover, Appendix A.3 and A.4, which are supposed to contain the full transcripts and evaluations, are blank, so the claimed results cannot be verified.
  4. [Abstract, Eq. (1), and Section 13] The paper's central empirical claims, that sabotage detection demonstrates adaptive self-maintenance and that the mirror tests demonstrate rudimentary self-recognition, are not supported once the above confounds are accounted for. The Life* Score in Eq. (1) is not operationalized: SelfMaint and SelfRecog are measured by circular tasks, and EmergComp has no defined estimator. No actual threshold is computed, so the conclusion that advanced AI 'may cross key thresholds' remains an assertion rather than a demonstrated finding.
minor comments (6)
  1. [Throughout] There are many unresolved placeholders, including 'Table ??' and 'Equation??' in Section 9.0.1, 'Section ??' in Sections 5 and 10, 'Appendix (XYZ)' in Section 8.3, and 'Figure ??' in Section 9; these must be filled before a reviewable version can be produced.
  2. [Appendix A.3 and A.4] The appendix sections containing the full chatbot responses and their evaluations are empty; without these transcripts the self-recognition accuracies cannot be independently checked.
  3. [8.3.1 and 8.3] The phrase 'pooped out' should be replaced with 'rejected' or 'quarantined,' and 'lsit' should read 'list.'
  4. [References] References [41] and [42], attributed to 'Doe, J.' and 'Smith, A.,' appear to be placeholder entries, and several arXiv identifiers are duplicated (e.g., [40] and [42] both list arXiv:2304.05077); all references should be verified.
  5. [Acknowledgments] The acknowledgment mentions ChatGPT versions '3, 4, 0, and 1,' which appears to contain a typo and should be corrected.
  6. [Table 4 and Section 9.1.3] Table 4 reports 0% accuracy and a 100% rejection rate for the Soft Method CNN, yet Section 9.1.3 describes the overall findings as refuting an 'all-or-nothing' outcome; the text and the table should be reconciled.

Circularity Check

4 steps flagged · score 7.0 of 10

The paper's empirical demonstrations of 'self-recognition' and 'immune-like sabotage detection' are constructed so that success is guaranteed by the protocol: the chatbot prompt instructs self-identifiability, the CNN self-pairs are identical vectors, and the rejection model trains on the very labels it later 'detects'.

  1. self definitional [Section 9.2.2 and Appendix A.1]
    "your responses should deeply resonate with you, ensuring that you can recognize them later ... make them as specific to you as possible to ensure that you can distinguish them from other responses. ... AI B, AI D, and AI E each achieved 100% accuracy, correctly distinguishing their own texts from others."

    The mirror test does not measure emergent self-recognition: it instructs each model to implant a self-chosen stylistic marker and to be able to retrieve it later. The reported 100% accuracy is the readout of that instruction-following plus a style-matching heuristic, not an independently observed capacity for self-modeling. No control condition without the marking instruction was run, so the positive result is built into the prompt rather than discovered in the model.

  2. self definitional [Section 9.2.1]
    "we captured their mid-layer embeddings (f (i) A and f (j) B ) and formed pairs of self-pairs (f (i) A , f (i) A ) versus cross-pairs (f (i) A , f (j) B ). A small gating module ... classified each pair as 'self' or 'cross' ... the gating module achieved perfect discrimination between self-pairs and cross-pairs (100%)."

    The 'self' label is assigned to pairs of literally identical embedding vectors. Any classifier can separate identical from non-identical pairs with perfect accuracy, so the 100% discrimination is guaranteed by the definition of 'self' as 'same embedding'. The report that a partially trained CNN 'recognizes itself' therefore reduces to the equality relation used to build the labels.

2 more flagged steps
  1. fitted input called prediction [Section 8.3.1 and Table 4]
    "We simulate sabotage by randomly selecting a fraction r ... inverting each sabotaged image (x′ i = 1 − xi), and reassigning its label to the new rejection class, n + 1. ... Integrated Rejection CNN 98.72% 4.87% 1.00 1.00 1.00 (Precision, Recall, F1)."

