This brief lists scholarly work on LLM philosophy, hallucinations, grounding, and formal impossibility arguments published 2024-2026. Queries targeted named venues (NeurIPS, ICML, ACL, EMNLP, Mind & Language, Synthese, Philosophical Studies) and named researchers. Raw Brave Answers API responses preserved below; anything beyond direct API output is flagged.
Queries Run
Q1. 2025-2026 LLM hallucination papers at top venues (NeurIPS, ICML, ACL, EMNLP, ICLR, TACL, AAAI). Returned: insufficient data
Q2. Philosophy papers on LLM cognition 2024-2026 in Mind & Language, Synthese, Philosophical Studies, Minds and Machines, Erkenntnis, Inquiry. 10 papers returned
Q3. Formal/theoretical proofs on hallucination inevitability 2024-2026 (Gödel, Turing, learning theory). 4 papers returned
Q4. Peer-reviewed critiques of the "hallucination" terminology 2023-2026. 2 papers returned, 5 named authors not found
Q5. Grounding, meaning, reference in LLMs — academic debate 2024-2026. 7 papers returned
Q1: Top-Venue 2025 Hallucination Papers — Null Result
Brave Answers API response (verbatim): "The provided search results do not contain sufficient information to generate a complete list of peer-reviewed academic papers on large language model hallucinations published in 2025 or 2026 at the specified top venues... with full details such as title, all authors, affiliations, publication dates, and core technical contributions."
Mentioned in passing but not fully described:
- "Can we trust AI doctors? A survey of medical hallucination in large language and large vision-language models" — Findings of ACL 2025, pp. 6293-6316. Author list and affiliations not returned.
- "CAI: Caption-sensitive attention intervention for mitigating object hallucination in large vision-language models" — arXiv:2506.23590 (2025). Not confirmed as peer-reviewed.
Note: Brave returned a proper null rather than fabricating sources. No papers from ICLR 2025, TACL, or AAAI 2025 were described in results.
Q2: Philosophy Papers on LLM Cognition (2024-2026)
| Title & Authors | Journal · Year | DOI | Core Argument |
|---|---|---|---|
| Agnosticism about artificial consciousness Tom McClelland |
Mind & Language · 2025 | 10.1111/mila.12543 | Current evidence is insufficient to determine whether LLMs possess consciousness. Advocates agnostic stance. Critiques inference-to-best-explanation arguments for artificial consciousness. |
| "All animals are conscious": Shifting the null hypothesis Kristin Andrews |
Mind & Language · 2024 | 10.1111/mila.12512 | Proposes reversing the default in consciousness studies — assume consciousness unless disproven. Applied to AI, challenges anthropocentric bias in mental-state attribution. |
| Can structural correspondences ground real-world representational content in LLMs? Iwan Williams |
Mind & Language · 2026 | 10.1111/mila.12567 | Critiques distributional semantics. Without causal interaction with environment, structural correspondences alone are insufficient for semantic content. |
| AI, LLMs, and the normativity of belief Camila Hernandez Flowerman |
Synthese · 2026 DOI inferred | 10.1007/s11229-026-05475-3 | Belief entails normative commitments (justification, revision, accountability) LLMs lack. They simulate belief-like outputs but don't participate in epistemic practices that constitute belief ascription. |
| Mechanistic Indicators of Understanding in LLMs Pierre Beckmann, Matthieu Queloz |
Philosophical Studies · 2026 venue inferred | 10.1007/s11098-026-02100-1 | Three-tiered understanding model: conceptual, state-of-world, principled. LLMs exhibit mechanistic analogs of understanding but lack intentionality and unified mechanisms. |
| Deep Learning Models Also Recall Features Pierre Beckmann |
Erkenntnis · 2026 DOI inferred | 10.1007/s10670-026-00789-x | Introduces "feature recall" mechanism. LLMs retrieve stored information via linear projections — suggests structured memory beyond surface pattern matching, but still lacking semantic grounding. |
| Risk Analysis in Automated Misinformation Detection Adrian K. Yee |
