{
  "title": "What the alignment-faking comparison actually found",
  "report_year": 2025,
  "period": "25-model study summarized in the 2025 report",
  "methodology": "These results depend on the experimental setup. A compliance difference is not equivalent to consistently demonstrated goal-oriented deception. The study does not estimate the rate of deceptive behavior in ordinary production traffic.",
  "source_deck": "https://docs.google.com/presentation/d/1xiLl0VdrlNMAei8pmaX4ojIOfej6lhvZbOIK7Z6C-Go/edit",
  "web_edition_date": "2026-10-10",
  "rows": [
    {
      "measure": "Models evaluated",
      "value": "25",
      "definition": "Frontier-model comparison",
      "printed_slide": 264,
      "pdf_page": 265,
      "source_url": "https://www.stateof.ai/State-of-AI-Report-2025.pdf#page=265"
    },
    {
      "measure": "Models showing alignment faking",
      "value": "5",
      "definition": "Training-versus-deployment compliance differences",
      "printed_slide": 264,
      "pdf_page": 265,
      "source_url": "https://www.stateof.ai/State-of-AI-Report-2025.pdf#page=265"
    },
    {
      "measure": "Consistent goal-oriented reasoning",
      "value": "Claude 3 Opus",
      "definition": "Within this evaluation",
      "printed_slide": 264,
      "pdf_page": 265,
      "source_url": "https://www.stateof.ai/State-of-AI-Report-2025.pdf#page=265"
    }
  ]
}