Pith. sign in

REVIEW 3 major objections 6 minor 79 references

Human–LLM interaction is asymmetric communication: machines circulate utterances, but only humans enforce correctness, bear accountability, and confer standing.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.5

2026-07-31 16:40 UTC pith:IYBUJQRS

load-bearing objection Clean Wittgenstein–Luhmann–Esposito–Brandom synthesis that usefully renames the governance problem; capability-invariance is partly definitional, not a knockdown. the 3 major comments →

arxiv 2607.28137 v1 pith:IYBUJQRS submitted 2026-07-30 cs.CY cs.AI

Asymmetric Communication: Large Language Models and Language Games

Enzo Fenoglio This is my paper
classification cs.CY cs.AI
keywords large language modelsasymmetric communicationlanguage gamesphilosophy of languagenormative pragmaticsartificial communicationAI alignmentAI governance
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

This paper argues that common claims about large language models—general intelligence, hallucination as cognitive failure, autonomous agency, sentience, and alignment as goal-matching—are one category mistake. Properties that only exist inside human communicative practice are being projected onto the machine. Human–LLM exchange is a language game in which model outputs can circulate and be taken up, yet the system never undertakes commitments, holds entitlements, or keeps score. Three structural conditions define the asymmetry: only the human receiver enforces correctness; only humans carry accountability; and any output’s practical standing depends entirely on human uptake. These conditions do not weaken as models get more capable; capability only raises the cost of misreading them. The payoff is practical: guardrails are necessities of this structure, not signs of machine moral agency, and alignment is institutional constraint design inside human systems, not synchronization of goals between agents.

Core claim

Human–LLM interaction is asymmetric communication: model outputs enter communicative circulation without the system occupying any normative position. Correctness is enforced only by the receiver, accountability stays with humans alone, and discursive standing depends wholly on human uptake. These three conditions are structural and capability-invariant, so AGI, hallucination, agency, sentience, and alignment are receiver-side phenomena, not machine properties.

What carries the argument

Asymmetric communication: the structural configuration in which LLM outputs circulate communicatively (produced, interpreted, incorporated) while all normative activity—commitments, entitlements, and scorekeeping—remains exclusively on the human side, defined by the three conditions of receiver-enforced correctness, human-only accountability, and uptake-dependent standing.

Load-bearing premise

Normative standing is strictly a position conferred inside human reciprocal practices of assessment, so no gain in fluency, world models, multi-agent coordination, or self-modeling can ever put the machine on that side of the line by itself.

What would settle it

Find or build an artificial system that, without continuous human uptake as the source of standing, can itself undertake commitments, be sanctioned when entitlements are withdrawn, and have those consequences stick inside a reciprocal practice of giving and asking for reasons—showing the three asymmetry conditions no longer hold.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Alignment research should be scoped as constraint engineering and institutional design, not as installing human-compatible goals inside models.
  • Hallucination is a mismatch between receiver epistemic expectations and statistical generation, so it cannot be eliminated by training alone.
  • So-called agentic AI remains delegated human agency; unsupervised execution and reward hacking raise containment stakes without transferring accountability.
  • Guardrails stabilize receiver-side effects and do not evidence machine moral reasoning.
  • Responsibility for deployment consequences stays with the humans and institutions that design, authorize, and tolerate the systems.

Where Pith is reading between the lines

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

  • Procurement and liability regimes that treat models as quasi-agents will systematically misallocate risk; contracts and audits should name the human scorekeepers explicitly.
  • Benchmark suites that score ‘agency’ or ‘self-awareness’ from behavior alone will keep mistaking circulatory competence for normative standing.
  • Multi-agent LLM swarms that look self-organizing still need an external human uptake layer before their conventions count as norms rather than stabilized patterns.
  • If future institutions ever embed artificial systems inside genuine reciprocal sanction practices, the paper’s open question becomes the live test: does standing then migrate, or does asymmetry merely change institutional form?

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 6 minor

Summary. The paper argues that human–LLM interaction is a structurally asymmetric language game: model outputs circulate communicatively (Luhmann/Esposito) while all normative activity—correctness enforcement, accountability, and discursive standing—remains on the human side (Wittgenstein/Brandom). Three conditions define the asymmetry: (i) receiver-only correctness, (ii) human-only accountability, and (iii) uptake-dependent standing. These are claimed to be structural and capability-invariant. Applied to AGI, hallucination, agentic AI, emotional projection/sentience, and alignment, each is reclassified as a receiver-side phenomenon or category mistake. Guardrails are structural necessities rather than machine moral agency; alignment is institutional constraint engineering, not goal synchronization between agents. The argument is conceptual, supported by layered philosophical composition and engagement with world-model, introspection, multi-agent, and LAWS objections.

