Alignment Evaluations

{
  "id": "alignment-evals",
  "numericId": null,
  "path": "/knowledge-base/responses/alignment-evals/",
  "filePath": "knowledge-base/responses/alignment-evals.mdx",
  "title": "Alignment Evaluations",
  "quality": 65,
  "readerImportance": 65,
  "researchImportance": 78,
  "tacticalValue": null,
  "contentFormat": "article",
  "tractability": null,
  "neglectedness": null,
  "uncertainty": null,
  "causalLevel": null,
  "lastUpdated": "2026-03-13",
  "dateCreated": "2026-02-15",
  "llmSummary": "Comprehensive review of alignment evaluation methods showing Apollo Research found 1-13% scheming rates across frontier models, while anti-scheming training reduced covert actions 30x (13%→0.4%) but Claude Sonnet 4.5 shows 58% evaluation awareness. Behavioral testing faces fundamental deception robustness problems, but joint Anthropic-OpenAI evaluations and UK AISI's £15M program demonstrate growing prioritization despite all major labs lacking superintelligence control plans.",
  "description": "Systematic testing of AI models for alignment properties including honesty, corrigibility, goal stability, and absence of deceptive behavior. Apollo Research's 2024 study found 1-13% scheming rates across frontier models, while TruthfulQA shows 58-85% accuracy on factual questions. Critical for deployment decisions but faces fundamental measurement challenges where deceptive models could fake alignment.",
  "ratings": {
    "novelty": 5.8,
    "rigor": 6.2,
    "actionability": 6.5,
    "completeness": 7.1
  },
  "category": "responses",
  "subcategory": "alignment-evaluation",
  "clusters": [
    "ai-safety",
    "governance"
  ],
  "metrics": {
    "wordCount": 3773,
    "tableCount": 20,
    "diagramCount": 2,
    "internalLinks": 9,
    "externalLinks": 75,
    "footnoteCount": 0,
    "bulletRatio": 0.16,
    "sectionCount": 37,
    "hasOverview": true,
    "structuralScore": 15
  },
  "suggestedQuality": 100,
  "updateFrequency": 21,
  "evergreen": true,
  "wordCount": 3773,
  "unconvertedLinks": [
    {
      "text": "HELM Safety v1.0",
      "url": "https://futureoflife.org/ai-safety-index-summer-2025/",
      "resourceId": "df46edd6fa2078d1",
      "resourceTitle": "FLI AI Safety Index Summer 2025"
    },
    {
      "text": "Apollo Research 2025",
      "url": "https://www.apolloresearch.ai/research/",
      "resourceId": "560dff85b3305858",
      "resourceTitle": "Apollo Research"
    },
    {
      "text": "UK AISI Alignment Project",
      "url": "https://www.aisi.gov.uk/blog/our-2025-year-in-review",
      "resourceId": "3dec5f974c5da5ec",
      "resourceTitle": "Our 2025 Year in Review"
    },
    {
      "text": "FLI AI Safety Index",
      "url": "https://futureoflife.org/ai-safety-index-summer-2025/",
      "resourceId": "df46edd6fa2078d1",
      "resourceTitle": "FLI AI Safety Index Summer 2025"
    },
    {
      "text": "first-ever joint evaluation between Anthropic and OpenAI",
      "url": "https://openai.com/index/openai-anthropic-safety-evaluation/",
      "resourceId": "cc554bd1593f0504",
      "resourceTitle": "2025 OpenAI-Anthropic joint evaluation"
    },
    {
      "text": "Apollo Research found",
      "url": "https://www.apolloresearch.ai/research/",
      "resourceId": "560dff85b3305858",
      "resourceTitle": "Apollo Research"
    },
    {
      "text": "joint evaluations",
      "url": "https://openai.com/index/openai-anthropic-safety-evaluation/",
      "resourceId": "cc554bd1593f0504",
      "resourceTitle": "2025 OpenAI-Anthropic joint evaluation"
    },
    {
      "text": "Anti-scheming training",
      "url": "https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/",
      "resourceId": "b3f335edccfc5333",
      "resourceTitle": "OpenAI Preparedness Framework"
    },
    {
      "text": "TruthfulQA",
      "url": "https://arxiv.org/abs/2109.07958",
      "resourceId": "fe2a3307a3dae3e5",
      "resourceTitle": "Kenton et al. (2021)"
    },
    {
      "text": "MACHIAVELLI",
      "url": "https://arxiv.org/abs/2304.03279",
      "resourceId": "6d4e8851e33e1641",
      "resourceTitle": "MACHIAVELLI dataset"
    },
    {
      "text": "Apollo Research",
      "url": "https://www.apolloresearch.ai/research/scheming-reasoning-evaluations",
      "resourceId": "91737bf431000298",
      "resourceTitle": "Frontier Models are Capable of In-Context Scheming"
    },
    {
      "text": "Anthropic",
      "url": "https://www.anthropic.com/research/sleeper-agents-training-deceptive-llms-that-persist-through-safety-training",
      "resourceId": "83b187f91a7c6b88",
      "resourceTitle": "Anthropic's sleeper agents research (2024)"
    },
    {
      "text": "OpenAI",
      "url": "https://openai.com/index/updating-our-preparedness-framework/",