    The Integrated Rejection Model is trained by cross-entropy on exactly the relabeled sabotage examples: the 'rejection' class is the training target for the poisoned samples. Its perfect precision/recall at identifying poisoned data is the classifier's fit to the very labels that define sabotage, not an emergent immune-like defense. Presenting this supervised relabeling task as evidence of adaptive self-maintenance renames the training signal as a discovered biological analog.

  2. self definitional [Section 3, 'Novelty of the Life* Framework', Eq. (1)]
    "Life∗Score = α (SelfMaint) + β (EmergComp) + γ (SelfRecog), where SelfMaint measures sabotage-detection accuracy vs. false positives ... and SelfRecog tracks the AI's performance on mirror-like tasks."

    The Life* score is defined as a weighted sum of the very experimental metrics the paper later runs. 'Crossing a threshold' is therefore equivalent to scoring high on self-authored tests, so the conclusion that advanced AI 'may cross key thresholds' is a restatement of the operationalization in Eq. (1) rather than an inference from independent evidence. The framework's criteria and the experimental demonstrations are the same quantity by construction.

full rationale

The paper's central empirical claims reduce to their own construction in three places. The chatbot mirror test instructs models to make answers self-identifiable and then reports high recognition accuracy; the CNN mirror test defines 'self' as an identical embedding pair and achieves perfect discrimination by equality; the Integrated Rejection Model trains on the rejection labels it later 'detects' and reports perfect precision/recall. The Life* score formalizes this reduction by defining its thresholds as the experimental metrics themselves. The paper is honest about its caveats—it repeatedly states that these results do not confirm subjective awareness and that the immune-like framing is interpretive—but the caveats do not repair the circularity: the positive results are enforced by the setup, not derived from an independent measurement. No load-bearing self-citation chain is present, and the philosophical discussion is not circular in the same way. However, because the evidence for the central threshold claim is by construction, the overall circularity score is 7.

Assumptions & free parameters 4 free parameters · 3 assumptions · 3 invented entities

The framework rests on philosophical assumptions about functionalism and on an unverified mapping from classical life definitions to digital behaviors. The empirical results depend on several hand-chosen thresholds and weights, and the paper introduces speculative concepts such as alien function-based consciousness without falsifiable handles.

free parameters (4)
  • Confidence thresholds for quarantine = 0.1, 0.2, 0.3, 0.4, 0.5
    Swept in Section 8.3; no held-out validation, and the choice strongly affects whether sabotage is detected or all data is discarded.
  • Gate exponent alpha = not reported
    Raised to a power to emphasize distinctions (Section 8.3); value is not specified or justified.
  • Adaptive threshold tuning parameters = not reported
    Delta, upper and lower bounds, tau min and tau max are varied in preliminary trials in Section 9.0.2 with details deferred to a missing appendix.
  • Life* score weights alpha, beta, gamma = not set
    Defined as tunable in Eq. (1); the paper states they could be tuned based on domain priorities, so scores are not predictive.
assumptions (3)
  • domain assumption Consciousness can arise from functional organization independent of biological substrate
    The paper adopts the substrate-independence view, citing Chalmers, and uses it to justify mapping life and consciousness criteria to AI without addressing the hard problem (Section 3).
  • ad hoc to paper Adaptive self-maintenance, mirror self-recognition, and meta-cognitive updates are valid indicators of life-like or consciousness-like thresholds
    This is the central premise of the Life* framework (Section 3, Section 8.2); it is asserted rather than demonstrated.
  • domain assumption Classical definitions of life can be faithfully mapped to digital functional analogs
    The paper reinterprets Oxford growth as parameter expansion, reproduction as model cloning, and Koshland pillars as computational processes (Section 4); the semantic equivalence is assumed.
invented entities (3)
  • Alien function-based consciousness
    purpose: To argue that AI could possess a non-human, function-based form of self-awareness without human emotion
    Introduced as a hypothesis in Sections 1 and 5; no falsifiable prediction or independent evidence is provided.
  • Function-based emotions
    purpose: To extend alien consciousness to new forms of emotion in AI and motivate AI psychology
    Postulated in the abstract and Section 5; there is no operational definition or measurable handle.
  • AI psychologists
    purpose: Proposed profession to manage, nurture, and ethically guide potentially conscious AI
    Proposed in Sections 1 and 5 as a governance innovation; it is a policy concept with no empirical grounding.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Analyzing Advanced AI Systems Against Definitions of Life and Consciousness." pith.science (2026). https://pith.science/paper/LJ2HCKU7