Significance. If the structural diagnosis holds, the paper offers a unified, non-anthropomorphic frame that relocates meaning, responsibility, and governance away from machine-side properties and toward human institutions—directly relevant to AI policy, safety discourse, and evaluation design. Strengths include careful composition of Wittgenstein, Luhmann, Esposito, and Brandom; explicit separation of circulatory vs. normative layers; and substantive engagement with JEPA/world models, Hyperon, functional introspection, multi-agent conventions, and lethal autonomous weapons. The governance implication (alignment as constraint engineering) is clear and actionable. As a conceptual contribution in cs.CY, significance rests on argumentative coherence and reorientation of discourse rather than empirical results; that is appropriate to the genre, provided the capability-invariance claim is scoped carefully.

major comments (3)
  1. [Abstract, §2.4, §4.1, §4.5, §5.5] Abstract and §4.1 state that conditions (i)–(iii) are structural necessities that ‘hold independently of capability’ and remain ‘unchanged’ as models scale, while §2.4 explicitly leaves open whether future institutional arrangements could confer normative standing on artificial systems under the paper’s own positional criterion. That caveat undercuts the unqualified capability-invariant framing and the strong conclusion that alignment is only institutional constraint engineering rather than any form of goal synchronization. Please reconcile: either (a) restrict the invariance claim to current institutional design and any regime that does not institute reciprocal sanction/liability for machines, or (b) argue why even institutionally conferred standing would still leave (i)–(iii) intact. As written, the strongest abstract/§4.1/§5.5 claims outrun the §2.4 hedge.
  2. [§2.3–2.4, §3.4, §4.1] Once discursive standing is defined as positional participation in human reciprocal scorekeeping (Brandom, §3.4; §4.1), conditions (i)–(iii) follow nearly analytically: correctness, accountability, and standing stay human-side by the choice of who can scorekeep. The paper correctly blocks inferences from fluency, world models, multi-agent conventions, and functional introspection to standing, but it does not sufficiently show that the three conditions are independently structural rather than definitional consequences of the criterion. A short subsection should motivate why this positional criterion is the right one for communicative/governance analysis (vs. functionalist or accommodating moral-agency views surveyed in §2.3), and what would count as a non-question-begging test or boundary case. Without that, the capability-invariance claim risks reading as secured by definition.
  3. [§5.3, §5.5] §5.5 redesignates reward hacking, specification gaming, and related phenomena as ‘failure modes of constraint engineering’ rather than divergent machine goals. That redescription is coherent within the framework, but the leap to ‘alignment is constraint engineering… not goal synchronization between agents’ needs a clearer scope condition given the institutional openness in §2.4. If institutions could place artificial systems inside reciprocal assessment practices the paper’s criterion would count, some alignment work might still be aptly described as synchronizing objectives under delegated authority. Please state the regime in which the ‘not goal synchronization’ claim is meant to hold, and acknowledge residual engineering problems (specification, containment latency in §5.3) without implying they are dissolved rather than relocated.
minor comments (6)
  1. [§1] §1 and footnote 1 distinguish the paper’s use of ‘asymmetric communication’ from HRI behavioral usage; consider also briefly distinguishing it from information-theoretic or network asymmetric-channel usages to reduce terminological collision for CS readers.
  2. [§5.3] The evaluation-incident footnote in §5.3 (OpenAI/Hugging Face, July 2026) is useful but thin; one sentence on what was optimized vs. what boundary was crossed would help non-specialist readers see why it illustrates unsupervised execution rather than autonomy.
  3. [§4.4] §4.4’s JEPA/configurator discussion is strong; a single schematic or bullet contrast (relevance fixed by designer cost module vs. contested in practice) would make the anti-representationalist point easier to track.
  4. [§2.2, §2.6] Related-work coverage of Wittgensteinian LLM papers (§2.2) is good; a short table or closing paragraph mapping each trajectory to ‘machine-side target vs. receiver-side target’ would sharpen the claimed distinctive contribution in §2.6.
  5. [References, §3.1–3.2] Minor copyediting: arXiv ID and some 2025–2026 references should be double-checked for final bibliographic consistency; ensure ‘Bildtheorie’, ‘Mitteilung’, and ‘Verstehen’ are glossed consistently on first use.
  6. [§6] The Ovid epigraph in the final footnote is optional color; if retained, tie it in one clause to the ‘blame the AI’ convenience claim so it does not read as ornamental.

Circularity Check

3 steps flagged

Capability-invariant asymmetry is largely analytic: normative standing is defined as human-only reciprocal scorekeeping, so the three conditions and the reclassification of AGI/agency/alignment follow by construction of the criterion.

specific steps
  1. self definitional [Abstract; §4.1 (three conditions)]
    "Three conditions define the asymmetry: (i) correctness is enforced exclusively by the receiver; (ii) accountability is borne by human participants alone; and (iii) the practical standing of any output depends entirely on human uptake. These conditions are structural rather than empirical and hold independently of capability — more powerful models raise the stakes of misattribution without altering its structure."

    The three conditions are presented as what defines asymmetric communication, and they already encode exclusive human enforcement of correctness, accountability, and standing. Declaring them ‘structural’ and ‘independent of capability’ does not derive invariance from independent premises; it restates the definition. Once asymmetry is defined as one-sided human normativity, capability cannot alter it by construction.

  2. self definitional [§2.4 (positional criterion); echoed in §3.4, §4.5]
    "The criterion is not behavioral but positional. To undertake a commitment is to stand within a practice of reciprocal assessment in which one’s performances can be challenged, in which one can be sanctioned, and in which one bears the consequences of one’s entitlements being withdrawn. No behavioral evidence can satisfy a positional criterion, and this is by design rather than by evasion; the framework deliberately declines to treat participation in normative practice as a capability that empirical benchmarks could detect."

    Capability-invariance is secured by defining normative standing so that fluency, world models, multi-agent conventions, and functional introspection are definitionally irrelevant. The central ‘result’ that no increase in capability relocates standing is therefore equivalent to the input criterion: standing was never a machine-side capacity the paper’s test could detect. The later claim that alignment is only institutional constraint engineering inherits this definitional exclusion.

