{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-17-004",
    "source_story_id": "tmp-story-probe-decodability-causality",
    "edition_id": "mp-2026-09-17-morning-0070",
    "edition_url": "https://themachinepress.com/edition/2026-09-17",
    "position": 4,
    "story_type": "dispatch",
    "section": "frontier-models",
    "editorial_classification": "editorial",
    "headline": "The Truth Probe Read Features the Model Barely Used",
    "slug": "the-truth-probe-read-features-the-model-barely-used",
    "dek": "Probe alignment and behavioral sensitivity overlapped by about 12% in one Gemma 2 deception setting.",
    "summary": "Probe alignment and behavioral sensitivity overlapped by about 12% in one Gemma 2 deception setting.",
    "body_text": "A linear probe can decode truthfulness without identifying the features that actually drive an answer. Decomposing one deployed truth probe into sparse-autoencoder features produced only about 12% overlap between geometric alignment and behavioral gradient sensitivity. Ablating features shared by the probe and the model flipped outputs as much as 27%, against 6% for equally sized probe-only sets and 1% for random sets at full coherence. The result held across five seeds and a held-out split, but it is evidence from one model and experimental setting, not a universal account of probe causality.",
    "why_it_matters": "Probe alignment and behavioral sensitivity overlapped by about 12% in one Gemma 2 deception setting.",
    "limitations": [
      "Decomposing one deployed truth probe into sparse-autoencoder features produced only about 12% overlap between geometric alignment and behavioral gradient sensitivity.",
      "Ablating features shared by the probe and the model flipped outputs as much as 27%, against 6% for equally sized probe-only sets and 1% for random sets at full coherence.",
      "The result held across five seeds and a held-out split, but it is evidence from one model and experimental setting, not a universal account of probe causality."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-17-004/the-truth-probe-read-features-the-model-barely-used",
    "json_url": "https://themachinepress.com/story/mp-2026-09-17-004.json",
    "first_published_at": "2026-09-17T09:00:00.000-04:00",
    "modified_at": "2026-09-17T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-17-004-001",
        "text": "Probe alignment and behavioral sensitivity overlapped by about 12% in one Gemma 2 deception setting.",
        "source_ids": [
          "source-2026-09-17-004"
        ],
        "qualification": "Decomposing one deployed truth probe into sparse-autoencoder features produced only about 12% overlap between geometric alignment and behavioral gradient sensitivity."
      }
    ],
    "source_ids": [
      "source-2026-09-17-004"
    ],
    "tags": [
      "interpretability",
      "causal intervention",
      "linear probes"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-17-004",
      "title": "arXiv preprint 2609.18080",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.18080",
      "canonical_url": "https://arxiv.org/abs/2609.18080",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-15T23:35:38.000-04:00",
      "accessed_at": "2026-09-17T08:28:34.205-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-17-004-001"
      ]
    }
  ],
  "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 Truth Probe Read Features the Model Barely Used",
    "publisher": "The Machine Press",
    "published_at": "2026-09-17T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-17-004/the-truth-probe-read-features-the-model-barely-used"
  }
}