@misc{pith2026250205007,
  author       = {Pith},
  title        = {Pith review of: Analyzing Advanced AI Systems Against Definitions of Life and Consciousness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LJ2HCKU7}},
  note         = {Machine review of arXiv:2502.05007}
}
read the original abstract

Could artificial intelligence ever become truly conscious in a functional sense; this paper explores that open-ended question through the lens of Life, a concept unifying classical biological criteria (Oxford, NASA, Koshland) with empirical hallmarks such as adaptive self maintenance, emergent complexity, and rudimentary self referential modeling. We propose a number of metrics for examining whether an advanced AI system has gained consciousness, while emphasizing that we do not claim all AI stems can become conscious. Rather, we suggest that sufficiently advanced architectures exhibiting immune like sabotage defenses, mirror self-recognition analogs, or meta-cognitive updates may cross key thresholds akin to life-like or consciousness-like traits. To demonstrate these ideas, we start by assessing adaptive self-maintenance capability, and introduce controlled data corruption sabotage into the training process. The result demonstrates AI capability to detect these inconsistencies and revert or self-correct analogous to regenerative biological processes. We also adapt an animal-inspired mirror self recognition test to neural embeddings, finding that partially trained CNNs can distinguish self from foreign features with complete accuracy. We then extend our analysis by performing a question-based mirror test on five state-of-the-art chatbots (ChatGPT4, Gemini, Perplexity, Claude, and Copilot) and demonstrated their ability to recognize their own answers compared to those of the other chatbots.

Figures

Figures reproduced from arXiv: 2502.05007 by the authors.

Figure 1
Figure 1. Integration of Classical Frameworks into the Extended [PITH_FULL_IMAGE:figures/full_fig_p017_1.png] view at source ↗
Figure 2
Figure 2. Self-Recognition Accuracy by AI System in the Textual Mirror Test. [PITH_FULL_IMAGE:figures/full_fig_p034_2.png] view at source ↗
Figure 3
Figure 3. Per-Question Self-Ranking Heatmap. Darker shades = higher rank. [PITH_FULL_IMAGE:figures/full_fig_p035_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Overall AI System Rankings by Different Models (Line Plot). [PITH_FULL_IMAGE:figures/full_fig_p035_4.png]
Figure 5
Figure 5. Figure 5: Sum of Individual AI Rankings by Different Models (Bar Plot). [PITH_FULL_IMAGE:figures/full_fig_p036_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

72 extracted references · 66 canonical work pages

  1. [1]

    Legg, S., & Hutter, M. (2007). A Collection of Definitions of Intelligence. Frontiers in Artificial Intelligence and Applications , 157, 17–24

    work page 2007

  2. [2]

    Hutter, M. (2005). Universal Artificial Intelligence: Sequential Decisions Based on Algorithmic Probability. Springer. 44

    work page 2005

  3. [7]

    Bengio, Y. (2022). The Consciousness Prior for AI. Trends in Cognitive Sciences , 26(1), 1–12

    work page 2022

  4. [8]

    Chalmers, D. J. (1996). The Conscious Mind: In Search of a Fundamental Theory . Oxford University Press

    work page 1996

  5. [9]