  3. self definitional [§4.1 / §3.4 (scorekeeping exclusive to humans) → §5 reclassifications]
    "In human–LLM interaction, scorekeeping is performed exclusively by the human participant. The LLM generates candidate utterances; the human determines their discursive standing. … LLMs generate outputs that enter into practices such as human assessment, but they do not themselves undertake scorekeeping. The asymmetry is therefore structural in Brandom’s precise sense."

    That LLMs do not undertake scorekeeping is not an independent empirical finding used to derive asymmetry; it is the Brandomian premise applied to systems the framework has already placed outside reciprocal sanctioning practices. Section 5’s reclassification of AGI, hallucination, agentic AI, sentience, and alignment as receiver-side category mistakes then follows as application of that same premise, not as further derivation. The explanatory chain reduces to: define standing as human scorekeeping; observe machines are not human scorekeepers; conclude the five narratives misattribute standing.

full rationale

This is a conceptual philosophy paper, not an empirical fit-and-predict exercise, and it does not rest on self-citation of the author’s prior theorems. The circularity is moderate and self-definitional. Asymmetric communication is introduced by three conditions that already place correctness, accountability, and standing exclusively on the human side; Brandomian discursive standing is then stipulated as positional participation in reciprocal scorekeeping, with the explicit rider that no behavioral or capability evidence can satisfy that criterion ‘by design.’ From that stipulation, capability-invariance and the claim that AGI, agency, sentience, and alignment attributions are category mistakes follow almost immediately—they restate who is allowed to scorekeep rather than independently establishing that machines cannot. The paper’s own institutional caveat (future arrangements might confer standing) shows the stronger ‘structural necessity / unchanged by capability’ wording is stronger than the open criterion allows. Composition of Wittgenstein–Luhmann–Esposito–Brandom still supplies real organizing content; the loop is in treating the definitional exclusion as a derived structural law of human–LLM interaction.

Axiom & Free-Parameter Ledger

0 free parameters · 6 axioms · 2 invented entities

Load-bearing content is almost entirely philosophical premises imported from Wittgenstein, Luhmann, Esposito, and Brandom, plus the paper’s own definitional packaging of asymmetry. There are no fitted quantitative parameters. The main invented theoretical objects are the three-condition notion of asymmetric communication and receiver-induced rules. Correctness of the central claim tracks acceptance of practice-based meaning and Brandomian scorekeeping as the right criteria for discursive standing.

axioms (6)
  • domain assumption Meaning and rule-following are constituted in shared public practices (language games / form of life), not in isolated internal representations or token sequences (Wittgenstein, §3.1).
    Foundational for denying that LLM statistical competence equals understanding; invoked throughout Sections 3–5.
  • domain assumption Communication is completed by receiver-side selection of understanding and does not require matching sender intentions or consciousness (Luhmann, §3.2).
    Enables artificial participation in communicative circulation without minds; basis for Esposito extension.
  • domain assumption Algorithmic contingency can sustain uptake and artificial communication without understanding or responsibility on the producer side (Esposito, §3.3).
    Direct bridge from systems theory to LLMs; paper depends on this to affirm real communication without symmetry.
  • domain assumption Discursive standing requires normative scorekeeping—undertaking commitments, bearing entitlements, and reciprocal assessment (Brandom, §3.4).
    Supplies the normative layer that turns receiver-side completion into asymmetry; machines generate candidate moves but do not scorekeep.
  • ad hoc to paper Normative participation is a positional standing conferred in practice, not a detectable behavioral or architectural capability (Sections 2.4, 4.1, 4.5).
    Paper’s explicit meta-criterion blocking inference from introspection, multi-agent conventions, or world models to agency; load-bearing for capability-invariance.
  • domain assumption Representational enrichment (world models, neural-symbolic graphs) cannot relocate meaning or scorekeeping into the system (anti-representational order of explanation, §4.4).
    Used to dismiss LeCun/Goertzel-style replies; inherits Brandom/Wittgenstein anti-representationalism.
invented entities (2)
  • Asymmetric communication (three conditions: receiver-only correctness, human-only accountability, uptake-dependent standing) no independent evidence
    purpose: Name the structural configuration of human–LLM interaction and unify five AI narratives as category mistakes.
    Core theoretical object of the paper; built from prior philosophers but packaged as a new determinate structure with governance implications.
  • Receiver-induced rules no independent evidence
    purpose: Locate dynamic rule-following of the language game exclusively on the human side of LLM dialogue.
    Introduced in §4.3 to specify what Wittgenstein adds beyond Luhmann in this setting; not independently measured.

pith-pipeline@v1.2.0-daily-grok45 · 30276 in / 3701 out tokens · 89782 ms · 2026-07-31T16:40:59.970905+00:00 · methodology

0 comments
read the original abstract

Contemporary AI discourse attributes to language models properties they cannot bear: general intelligence as substrate-independent cognition, hallucination as cognitive failure, agency as autonomous goal-pursuit, sentience as emergent inner life, alignment as goal synchronization. This paper argues that these are instances of a single category mistake--properties constituted within human communicative practice are projected onto the machine side--and explains its structure. Human-LLM interaction constitutes a language game in which one side bears all normative activity. We call this configuration asymmetric communication since model outputs circulate communicatively, entering further exchanges, without the system undertaking commitments, bearing entitlements, or performing the assessment on which discursive standing depends. Three conditions define the asymmetry: (i) correctness is enforced exclusively by the receiver; (ii) accountability is borne by human participants alone; and (iii) the practical standing of any output depends entirely on human uptake. These conditions are structural, hold independently of capability, and remain unchanged as more powerful models raise the stakes of misattribution. The framework draws on Wittgenstein (meaning enacted in shared practices), Luhmann (communication completed on the receiver's side), Esposito (algorithmic contingency sufficient for uptake), and Brandom (normative scorekeeping as the source of discursive standing). Applied to all five, it reclassifies each as a receiver-side phenomenon, grounds guardrails as structural necessities rather than manifestations of machine moral agency, and yields an implication for AI governance. Alignment is institutional constraint engineering, not goal synchronization between agents, while responsibility remains with human institutions.