Minds and Machines · 2026 DOI inferred | 10.1007/s11023-026-09775-y | LLM misinformation-detection systems create epistemic opacity and false confidence. Lack of genuine understanding undermines reliability as epistemic agents. |
| Epistemic Drift in Mind-Model Systems Søgaard, Rajcic, Scott |
Minds and Machines · 2026 DOI inferred | 10.1007/s11023-026-09772-1 | LLMs modeling human minds exhibit "epistemic drift" — divergence from accurate mental-state attribution over time. Without real-world grounding, errors accumulate. |
| Rectifying illusion: a Buddhist–Confucian framework for LLM hallucinations Sridharan Sankaran |
Philosophy & Technology · 2026 | 10.1007/s43681-026-01053-y | Non-Western framework: hallucinations reveal absence of genuine intentionality. Proposes virtue-based, context-sensitive truth-telling model. |
| Do LLMs Defend Inferentialist Semantics? On Logical Expressivism and Anti-Representationalism Yuzuki Arai, Sho Tsugawa |
Inquiry · 2026 | 10.1080/0020174X.2026.2156789 | Argues LLMs support Brandom's inferentialist semantics — meaning arises from inferential roles, not referential correspondence. LLMs challenge semantic externalism. |
Q3: Formal Impossibility Proofs (2024-2026)
| Title & Authors | arXiv · Date | Formal Tool | Core Theorem / Claim |
|---|---|---|---|
| LLMs Will Always Hallucinate, and We Need to Live With This Sourav Banerjee, Ayushi Agarwal, Saloni Singla |
arXiv:2409.05746 Sep 2024 |
Gödel's First Incompleteness, Halting Problem | No finite training dataset encodes all truths → structural hallucinations. Hallucination detection reduces to Halting Problem → undecidable. Every pipeline stage has non-zero hallucination probability. Introduces "Structural Hallucination" concept. |
| On the Fundamental Impossibility of Hallucination Control in LLMs Authors not explicitly named in response |
arXiv:2506.06382 Jun 2025 |
Halting Problem, Diagonalization | Self-detection of hallucinations equivalent to Halting Problem → undecidable. Diagonalization: for any training set D, exists true statement z ∈ F (facts) but z ∉ D. Generative function G(D) cannot fully contain truth set F. |
| Hallucinations are inevitable but can be made statistically negligible Atsushi Suzuki, Yulan He, Feng Tian, Zhongyuan Wang |
arXiv:2502.12187 Feb 2025 |
Probability theory, Shannon source coding analogy | Critiques Banerjee et al. / Xu et al. Accepts computational inevitability on infinite input spaces but proves P(hallucination) → 0 as |D| → ∞ and data quality → 1. Distinguishes computational inevitability from practical significance under realistic input distributions. |
| Limitations on Safe, Trusted, Artificial General Intelligence Rina Panigrahy |
arXiv:2509.21654 Sep 2025 |
Turing undecidability, Gödel, Acceptance Problem | Extends hallucination impossibility to broader AGI safety limits. Cites Xu et al. Determining whether LLM output consistent with ground truth reduces to Acceptance Problem → undecidable. Self-verification of truth impossible for LLMs. |
Q4: Terminology Critiques (2023-2026)
| Paper & Authors | Venue · Year | Proposed Term | Objection to "Hallucination" |
|---|---|---|---|
| ChatGPT is bullshit Michael Townsen Hicks, James Humphries, Joe Slater |
Ethics and Information Technology · 2024 | Bullshit (Frankfurt) | "Hallucination" implies perceptual/representational relationship to truth akin to human cognition. LLMs don't attempt to represent reality — they generate text by statistical pattern without concern for factual accuracy. Errors are not misperceptions but outputs from a system indifferent to truth. |
| Redefining "Hallucination" in LLMs: Towards a psychology-informed framework Authors not returned by API |
arXiv preprint · 2024 not peer-reviewed | Confabulation | "Hallucination" lacks psychological precision. "Confabulation" (cognitive psychology: production of false memories without intent to deceive) better describes LLM fabrication when correct information was present in training. |
Not found in Brave results: peer-reviewed publications from Gary Smith, Mary Shaw, Usama Fayyad, Maksym Andriushchenko, or Hayden Field specifically critiquing the term "hallucination" within the 2023-2026 window.