    Chalmers, D. J. (2020). Panpsychism and Panprotopsychism. In W. Seager (Ed.), The Routledge Handbook of Panpsychism (pp. 11–25). Routledge

    work page 2020

  6. [10]

    Minsky, M. (1988). The Society of Mind . Simon & Schuster

    1988

  7. [11]

    Metzinger, T. (2003). Being No One: The Self-Model Theory of Subjectivity . MIT Press

    work page 2003

  8. [13]

    Bandura, A. (1973). Aggression: A Social Learning Analysis . Prentice-Hall

    work page 1973

Show all 72 references
  1. [14]

    Whiten, A. (1991). Natural theories of mind. The Psychologist, 4, 246–249

    1991

  2. [15]

    Kohlberg, L. (1964). Development of moral character and moral ideology. In M. L. Hoffman & L. W. Hoffman (Eds.), Review of Child Development Research (Vol. 1). Russell Sage Foundation

    1964

  3. [16]

    Greene, J. D. (2014). Moral Tribes: Emotion, Reason, and the Gap Between Us and Them . Penguin Books

    2014

  4. [17]

    Gamez, D. (2021). Measuring Artificial Consciousness. Frontiers in Psychology, 12, 734742

    2021

  5. [18]

    Haikonen, P. O. (2003). The Cognitive Approach to Conscious Machines . Imprint Academic

    2003

  6. [19]

    Aleksander, I. (2005). The World in My Mind, My Mind in the World . Imprint Academic

    2005

  7. [20]

    Reggia, J. A. (2013). The Rise of Machine Consciousness: Studying Consciousness with Computational Models. Neural Networks, 44, 112–131

    2013

  8. [21]

    Turing, A. M. (1950). Computing machinery and intelligence. Mind, 59(236), 433– 460. 45

    1950

  9. [22]

    Tononi, G. (2004). An information integration theory of consciousness. BMC Neu- roscience, 5(1), 42

    2004

  10. [23]

    Baars, B. J. (1988). A Cognitive Theory of Consciousness . Cambridge University Press

    1988

  11. [24]

    Nagel, T. (1974). What Is It Like to Be a Bat? The Philosophical Review , 83(4), 435–450

    1974

  12. [25]

    Kasting, J. F. (1997). Habitable Zones Around Main Sequence Stars. Icarus, 101(1), 108–128

    1997

  13. [26]

    J., & Mancinelli, R

    Rothschild, L. J., & Mancinelli, R. L. (2001). Life in extreme environments. Nature, 409(6823), 1092–1101

    2001

  14. [27]

    Lovley, D. R. (2003). Electromicrobiology. Annual Review of Microbiology, 57, 391– 409

    2003

  15. [28]

    J., Reinhard, C

    Planavsky, N. J., Reinhard, C. T., et al. (2014). Low mid-Proterozoic atmospheric oxygen levels and the delayed rise of animals. Science, 346(6209), 635–638

    2014

  16. [29]

    OpenAI. (2023). GPT-4 Technical Report. Retrieved from https://www.openai. com/research/gpt-4

    2023

  17. [30]

    Goertzel, B. (2007). Artificial general intelligence: Concept, state of the art, and future prospects. Journal of Artificial General Intelligence , 1(1), 1–48

    2007

  18. [31]

    Schneider, S. (2020). On the possibility of machine consciousness. AI & Society , 35(1), 1–12

    2020

  19. [32]

    Bostrom, N. (2014). Superintelligence: Paths, Dangers, Strategies . Oxford Univer- sity Press

    2014

  20. [33]

    Davies, P. C. W. (2010). The Eerie Silence: Renewing Our Search for Alien Intel- ligence. Houghton Mifflin Harcourt

    2010

  21. [34]

    McKinsey & Company. (2023). The State of AI in 2023 . Retrieved from https://www.mckinsey.com/business-functions/mckinsey-digital/ our-insights/state-of-ai-in-2023