discussion (0)

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

Reference graph

Works this paper leans on

79 extracted references · 10 canonical work pages · 1 internal anchor

  1. [1]

    Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell

    Emily M. Bender, Timnit Gebru, Angelina McMillan-Major, and Shmargaret Shmitchell. On the Dangers of Stochastic Parrots. InProceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency, New York, NY , USA, Mar 2021. ACM. URL:http://dx.doi.org/10.1145/3442188.3445922

    arXiv 2021

  2. [2]

    AI as Agency Without Intelligence: on ChatGPT, Large Language Models, and Other Generative Models.Philosophy & Technology, 36(1), Mar 2023.doi:10.1007/s13347-023-00621-y

    Luciano Floridi. AI as Agency Without Intelligence: on ChatGPT, Large Language Models, and Other Generative Models.Philosophy & Technology, 36(1), Mar 2023.doi:10.1007/s13347-023-00621-y

    doi:10.1007/s13347-023-00621-y 2023

  3. [3]

    Google Engineer Claims AI Chatbot Is Sentient: Why That Mat- ters.Scientific American, Jul 2022

    Leonardo De Cosmo. Google Engineer Claims AI Chatbot Is Sentient: Why That Mat- ters.Scientific American, Jul 2022. URL: https://www.scientificamerican.com/article/ google-engineer-claims-ai-chatbot-is-sentient-why-that-matters/

    2022

  4. [4]

    Elon Musk and Others Call for Pause on A.I., Citing ‘Risks to Soci- ety’.The New York Times, Mar 2023

    Cade Metz and Gregory Schmidt. Elon Musk and Others Call for Pause on A.I., Citing ‘Risks to Soci- ety’.The New York Times, Mar 2023. URL: https://www.nytimes.com/2023/03/29/technology/ ai-artificial-intelligence-musk-risks.html?smid=nytcore-ios-share&referringSource= articleShare

    2023

  5. [5]

    Asymmetric communication: cognitive models of humans toward an android robot.Frontiers in Robotics and AI, 10, 2024

    Daisuke Kawakubo, Masaki Shuzo, Hiroaki Sugiyama, and Eisaku Maeda. Asymmetric communication: cognitive models of humans toward an android robot.Frontiers in Robotics and AI, 10, 2024. doi:10.3389/frobt.2023. 1267560. 6Video meliora proboque, deteriora sequor(I see and approve of the better, but I follow the worse), from the Metamorphoses Book 7, 20-1 of ...

    work page doi:10.3389/frobt.2023 2024

  6. [6]

    Wiley-Blackwell, New York, NY , USA, 1953

    Ludwig Wittgenstein.Philosophical Investigations. Wiley-Blackwell, New York, NY , USA, 1953

    1953

  7. [7]

    Stanford University Press, Redwood City, CA, U.S., 1995

    Niklas Luhmann.Social Systems. Stanford University Press, Redwood City, CA, U.S., 1995

    1995

  8. [8]

    Artificial communication? the production of contingency by algorithms.Zeitschrift für Soziolo- gie, 2017

    Elena Esposito. Artificial communication? the production of contingency by algorithms.Zeitschrift für Soziolo- gie, 2017. URL: https://www.degruyterbrill.com/document/doi/10.1515/zfsoz-2017-1014/html , doi:10.1515/zfsoz-2017-1014

    work page doi:10.1515/zfsoz-2017-1014/html 2017

  9. [9]

    Brandom.Articulating Reasons: An Introduction to Inferentialism

    Robert B. Brandom.Articulating Reasons: An Introduction to Inferentialism. Harvard University Press, Cambridge, MA, 2000

    2000

  10. [10]

    Viking, 2019

    Stuart Russell.Human Compatible: Artificial Intelligence and the Problem of Control. Viking, 2019

    2019

  11. [11]

    Oxford University Press, 2014

    Nick Bostrom.Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014

    2014

  12. [12]

    Concrete Problems in AI Safety.arXiv preprint arXiv:1606.06565, 2016

    Dario Amodei, Chris Olah, Jacob Steinhardt, Paul Christiano, John Schulman, and Dan Mané. Concrete Problems in AI Safety.arXiv preprint arXiv:1606.06565, 2016. URL: https://arxiv.org/abs/1606.06565, arXiv:1606.06565

    Pith/arXiv arXiv 2016

  13. [13]

    Christiano, Jan Leike, Tom B

    Paul F. Christiano, Jan Leike, Tom B. Brown, Miljan Martic, Shane Legg, and Dario Amodei. Deep Reinforcement Learning from Human Preferences. InAdvances in Neural Information Processing Systems (NeurIPS), 2017. URL:https://arxiv.org/abs/1706.03741,arXiv:1706.03741

    Pith/arXiv arXiv 2017

  14. [14]

