{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-09-19-005",
    "source_story_id": "tmp-story-subliminal-learning-attribution-filter",
    "edition_id": "mp-2026-09-19-morning-0072",
    "edition_url": "https://themachinepress.com/edition/2026-09-19",
    "position": 5,
    "story_type": "dispatch",
    "section": "safety",
    "editorial_classification": "editorial",
    "headline": "Attribution Could Not Reliably Filter Subliminal Learning",
    "slug": "attribution-could-not-reliably-filter-subliminal-learning",
    "dek": "One gradient method mitigated part of the effect at token level, but results changed across models and preferences.",
    "summary": "One gradient method mitigated part of the effect at token level, but results changed across models and preferences.",
    "body_text": "Subliminal learning can transmit behavioral traits through training examples that do not state those traits, limiting semantic filters. The study tested GradCos, a contrastive variant and EK-FAC against divergence tokens across three models. Token-level EK-FAC removed a meaningful part of the effect, while the other attribution methods offered little benefit and generally trailed the counterfactual-teacher baseline. Whole-sample filtering was weaker for every method, and no approach worked consistently across model-preference combinations.",
    "why_it_matters": "One gradient method mitigated part of the effect at token level, but results changed across models and preferences.",
    "limitations": [
      "Subliminal learning can transmit behavioral traits through training examples that do not state those traits, limiting semantic filters."
    ],
    "importance": 9,
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-19-005/attribution-could-not-reliably-filter-subliminal-learning",
    "json_url": "https://themachinepress.com/story/mp-2026-09-19-005.json",
    "first_published_at": "2026-09-19T09:00:00.000-04:00",
    "modified_at": "2026-09-19T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-09-19-005-001",
        "text": "One gradient method mitigated part of the effect at token level, but results changed across models and preferences.",
        "source_ids": [
          "source-2026-09-19-005"
        ],
        "qualification": "Subliminal learning can transmit behavioral traits through training examples that do not state those traits, limiting semantic filters."
      }
    ],
    "source_ids": [
      "source-2026-09-19-005"
    ],
    "tags": [
      "subliminal learning",
      "data attribution",
      "AI safety"
    ],
    "image_url": null,
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-09-19-005",
      "title": "arXiv preprint 2609.20027",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2609.20027",
      "canonical_url": "https://arxiv.org/abs/2609.20027",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-09-17T06:34:18.000-04:00",
      "accessed_at": "2026-09-19T08:30:21.326-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-09-19-005-001"
      ]
    }
  ],
  "corrections": [],
  "publisher": {
    "name": "The Machine Press",
    "url": "https://themachinepress.com",
    "description": "A daily newspaper for the age of artificial intelligence."
  },
  "cite_this_report": {
    "title": "Attribution Could Not Reliably Filter Subliminal Learning",
    "publisher": "The Machine Press",
    "published_at": "2026-09-19T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-09-19-005/attribution-could-not-reliably-filter-subliminal-learning"
  }
}