{
  "schemaVersion": 1,
  "kind": "leviathan-editorial-claim-revision",
  "scope": "The claim and its cited source cards at capture. External papers and linked project pages are not frozen by this record. A digest identifies bytes, not truth or endorsement.",
  "claim": {
    "id": "C01",
    "title": "What does human language carry into AI?",
    "statement": "Some learned internal representations associated with human concepts can causally shape model behavior.",
    "kind": "Empirical question",
    "status": "Supported within a defined scope",
    "summary": "Human language carries ways of interpreting the world: care, fear, ambition, prejudice, and cooperation. Research offers ways to investigate how some concepts become working structures inside models. Leviathan asks how communities could examine that inheritance and test whether explicit, contextual relationships among their concepts and values help agents act more consistently.",
    "definitions": [
      {
        "term": "Internal representation",
        "meaning": "A pattern of model activity associated with information the model uses."
      },
      {
        "term": "Causal intervention",
        "meaning": "Changing an internal pattern and testing whether the model's behavior changes."
      }
    ],
    "values": [
      "Understand the systems that increasingly shape our lives.",
      "Make the influence of inherited human concepts open to scrutiny."
    ],
    "reasoning": [
      "If a learned concept changes a model's decisions, understanding its role matters for both capability and safety.",
      "This connects the study of language and culture to mechanisms we can investigate. A human name for a representation remains a hypothesis about its function.",
      "Leviathan proposes local meaning kernels that connect concepts, principles, rules, and their versions. Those explicit records are a design layer; they are distinct from learned internal model representations.",
      "A proposed test would ask whether that context changes interpretation and action on unfamiliar cases, including when a community revises a concept or disagrees with another community. The cited findings motivate this question; they do not answer it."
    ],
    "limits": "The findings concern particular concepts, models and experimental settings. They do not establish that all culture is contained in language or that every model operation can be reduced to a human concept.",
    "strongestObjection": "A representation extracted using human labels may partly reflect the researcher's choice of categories. An intervention can change behavior without proving that a model uses the concept as a person does. Similarly, supplying a detailed value context could produce fluent agreement without reliable changes in action.",
    "whatWouldChangeOurView": "Independent interventions that fail to reproduce the reported effects, or controls showing that unrelated directions explain them equally well, would weaken the claim. Replication across models, languages and contexts would strengthen it.",
    "nextQuestion": "Which effects survive different concept lists, model families, languages, and post-training methods? For our proposed local kernels, compare ordinary instructions with explicit concept–principle relationships at a comparable total budget. Test behavior when meanings change, values conflict, and unfamiliar cases appear; record failures as well as agreement.",
    "evidence": [
      {
        "sourceId": "R-AI-01",
        "relation": "supports",
        "reason": "Interventions on emotion-related representations change measured behavior."
      },
      {
        "sourceId": "R-AI-06",
        "relation": "supports-part",
        "reason": "A different base model partly reproduces the representation geometry; causal behavioral effects were not tested."
      },
      {
        "sourceId": "R-AI-09",
        "relation": "supports",
        "reason": "Minimal reward training recruits pre-existing representations that affect behavior."
      },
      {
        "sourceId": "R-AI-03",
        "relation": "context",
        "reason": "Some models have limited access to their internal representations; this does not validate every self-report."
      }
    ],
    "history": [
      {
        "date": "2026-09-30",
        "note": "Added the functional-representation claim with a partial replication and its limits. Clarified that 171 is the study's preselected concept list."
      },
      {
        "date": "2026-10-02",
        "note": "Connected the functional-representation findings to a proposed test of local meaning and value contexts. Distinguished explicit kernels from learned model representations and verbal agreement from behavior. Empirical statement, support status, and source reviews are unchanged."