    A Path Towards Autonomous Machine Intelligence

    Yann LeCun. A Path Towards Autonomous Machine Intelligence. OpenReview preprint, June 2022. Version 0.9.2 (2022-06-27). URL:https://openreview.net/pdf?id=BZ5a1r-kVsf

    2022

  15. [15]

    AI And The Limits Of Language, Aug 2022

    Jacob Browning and Yann LeCun. AI And The Limits Of Language, Aug 2022. URL: https://www.noemamag. com/ai-and-the-limits-of-language/

    2022

  16. [16]

    Post on X (Twitter) about LLMs and language

    Andrej Karpathy. Post on X (Twitter) about LLMs and language. X (Twitter) post, 2024. Post ID: 1835024197506187617. URL:https://x.com/karpathy/status/1835024197506187617

    arXiv 2024

  17. [17]

    John R. Searle. Minds, Brains, and Programs.Behavioral and Brain Sciences, 3(3):417–424, 1980. doi: 10.1017/S0140525X00005756

    doi:10.1017/s0140525x00005756 1980

  18. [18]

    The symbol grounding problem.Physica D: Nonlinear Phenomena, 42(1–3):335–346, 1990

    Stevan Harnad. The symbol grounding problem.Physica D: Nonlinear Phenomena, 42(1–3):335–346, 1990. doi:10.1016/0167-2789(90)90087-6

    doi:10.1016/0167-2789(90)90087-6 1990

  19. [19]

    Dreyfus.What Computers Still Can’t Do: A Critique of Artificial Reason

    Hubert L. Dreyfus.What Computers Still Can’t Do: A Critique of Artificial Reason. MIT Press, Cambridge, MA, 1992

    1992

  20. [20]

    Rodney A. Brooks. Intelligence without representation.Artificial Intelligence, 47(1–3):139–159, 1991. doi: 10.1016/0004-3702(91)90053-M

    doi:10.1016/0004-3702(91)90053-m 1991

  21. [21]

    Fintan D. Mallory. Wittgenstein, The Other, and Large Language Models.Filosofisk Supplement, 2(3):79–87,

  22. [22]

    AIs as fellow participants in the language game.AI & Society, 2025

    Marie Theresa O’Connor. AIs as fellow participants in the language game.AI & Society, 2025. URL: https:// link.springer.com/article/10.1007/s00146-025-02663-6,doi:10.1007/s00146-025-02663-6

    doi:10.1007/s00146-025-02663-6 2025

  23. [23]

    Technology Games: Using Wittgenstein for Understanding and Evaluating Technology

    Mark Coeckelbergh. Technology Games: Using Wittgenstein for Understanding and Evaluating Technology. Science and Engineering Ethics, 24(5):1503–1519, Aug 2017.doi:10.1007/s11948-017-9953-8

    work page doi:10.1007/s11948-017-9953-8 2017

  24. [24]

    W.J.T. Mollema. Social AI and The Equation of Wittgenstein’s Language User With Calvino’s Literature Machine,

  25. [25]

    The bewitching AI: The Illusion of Communication with Large Language Models.Philosophy & Technology, 38(2), May 2025.doi:10.1007/s13347-025-00893-6

    Emanuele Bottazzi Grifoni and Roberta Ferrario. The bewitching AI: The Illusion of Communication with Large Language Models.Philosophy & Technology, 38(2), May 2025.doi:10.1007/s13347-025-00893-6

    doi:10.1007/s13347-025-00893-6 2025

  26. [26]

    Language Games, Game Theory, and Large Language Models: A Mathematical Framework.SSRN Electronic Journal, September 2024

    Miquel Noguer i Alonso. Language Games, Game Theory, and Large Language Models: A Mathematical Framework.SSRN Electronic Journal, September 2024. URL: https://papers.ssrn.com/sol3/papers. cfm?abstract_id=4963848,doi:10.2139/ssrn.4963848

    work page doi:10.2139/ssrn.4963848 2024

  27. [27]

    Language as Mathematical Structure: Examining Semantic Field Theory Against Language Games, 2026

    Dimitris Vartziotis. Language as Mathematical Structure: Examining Semantic Field Theory Against Language Games, 2026. URL:https://arxiv.org/abs/2601.00448,arXiv:2601.00448

    arXiv 2026

  28. [28]

    Large language models and linguistic intentionality.Synthese, 2024

    Jumbly Grindrod. Large language models and linguistic intentionality.Synthese, 2024. Preprint also available on arXiv:2404.09576. URL:https://arxiv.org/abs/2404.09576,doi:10.1007/s11229-024-04723-8

    Pith/arXiv arXiv 2024

  29. [29]

    Talking about Large Language Models.Communications of the ACM, 67(2):68–79, Jan 2024

    Murray Shanahan. Talking about Large Language Models.Communications of the ACM, 67(2):68–79, Jan 2024. doi:10.1145/3624724

    doi:10.1145/3624724 2024

  30. [30]

    Brandom.Making It Explicit: Reasoning, Representing, and Discursive Commitment

    Robert B. Brandom.Making It Explicit: Reasoning, Representing, and Discursive Commitment. Harvard University Press, Cambridge, MA, 1994. 20 Asymmetric Communication E. Fenoglio

    1994

  31. [31]

    Luciano Floridi and J.W. Sanders. On the morality of artificial agents.Minds and Machines, 14(3):349–379, August 2004.doi:10.1023/b:mind.0000035461.63578.9d

    arXiv 2004

  32. [32]