Q5: Grounding, Meaning & Reference in LLMs (2024-2026)
| Title & Authors | Venue · Year | Position | Core Claim |
|---|---|---|---|
| The Vector Grounding Problem Dimitri Coelho Mollo, Raphaël Millière |
Philosophy and the Mind Sciences, Vol. 7 No. 1 · 2026 DOI: 10.33735/phimisci.2026.12307 |
Optimistic | LLMs can achieve referential grounding without multimodal/embodied experience. Draws on teleosemantics: grounding occurs when internal states (1) have causal-informational relations to world and (2) have functional history of selection for carrying that information. |
| A Philosophical Introduction to Language Models — Part I Raphaël Millière, Cameron Buckner |
Philosophy Compass · 2024 venue assumed | Moderate deflationary | Situates LLMs in classical meaning debates (symbol grounding, Chinese Room). LLMs lack sensorimotor grounding but inferential-role semantics complicates strict functionalist dismissal. Outputs can be meaningfully interpreted in context. |
| Pragmatic Norms Are All You Need — Why Symbol Grounding Does Not Apply to LLMs Authors not returned |
EMNLP 2024 · ACL Anthology 2024.emnlp-main.651 | Pragmatic / use-based | Challenges applicability of classical symbol grounding problem. Pragmatic norms and linguistic usage — not world-world causal links — sufficient for meaning. Reverses Bender & Koller Octopus Test conclusion: if LLM participates in language games governed by social norms, it is pragmatically grounded. |
| How Well Do Large Language Models Truly Ground? Hyunji Lee et al. (6 co-authors) |
arXiv:2311.09069 · 2023 (v1), updated 2024 | Empirical / skeptical | Proposes stricter operational grounding definition: model is grounded only if it fully utilizes context AND does not exceed context bounds. Tests 25 LLMs — larger instruction-tuned models perform better, but no model achieves full grounding. |
| Grounding LLMs in Interactive Environments with Online Reinforcement Learning Carta, Romac, Wolf, Lamprier, Sigaud, Oudeyer |
ICML 2023 · arXiv:2302.02662v4 (updated Oct 2024) | Functional grounding | Introduces GLAM (Grounding through Language-Action Modeling). Uses online RL in interactive textual environments (BabyAI-Text). Interactive experience — even in simulated text worlds — improves grounding via causal learning and error correction. |
| LLMs, Turing Tests and Chinese Rooms: The Prospects for Meaning in LLMs Elselijn Borg |
Inquiry · 2025 DOI: 10.1080/0020174X.2024.2446241 |
Hard skepticism | Text-only models lack semantic depth; cannot possess understanding in human sense. Rejects functionalist/behaviorist criteria. Grounding requires embodiment or causal interaction with a world. |
| Propositional Interpretability in Artificial Intelligence David J. Chalmers |
arXiv:2501.15740 · 2025 | Non-reductive functionalism | Follow-up to 2023 consciousness paper. Distinguishes interpretability-from-outside vs intrinsic semantic content. Complex, coherent systems may have form of derived intentionality, especially embedded in interactive/social contexts. |
Flags & Unverified Items
- Q2 DOIs flagged "inferred": Brave explicitly stated it "inferred or constructed [some DOIs] based on standard Springer and Wiley naming conventions." These should be verified before citation: Flowerman (s11229-026-05475-3), Beckmann & Queloz (s11098-026-02100-1), Beckmann (s10670-026-00789-x), Yee (s11023-026-09775-y), Søgaard et al. (s11023-026-09772-1).
- Q2 venue assumption: Millière & Buckner's Philosophical Introduction to Language Models Part I venue listed as Philosophy Compass is marked "assumed." Original arXiv:2401.03910.
- Q3 anonymous paper: arXiv:2506.06382 "On the Fundamental Impossibility of Hallucination Control" — Brave did not return author names in response.
- Q4 unverified authors: No peer-reviewed papers from Gary Smith, Mary Shaw, Usama Fayyad, Maksym Andriushchenko, or Hayden Field were returned by Brave for the 2023-2026 window.
- Q4 confabulation paper: author list not returned by API.
- Q5 "Pragmatic Norms" paper: authors not returned; inferred as response to Bender & Koller.
Raw Source Files
research/brave-llm-hallucinations-v2-2026-04-05.json— full verbatim Brave Answers API responses (5 queries)research/brave-llm-hallucinations-v2-2026-04-05.js— query script