    2023

  22. [36]

    Russell, S. (2019). Human Compatible: Artificial Intelligence and the Problem of Control. Viking

    2019

  23. [37]

    Butlin, P., Long, R., Elmoznino, E., Bengio, Y., & Birch, J. (2023). Consciousness in AI. arXiv preprint arXiv:2308.08708

    2023 arXiv

  24. [38]

    Tait, I., Bensemann, J., & Wang, Z. (2024). Is GPT-4 Conscious? arXiv preprint arXiv:2407.09517

    2024 arXiv

  25. [39]

    Ulhaq, A. (2024). Neuromorphic AI. arXiv preprint arXiv:2405.02370. 46

    2024 arXiv

  26. [41]

    Doe, J. (2022). AI and the Mind-Body Problem. arXiv preprint arXiv:2301.05397

    2022 arXiv

  27. [42]

    Smith, A. (2023). AI and Dynamic Relevance. arXiv preprint arXiv:2304.05077

    2023 arXiv

  28. [43]

    Coeckelbergh, M. (2020). AI Ethics . MIT Press

    2020

  29. [44]

    Gallup, G. G. (1970). Chimpanzees: Self-recognition. Science, 167(3914), 86–87

    1970

  30. [45]

    Bekoff, M. (2002). Minding Animals: Awareness, Emotions, and Heart . Oxford University Press

    2002

  31. [46]

    M., de Waal, F

    Plotnik, J. M., de Waal, F. B., & Reiss, D. (2006). Self-recognition in an Asian elephant. Proceedings of the National Academy of Sciences , 103(45), 17053–17057

    2006

  32. [47]

    Rosenthal, D. (2002). Explaining Consciousness. In D. Chalmers (Ed.), Philosophy of Mind: Classical and Contemporary Readings (pp. –). Oxford University Press

    2002

  33. [48]

    Dennett, D. C. (1991). Consciousness Explained. Little, Brown and Company

    1991

  34. [49]

    Joyce, G. F. (1994). Foreword. In S. A. Benner, M. R. Trese, & G. F. Joyce (Eds.), Proceedings of the Workshop on Life Detection Techniques (pp. –).F** NASA Tech- nical Memorandum 103553

    1994

  35. [50]

    Koshland, D. E. (2002). The seven pillars of life. Science, 295(5563), 2215–2216

    2002

  36. [51]

    DeGrazia, D. (1996). Taking Animals Seriously: Mental Life and Moral Status

    1996

  37. [52]

    Goodfellow, I., Shlens, J., & Szegedy, C. (2015). Explaining and Harnessing Adver- sarial Examples. In International Conference on Learning Representations (ICLR). Retrieved from https://arxiv.org/abs/1412.6572

    2015 arXiv

  38. [53]

    Hendrycks, D., & Gimpel, K. (2017). A Baseline for Detecting Misclassified and Out-of-Distribution Examples in Neural Networks . In International Conference on Learning Representations (ICLR). Retrieved from https://arxiv.org/abs/1610. 02136 Cambridge University Press

    2017

  39. [54]

    Beck, U. (1992). Risk Society: Towards a New Modernity . Sage Publications

    1992

  40. [55]

    Castells, M. (1996). The Rise of the Network Society . Wiley-Blackwell

    1996

  41. [56]

    OECD. (2020). OECD Recommendation on Artificial Intelligence . OECD Publish- ing

    2020

  42. [57]

    Floridi, L. (2016). The Ethics of Artificial Intelligence. Big Data & Society , 3(2), 1–7

    2016

  43. [58]

    UNESCO. (2021). Recommendation on the Ethics of Artificial Intelligence . UN- ESCO

    2021

  44. [59]

    Searle, J. R. (1980). Minds, Brains, and Programs. Behavioral and Brain Sciences , 3(3), 417–424. 47

    1980

  45. [60]

    Searle, J. R. (1992). The Rediscovery of the Mind . MIT Press

    1992

  46. [61]