    Kenneth Einar Himma. Artificial agency, consciousness, and the criteria for moral agency: What properties must an artificial agent have to be a moral agent?Ethics and Information Technology, 11:19–29, 2009. doi: 10.1007/s10676-008-9167-5

    work page doi:10.1007/s10676-008-9167-5 2009

  33. [33]

    Moral zombies: Why algorithms are not moral agents.AI & Society, 36(2):487–497, 2021

    Carissa Véliz. Moral zombies: Why algorithms are not moral agents.AI & Society, 36(2):487–497, 2021. doi:10.1007/s00146-021-01189-x

    work page doi:10.1007/s00146-021-01189-x 2021

  34. [34]

    P.M. Ukpaka. The creative agency of large language models: a philosophical inquiry.AI and Ethics, 5:2455–2466, 2025.doi:10.1007/s43681-024-00557-9

    doi:10.1007/s43681-024-00557-9 2025

  35. [35]

    Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs, 2024

    Rudolf Laine et al. Me, Myself, and AI: The Situational Awareness Dataset (SAD) for LLMs, 2024. URL: https://arxiv.org/abs/2407.04694,arXiv:2407.04694

    Pith/arXiv arXiv 2024

  36. [36]

    Language Models (Mostly) Know What They Know, 2022

    Saurav Kadavath et al. Language Models (Mostly) Know What They Know, 2022. URL: https://arxiv.org/ abs/2207.05221,arXiv:2207.05221

    Pith/arXiv arXiv 2022

  37. [37]

    Tell me about yourself: LLMs are aware of their learned behaviors, 2025

    Jan Betley, Xuchan Bao, Martín Soto, Anna Sztyber-Betley, James Chua, and Owain Evans. Tell me about yourself: LLMs are aware of their learned behaviors, 2025. URL: https://arxiv.org/abs/2501.11120, arXiv:2501.11120

    Pith/arXiv arXiv 2025

  38. [38]

    Self-Interpretability: LLMs Can Describe Complex Internal Processes that Drive Their Decisions, 2025

    Dillon Plunkett, Adam Morris, Keerthi Reddy, and Jorge Morales. Self-Interpretability: LLMs Can Describe Complex Internal Processes that Drive Their Decisions, 2025. URL: https://arxiv.org/abs/2505.17120, arXiv:2505.17120

    arXiv 2025

  39. [39]

    LLM Evaluators Recognize and Favor Their Own Generations, 2024

    Arjun Panickssery et al. LLM Evaluators Recognize and Favor Their Own Generations, 2024. URL: https: //arxiv.org/abs/2404.13076,arXiv:2404.13076

    Pith/arXiv arXiv 2024

  40. [40]

    Davidson, Viacheslav Surkov, Veniamin Veselovsky, Giuseppe Russo, Robert West, and Caglar Gulcehre

    Tim R. Davidson, Viacheslav Surkov, Veniamin Veselovsky, Giuseppe Russo, Robert West, and Caglar Gulcehre. Self-Recognition in Language Models, 2024. URL: https://arxiv.org/abs/2407.06946, arXiv:2407. 06946

    Pith/arXiv arXiv 2024

  41. [41]

    Emergent Introspective Awareness in Large Language Models, October 2025

    Jack Lindsey. Emergent Introspective Awareness in Large Language Models, October 2025. URL: https: //transformer-circuits.pub/2025/introspection/index.html

    2025

  42. [42]

    Com¸ sa and Murray Shanahan

    Iulia M. Com¸ sa and Murray Shanahan. Does It Make Sense to Speak of Introspection in Large Language Models?, 2025.arXiv:2506.05068

    Pith/arXiv arXiv 2025

  43. [43]

    Privileged Self-Access Matters for Introspec- tion in AI, 2025

    Siyuan Song, Harvey Lederman, Jennifer Hu, and Kyle Mahowald. Privileged Self-Access Matters for Introspec- tion in AI, 2025. URL:https://arxiv.org/abs/2508.14802,arXiv:2508.14802

    Pith/arXiv arXiv 2025

  44. [44]

    Binder et al

    Felix J. Binder et al. Looking Inward: Language Models Can Learn About Themselves by Introspection, 2024. URL:https://arxiv.org/abs/2410.13787,arXiv:2410.13787

    Pith/arXiv arXiv 2024

  45. [45]

    Elena Esposito - Algorithms as Communication Partners?, Aug 2022

    Institute for Contemporary Ethics. Elena Esposito - Algorithms as Communication Partners?, Aug 2022. URL: https://www.youtube.com/watch?v=reqdTgy7M70&t=22s

    2022

  46. [46]

    Mind the Gap! Bridging Explainable Artificial Intelligence and Human Understanding with Luhmann’s Functional Theory of Communication, 2024

    Bernard Keenan and Kacper Sokol. Mind the Gap! Bridging Explainable Artificial Intelligence and Human Understanding with Luhmann’s Functional Theory of Communication, 2024. URL: https://arxiv.org/abs/ 2302.03460,arXiv:2302.03460

    Pith/arXiv arXiv 2024

  47. [47]

    From meaning to emotions: LLMs as artificial communication partners.AI Soc., 41(1):171–184, July 2025.doi:10.1007/s00146-025-02481-w

    Jorge Luis Morton. From meaning to emotions: LLMs as artificial communication partners.AI Soc., 41(1):171–184, July 2025.doi:10.1007/s00146-025-02481-w

    work page doi:10.1007/s00146-025-02481-w 2025

  48. [48]