      }
    ],
    "projectLinks": [
      {
        "label": "The founding letter",
        "href": "/whitenote"
      },
      {
        "label": "Local meaning kernels",
        "href": "/what-is"
      },
      {
        "label": "The language and learning work",
        "href": "/work"
      }
    ]
  },
  "sources": [
    {
      "id": "R-AI-01",
      "title": "Emotion Concepts and their Function in a Large Language Model",
      "url": "https://arxiv.org/abs/2604.07729",
      "authors": "Nicholas Sofroniew and 15 co-authors; Anthropic",
      "publishedAt": "2026-04-02",
      "version": "Research announcement: 2 April 2026; arXiv v1: 9 April 2026",
      "type": "Preprint",
      "reviewLevel": "Author summary",
      "finding": "Vectors associated with 171 preselected emotion concepts were extracted from Claude Sonnet 4.5. Interventions changed preferences and some alignment-related behaviors, supporting a functional role for these learned concepts.",
      "limitations": "The 171 concepts are a starting list, not 171 discovered feeling centers. The representations mainly track local context. The blackmail experiment used an early, unreleased snapshot; subjective feeling was not established.",
      "changes": "The geometry has a partial replication in another model in R-AI-06. That replication does not repeat the original behavioral interventions.",
      "checkedAt": "2026-09-30",
      "verificationStatus": "Primary source reviewed",
      "relatedUrls": [
        "https://www.anthropic.com/research/emotion-concepts-function",
        "https://transformer-circuits.pub/2026/emotions/index.html"
      ]
    },
    {
      "id": "R-AI-06",
      "title": "Replicating the Geometry of Emotion Representations in a Base Open-Weights Model",
      "url": "https://arxiv.org/html/2609.22208v1",
      "authors": "Adam Hollowell; University of North Carolina at Chapel Hill",
      "publishedAt": "2026-09-01",
      "version": "arXiv:2609.22208v1; submission date listed as 1 September 2026",
      "type": "Preprint",
      "reviewLevel": "Selected sections",
      "finding": "In base Gemma-2-27B, much of the valence geometry and clustering of 171 emotion vectors is recovered. Some structure is already present in token embeddings. The arousal axis does not meet all stability criteria.",
      "limitations": "This supports the representation finding, not a causal effect on behavior. The study uses Claude-generated fiction, one different model, reconstructed methods and a linear representation assumption.",
      "changes": "This is a partial geometric replication. The author does not count arousal as fully replicated and identifies structural-token confounds in some measurements.",
      "checkedAt": "2026-09-30",
      "verificationStatus": "Primary source reviewed",
      "relatedUrls": [
        "https://arxiv.org/abs/2609.22208"
      ]
    },
    {
      "id": "R-AI-09",
      "title": "How’s it going? Reinforcement learning in language models recruits a functional welfare axis",
      "url": "https://arxiv.org/abs/2605.30232",
      "authors": "Andy Q Han, David J. Chalmers and Pavel Izmailov",
      "publishedAt": "2026-05-28",
      "version": "arXiv v1",
      "type": "Preprint",
      "reviewLevel": "Abstract",
      "finding": "Reward and punishment vectors extracted after maze training influence behavior in other tasks. The authors' controls support the view that training recruits pre-existing representations rather than creating them from scratch.",
      "limitations": "Functional welfare here means an estimate of doing well or badly relative to goals. It is not a claim about experienced pleasure or pain. This review covers the abstract.",
      "changes": "The study adds evidence about functional internal representations. Its publication date is May 2026, even when later coverage draws attention to it.",
      "checkedAt": "2026-09-30",
      "verificationStatus": "Primary source reviewed",
      "relatedUrls": []
    },
    {
      "id": "R-AI-03",
      "title": "Emergent Introspective Awareness in Large Language Models",
      "url": "https://transformer-circuits.pub/2025/introspection/index.html",
      "authors": "Jack Lindsey; Anthropic",
      "publishedAt": "2025-10-29",
      "version": "First published: 29 October 2025; web revision: 1 January 2026; arXiv v1: 5 January 2026",
      "type": "Preprint",
      "reviewLevel": "Selected sections",
      "finding": "Some Claude models can identify injected concepts in certain settings and distinguish earlier internal representations from input text. Controls support a limited form of functional introspection.",
      "limitations": "Failures are common, and performance depends on context and post-training. The experiments do not establish human-like introspection, the reliability of every self-report, or subjective experience.",
      "changes": "The January web revision adds a control experiment and prompt corrections. The later arXiv submission is a separate publication event, not an independent replication.",
      "checkedAt": "2026-09-30",
      "verificationStatus": "Primary source reviewed",
      "relatedUrls": [
        "https://arxiv.org/abs/2601.01828",
        "https://www.anthropic.com/research/introspection"
      ]
    }
  ]
}