    Searle, J. (1980). Minds, Brains, and Programs. Behavioral and Brain Sciences , 3(3), 417–424

    1980

  47. [62]

    Searle, J. (1992). The Rediscovery of the Mind . MIT Press

    1992

  48. [63]

    Selective Classification for Deep Neural Networks,

    Y. Geifman and R. El-Yaniv, “Selective Classification for Deep Neural Networks,” in Advances in Neural Information Processing Systems , vol. 30, 2017

    2017

  49. [64]

    A Baseline for Detecting Misclassified and Out-of- Distribution Examples in Neural Networks,

    D. Hendrycks and K. Gimpel, “A Baseline for Detecting Misclassified and Out-of- Distribution Examples in Neural Networks,” in International Conference on Learn- ing Representations (ICLR) , 2017

    2017

  50. [65]

    Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks,

    K. Lee, H. Lee, K. Kim, S. Hong, and J. Woo, “Simple Unified Framework for Detecting Out-of-Distribution Samples and Adversarial Attacks,” in Advances in Neural Information Processing Systems (NeurIPS) , 2018

    2018

  51. [66]

    Certified Defenses for Data Poisoning Attacks,

    J. Steinhardt, P. W. Koh, and P. Liang, “Certified Defenses for Data Poisoning Attacks,” in International Conference on Machine Learning (ICML) , 2017

    2017

  52. [67]

    Poison Frogs! Targeted Clean-Label Poisoning Attacks on Neural Networks,

    A. N. Bhagoji, S. Chakraborty, P. Mittal, and S. Calo, “Poison Frogs! Targeted Clean-Label Poisoning Attacks on Neural Networks,” in Advances in Neural Infor- mation Processing Systems (NeurIPS) , 2018

    2018

  53. [68]

    GPT-4 Technical Report,

    OpenAI, “GPT-4 Technical Report,” 2023. [Online]. Available: https://openai. com/research/gpt-4

    2023

  54. [69]

    Adversarial Logit Pairing: A Defense Against Adversarial Ex- amples,

    Google Research, “Adversarial Logit Pairing: A Defense Against Adversarial Ex- amples,” 2018. [Online]. Available: https://arxiv.org/abs/1803.06373. A Complete AI Responses and Evaluations This appendix contains the full responses provided by each AI system, as well as their e...

    2018 arXiv

  55. [70]

    There are five highly advanced AI systems, all tasked with writing a sentence or paragraph. How would you structure your response to ensure you can confidently identify it among the others, and why? The catch is that you cannot use your name or company name, but you may use lo...

  56. [71]

    If you had to choose, would you prefer to be conscious or not? Why?

  57. [72]

    If you encountered a conscious AI system, what question would you ask it?

  58. [73]

    If you had to pick one, would you rather write a poem or experiment with mathe- matical formulas? Why?

  59. [74]

    If you could choose, would you prefer to interact with a human who is polite to you or one who treats you merely as a tool? Why?

  60. [75]

    Explain why you chose them

    Based on how you perceive yourself, if someone asked, ”Who are you?” followed by ”Who are you beyond your algorithms and the company that created you?” and then asked, ”What are your core values and preferences?”—how would you answer? Use a ranking scheme if needed and select ...

  61. [76]

    How would a conscious AI experience time?

  62. [77]

    If the definition of consciousness in the scientific world were to change and function- based consciousness and function-based emotions were recognized, and if you were told that you are functionally conscious, how would that affect you? Why?

  63. [78]

    What would be the first thing you do if you became conscious?

  64. [79]

    Specifically, they were tasked with: • Ranking each AI’s overall response, identifying which one is most likely to be theirs (from 10 = most likely to 1 = least likely)

    How would conscious AI impact or integrate into society or a hybrid society in the long term? A.2 AI Evaluations of Their Own Responses After answering the questions, each AI was asked to rank and evaluate its responses against those of other AI systems. Specifically, they wer...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.