{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-03-004",
    "source_story_id": "tmp-story-linguistic-illegibility-security",
    "edition_id": "mp-2026-09-03-morning-0056",
    "edition_url": "https://themachinepress.com/edition/2026-09-03",
    "position": 4,
    "story_type": "dispatch",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "The Model's Explanation May Not Describe Its Computation",
    "slug": "the-model-s-explanation-may-not-describe-its-computation",
    "dek": "A security argument says linguistic monitoring cannot supply complete guarantees when internal activation-space computation is only lossy-translated into words.",
    "summary": "A security argument says linguistic monitoring cannot supply complete guarantees when internal activation-space computation is only lossy-translated into words.",
    "body_text": "James Mickens names the gap “linguistic illegibility”: external text or language-like probes may fail to represent how a model produced an outcome. The paper argues that chain-of-thought monitoring, self-critique and linguistically defined activation probes therefore need a non-linguistic security floor. It proposes output taint tracking, robust virtualization and independent configuration audits as complementary isolation measures; this is a position and design argument, not a measured proof that every monitor fails.",
    "why_it_matters": "A security argument says linguistic monitoring cannot supply complete guarantees when internal activation-space computation is only lossy-translated into words.",
    "limitations": [
      "It proposes output taint tracking, robust virtualization and independent configuration audits as complementary isolation measures; this is a position and design argument, not a measured proof that every monitor fails."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-03-004/the-model-s-explanation-may-not-describe-its-computation",
    "json_url": "https://themachinepress.com/story/mp-2026-09-03-004.json",
    "first_published_at": "2026-09-03T09:00:00.000-04:00",
    "modified_at": "2026-09-03T14:20:28.218-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-03-004-001",
        "text": "A security argument says linguistic monitoring cannot supply complete guarantees when internal activation-space computation is only lossy-translated into words.",
        "source_ids": [
          "source-2026-09-03-004"
        ],
        "qualification": "It proposes output taint tracking, robust virtualization and independent configuration audits as complementary isolation measures; this is a position and design argument, not a measured proof that every monitor fails."
      }
    ],
    "source_ids": [
      "source-2026-09-03-004"
    ],
    "tags": [
      "LLM security",
      "monitoring",
      "sandboxing"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-03-004",
      "title": "arXiv preprint 2609.02852",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.02852",
      "canonical_url": "https://arxiv.org/abs/2609.02852",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-02T13:37:22.000-04:00",
      "accessed_at": "2026-09-03T14:20:28.218-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-03-004-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "The Model's Explanation May Not Describe Its Computation",
    "publisher": "The Machine Press",
    "published_at": "2026-09-03T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-03-004/the-model-s-explanation-may-not-describe-its-computation"
  }
}