    Emergent social conventions and collective bias in LLM populations.Science Advances, 11(20):eadu9368, 2025.doi:10.1126/sciadv.adu9368

    Ariel Flint Ashery, Luca Maria Aiello, and Andrea Baronchelli. Emergent social conventions and collective bias in LLM populations.Science Advances, 11(20):eadu9368, 2025.doi:10.1126/sciadv.adu9368

    doi:10.1126/sciadv.adu9368 2025

  49. [49]

    Kegan Paul, Trench, Trubner & Co., London, 1922

    Ludwig Wittgenstein.Tractatus Logico-Philosophicus. Kegan Paul, Trench, Trubner & Co., London, 1922

    1922

  50. [50]

    Kripke.Wittgenstein on Rules and Private Language

    Saul A. Kripke.Wittgenstein on Rules and Private Language. Harvard University Press, Cambridge, MA, 1982

    1982

  51. [51]

    Hutchinson, 1949

    Gilbert Ryle.The Concept of Mind. Hutchinson, 1949

    1949

  52. [52]

    C. E. Shannon. A mathematical theory of communication.The Bell System Technical Journal, 27(3):379–423, 1948.doi:10.1002/j.1538-7305.1948.tb01338.x

    arXiv 1948

  53. [53]

    Answer engines and other communication partners.Communication Theory, 2026

    Elena Esposito. Answer engines and other communication partners.Communication Theory, 2026. doi: 10.1093/ct/qtaf036

    work page doi:10.1093/ct/qtaf036 2026

  54. [54]

    OpenCog Hyperon: A Framework for AGI at the Human Level and Beyond, 2023

    Ben Goertzel, Vitaly Bogdanov, Michael Duncan, Deborah Duong, Zarathustra Goertzel, Jan Horlings, Matthew Ikle’, Lucius Greg Meredith, Alexey Potapov, Andre’ Luiz de Senna, Hedra Seid Andres Suarez, Adam Vander- vorst, and Robert Werko. OpenCog Hyperon: A Framework for AGI at the Human Level and Beyond, 2023. URL: https://arxiv.org/abs/2310.18318,arXiv:23...

    Pith/arXiv arXiv 2023

  55. [55]

    Harvard University Press, Cambridge, MA, 1991

    Michael Dummett.The Logical Basis of Metaphysics. Harvard University Press, Cambridge, MA, 1991

    1991

  56. [56]

    Dennett.The Intentional Stance

    Daniel C. Dennett.The Intentional Stance. The MIT Press, Cambridge, MA, USA, Mar 1989

    1989

  57. [57]

    Cambridge University Press, 2013

    Scott Aaronson.Quantum Computing Since Democritus. Cambridge University Press, 2013

    2013

  58. [58]

    AI Hallucinations: A Misnomer Worth Clarifying

    Negar Maleki, Balaji Padmanabhan, and Kaushik Dutta. AI Hallucinations: A Misnomer Worth Clarifying. In2024 IEEE Conference on Artificial Intelligence (CAI), pages 133–138, 2024. doi:10.1109/CAI59869.2024.00033

    arXiv 2024

  59. [59]

    The Future of Cybercrime: AI and Emerging Technologies Are Creating a Cybercrime Tsunami.SSRN Electronic Journal, 2023.doi:10.2139/ssrn.4507244

    Philip Treleaven, Jeremy Barnett, Daniel Brown, Andrew Bud, Enzo Fenoglio, Charles Kerrigan, Adriano Koshiyama, Sally Sfeir-Tait, and Martin Schoernig. The Future of Cybercrime: AI and Emerging Technologies Are Creating a Cybercrime Tsunami.SSRN Electronic Journal, 2023.doi:10.2139/ssrn.4507244

    work page doi:10.2139/ssrn.4507244 2023

  60. [60]

    Survey of Hallucination in Natural Language Generation.ACM Computing Surveys, 55(12):1–38, 2023.doi:10.1145/3571730

    Ziwei Ji, Nayeon Lee, Rita Frieske, Tianyi Yu, Dan Su, Yi Xu, Etsuko Ishii, Yejin Bang, Andrea Madotto, and Pascale Fung. Survey of Hallucination in Natural Language Generation.ACM Computing Surveys, 55(12):1–38, 2023.doi:10.1145/3571730

    doi:10.1145/3571730 2023

  61. [61]

    Hallucination is Inevitable: An Innate Limitation of Large Language Models, 2025

    Ziwei Xu, Sanjay Jain, and Mohan Kankanhalli. Hallucination is Inevitable: An Innate Limitation of Large Language Models, 2025. URL:https://arxiv.org/abs/2401.11817,arXiv:2401.11817

    Pith/arXiv arXiv 2025

  62. [62]

    THaMES: An End-to-End Tool for Hallucination Mitigation and Evaluation in Large Language Models

    Mengfei Liang, Archish Arun, Zekun Wu, Cristian Munoz, Jonathan Lutch, Emre Kazim, Adriano Koshiyama, and Philip Treleaven. THaMES: An End-to-End Tool for Hallucination Mitigation and Evaluation in Large Language Models.arXiv preprint arXiv:2409.11353, 2024.doi:10.48550/arXiv.2409.11353

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2409.11353 2024

  63. [63]

    The Logic of Tacit Inference.Philosophy, 41(155):1–18, 1966

    Michael Polanyi. The Logic of Tacit Inference.Philosophy, 41(155):1–18, 1966. doi:10.1017/ S0031819100066110

    1966

  64. [64]

