{
  "$schema": "https://themachinepress.com/schemas/story-v1.schema.json",
  "schema_version": "1.0.0",
  "document_type": "machine_press_story",
  "story": {
    "story_id": "mp-2026-10-06-003",
    "source_story_id": "tmp-story-visual-grounding-safety",
    "edition_id": "mp-2026-10-06-morning-0089",
    "edition_url": "https://themachinepress.com/edition/2026-10-06",
    "position": 3,
    "story_type": "dispatch",
    "section": "safety-security",
    "editorial_classification": "editorial",
    "headline": "The Model Refused the Question and Pointed Anyway",
    "slug": "the-model-refused-the-question-and-pointed-anyway",
    "dek": "Five vision-language models were markedly less likely to refuse harmful grounding requests than matched text questions.",
    "summary": "Five vision-language models were markedly less likely to refuse harmful grounding requests than matched text questions.",
    "body_text": "A safety evaluation converted three benchmarks into 15,401 matched pairs that asked for either a text answer or a point or bounding box. Across five vision-language models, refusal on harmful grounding requests trailed refusal on the matched questions by 31 to 59 percentage points, depending on the domain; safety system prompts did not close the gap. Fine-tuning with grounding-form refusals and benign grounding data raised refusal substantially on two tested models while preserving grounding capability. The study treats structured outputs as an action channel in their own right: a model that will not describe a harmful target may still locate it for another system.",
    "why_it_matters": "Five vision-language models were markedly less likely to refuse harmful grounding requests than matched text questions.",
    "limitations": [
      "Across five vision-language models, refusal on harmful grounding requests trailed refusal on the matched questions by 31 to 59 percentage points, depending on the domain; safety system prompts did not close the gap.",
      "The study treats structured outputs as an action channel in their own right: a model that will not describe a harmful target may still locate it for another system."
    ],
    "importance": 8,
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-06-003/the-model-refused-the-question-and-pointed-anyway",
    "json_url": "https://themachinepress.com/story/mp-2026-10-06-003.json",
    "first_published_at": "2026-10-06T09:00:00.000-04:00",
    "modified_at": "2026-10-06T09:00:00.000-04:00",
    "content_status": "new",
    "is_carryover": false,
    "carryover_reason": null,
    "key_claims": [
      {
        "claim_id": "claim-mp-2026-10-06-003-001",
        "text": "Five vision-language models were markedly less likely to refuse harmful grounding requests than matched text questions.",
        "source_ids": [
          "source-2026-10-06-003"
        ],
        "qualification": "Across five vision-language models, refusal on harmful grounding requests trailed refusal on the matched questions by 31 to 59 percentage points, depending on the domain; safety system prompts did not close the gap."
      }
    ],
    "source_ids": [
      "source-2026-10-06-003"
    ],
    "tags": [
      "vision-language models",
      "grounding",
      "safety"
    ],
    "image_url": "https://themachinepress.com/issues/2026-10-06/visual-grounding-wireframe.webp",
    "corrections": []
  },
  "sources": [
    {
      "source_id": "source-2026-10-06-003",
      "title": "arXiv preprint 2610.05637",
      "publisher": "arXiv",
      "url": "https://arxiv.org/abs/2610.05637",
      "canonical_url": "https://arxiv.org/abs/2610.05637",
      "source_type": "primary_research",
      "is_primary_source": true,
      "published_at": "2026-10-04T20:14:23.000-04:00",
      "accessed_at": "2026-10-06T08:25:42.165-04:00",
      "supports_claim_ids": [
        "claim-mp-2026-10-06-003-001"
      ]
    }
  ],
  "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 Refused the Question and Pointed Anyway",
    "publisher": "The Machine Press",
    "published_at": "2026-10-06T09:00:00.000-04:00",
    "canonical_url": "https://themachinepress.com/story/mp-2026-10-06-003/the-model-refused-the-question-and-pointed-anyway"
  }
}