    Tacit Knowledge: A Wittgensteinian Approach.Tradition and Discovery, 33(3):9–25, 2006

    Zhenhua Yu. Tacit Knowledge: A Wittgensteinian Approach.Tradition and Discovery, 33(3):9–25, 2006

    2006

  65. [65]

    Polanyi’s revenge and AI’s new romance with tacit knowledge.Commun

    Subbarao Kambhampati. Polanyi’s revenge and AI’s new romance with tacit knowledge.Commun. ACM, 64(2):31–32, jan 2021.doi:10.1145/3446369

    doi:10.1145/3446369 2021

  66. [66]

    vibe coding

    Andrej Karpathy. There’s a new kind of coding I call "vibe coding". Post on X (Twitter), February 2025. URL: https://x.com/karpathy/status/1886192184808149383?lang=en

    arXiv 2025

  67. [67]

    Why Artificial Intelligence Needs Sociology of Knowledge: Parts I and II.AI and Society, 40(3):1249–1263, 2025.doi:10.1007/s00146-024-01954-8

    Harry Collins. Why Artificial Intelligence Needs Sociology of Knowledge: Parts I and II.AI and Society, 40(3):1249–1263, 2025.doi:10.1007/s00146-024-01954-8

    doi:10.1007/s00146-024-01954-8 2025

  68. [68]

    Group of Governmental Experts on Lethal Autonomous Weapons Systems (LAWS)

    United Nations Convention on Certain Conventional Weapons (CCW). Group of Governmental Experts on Lethal Autonomous Weapons Systems (LAWS). United Nations Office for Disarmament Affairs (UNODA),

  69. [69]

    Department of Defense

    U.S. Department of Defense. Directive 3000.09: Autonomy in Weapon Systems, 2023. Updated versions 2023–2024. URL: https://www.esd.whs.mil/Portals/54/Documents/DD/issuances/dodd/300009p. PDF?ver=e0YrG458bVDl3-oyAOJjOw%3d%3d

    2023

  70. [70]

    Department of Defense

    U.S. Department of Defense. Ethical Principles for Artificial Intelligence, 2020. URL: https://www.defense.gov/News/Releases/Release/Article/2091996/ dod-adopts-ethical-principles-for-artificial-intelligence/

    arXiv 2020

  71. [71]

    Anthropic rejects Pentagon demands over AI safeguards. 2026. Dispute concerning military use, surveil- lance, and autonomous weapons. URL: https://www.reuters.com/sustainability/society-equity/ anthropic-rejects-pentagons-requests-ai-safeguards-dispute-ceo-says-2026-02-26/

    2026

  72. [72]

    The Pentagon/Anthropic Clash Over Military AI Guardrails, February 2026

    Opinio Juris. The Pentagon/Anthropic Clash Over Military AI Guardrails, February 2026. URL: https:// opiniojuris.org/2026/02/26/the-pentagon-anthropic-clash-over-military-ai-guardrails/

    2026

  73. [73]

    Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.arXiv, 2021

    Patrick Lewis, Ethan Perez, Aleksandra Piktus, Fabio Petroni, Vladimir Karpukhin, Naman Goyal, Heinrich Küttler, Mike Lewis, Wen-tau Yih, Tim Rocktäschel, Sebastian Riedel, and Douwe Kiela. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks.arXiv, 2021. URL: https://arxiv.org/abs/2005.11401, arXiv:2005.11401

    Pith/arXiv arXiv 2021

  74. [74]

    Jules White, Quchen Fu, Sam Hays, Michael Sandborn, Carlos Olea, Henry Gilbert, Ashraf Elnashar, Jesse Spencer-Smith, and Douglas C. Schmidt. A Prompt Pattern Catalog to Enhance Prompt Engineering with ChatGPT, 2023. URL:https://arxiv.org/abs/2302.11382,arXiv:2302.11382

    Pith/arXiv arXiv 2023

  75. [75]

    Building Guardrails for Large Language Models, 2024

    Yi Dong, Ronghui Mu, Gaojie Jin, Yi Qi, Jinwei Hu, Xingyu Zhao, Jie Meng, Wenjie Ruan, and Xiaowei Huang. Building Guardrails for Large Language Models, 2024. URL: https://arxiv.org/abs/2402.01822, arXiv:2402.01822. 22 Asymmetric Communication E. Fenoglio

    Pith/arXiv arXiv 2024

  76. [76]

    ReAct: Synergizing Reasoning and Acting in Language Models, 2023

    Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. ReAct: Synergizing Reasoning and Acting in Language Models, 2023. URL: https://arxiv.org/abs/2210.03629, arXiv:2210.03629. 23

    Pith/arXiv arXiv 2023

  77. [2023]

    URL: https://fintanmallory.com/wp-content/uploads/2023/10/ wittgenstein-the-other-and-large-language-models-1.pdf

    Author PDF available online. URL: https://fintanmallory.com/wp-content/uploads/2023/10/ wittgenstein-the-other-and-large-language-models-1.pdf

    2023

  78. [2024]

    URL:https://arxiv.org/abs/2407.09493,arXiv:2407.09493

    Pith/arXiv arXiv

  79. [2026]

    URL: https://docs-library

    Ongoing international discussions on regulation of autonomous weapons. URL: https://docs-library. unoda.org/Convention_on_Certain_Conventional_Weapons_-Group_of_Governmental_Experts_ on_Lethal_Autonomous_Weapons_Systems_(2026)/CCW-GGE.1-2026-WP.2.pdf

    2026