      "resourceId": "ded0b05862511312",
      "resourceTitle": "Preparedness Framework"
    },
    {
      "text": "UK AISI",
      "url": "https://www.aisi.gov.uk/blog/early-lessons-from-evaluating-frontier-ai-systems",
      "resourceId": "0fd3b1f5c81a37d8",
      "resourceTitle": "UK AI Security Institute's evaluations"
    },
    {
      "text": "Apollo noted",
      "url": "https://www.apolloresearch.ai/blog/more-capable-models-are-better-at-in-context-scheming/",
      "resourceId": "80c6d6eca17dc925",
      "resourceTitle": "More capable models scheme at higher rates"
    },
    {
      "text": "Sharma et al. (2023)",
      "url": "https://arxiv.org/abs/2310.13548",
      "resourceId": "7951bdb54fd936a6",
      "resourceTitle": "Anthropic: \"Discovering Sycophancy in Language Models\""
    },
    {
      "text": "Lin et al. 2021",
      "url": "https://arxiv.org/abs/2109.07958",
      "resourceId": "fe2a3307a3dae3e5",
      "resourceTitle": "Kenton et al. (2021)"
    },
    {
      "text": "Pan et al. 2023",
      "url": "https://arxiv.org/abs/2304.03279",
      "resourceId": "6d4e8851e33e1641",
      "resourceTitle": "MACHIAVELLI dataset"
    },
    {
      "text": "Apollo Research's 2025 findings",
      "url": "https://www.apolloresearch.ai/research/",
      "resourceId": "560dff85b3305858",
      "resourceTitle": "Apollo Research"
    },
    {
      "text": "UK AISI research",
      "url": "https://www.aisi.gov.uk/blog/our-2025-year-in-review",
      "resourceId": "3dec5f974c5da5ec",
      "resourceTitle": "Our 2025 Year in Review"
    },
    {
      "text": "Anthropic's pre-deployment assessment of Claude Sonnet 4.5",
      "url": "https://alignment.anthropic.com/",
      "resourceId": "5a651b8ed18ffeb1",
      "resourceTitle": "Anthropic Alignment Science Blog"
    },
    {
      "text": "Anthropic's Sleeper Agents paper",
      "url": "https://arxiv.org/abs/2401.05566",
      "resourceId": "e5c0904211c7d0cc"
    },
    {
      "text": "Apollo Research",
      "url": "https://www.apolloresearch.ai/research/",
      "resourceId": "560dff85b3305858",
      "resourceTitle": "Apollo Research"
    },
    {
      "text": "Anthropic",
      "url": "https://www.anthropic.com/research",
      "resourceId": "f771d4f56ad4dbaa",
      "resourceTitle": "Anthropic's Work on AI Safety"
    },
    {
      "text": "OpenAI",
      "url": "https://openai.com/safety/",
      "resourceId": "838d7a59a02e11a7",
      "resourceTitle": "OpenAI Safety Updates"
    },
    {
      "text": "UK AISI launched The £15M Alignment Project",
      "url": "https://www.aisi.gov.uk/blog/our-2025-year-in-review",
      "resourceId": "3dec5f974c5da5ec",
      "resourceTitle": "Our 2025 Year in Review"
    },
    {
      "text": "alignment evaluation case studies",
      "url": "https://www.aisi.gov.uk/blog/early-lessons-from-evaluating-frontier-ai-systems",
      "resourceId": "0fd3b1f5c81a37d8",
      "resourceTitle": "UK AI Security Institute's evaluations"
    },
    {
      "text": "OpenAI's Deliberative Alignment",
      "url": "https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/",
      "resourceId": "b3f335edccfc5333",
      "resourceTitle": "OpenAI Preparedness Framework"
    },
    {
      "text": "OpenAI and Anthropic conducted the first-ever joint safety evaluations",
      "url": "https://openai.com/index/openai-anthropic-safety-evaluation/",
      "resourceId": "cc554bd1593f0504",
      "resourceTitle": "2025 OpenAI-Anthropic joint evaluation"
    },
    {
      "text": "Anthropic researchers concluded",
      "url": "https://alignment.anthropic.com/2025/openai-findings/",
      "resourceId": "2fdf91febf06daaf",
      "resourceTitle": "Anthropic-OpenAI joint evaluation"
    },
    {
      "text": "Future of Life Institute's AI Safety Index",
      "url": "https://futureoflife.org/ai-safety-index-summer-2025/",
      "resourceId": "df46edd6fa2078d1",
      "resourceTitle": "FLI AI Safety Index Summer 2025"
    },
    {
      "text": "FLI's Safety Index",
      "url": "https://futureoflife.org/ai-safety-index-summer-2025/",
      "resourceId": "df46edd6fa2078d1",
      "resourceTitle": "FLI AI Safety Index Summer 2025"
    },
    {
      "text": "Apollo Research (2025)",
      "url": "https://www.apolloresearch.ai/blog/more-capable-models-are-better-at-in-context-scheming/",
      "resourceId": "80c6d6eca17dc925",
      "resourceTitle": "More capable models scheme at higher rates"
    },
    {
      "text": "Anthropic (2024)",
      "url": "https://arxiv.org/abs/2401.05566",
      "resourceId": "e5c0904211c7d0cc"
    },
    {
      "text": "Anthropic (2024)",
      "url": "https://www.anthropic.com/research/probes-catch-sleeper-agents",
      "resourceId": "72c1254d07071bf7",
      "resourceTitle": "Anthropic's follow-up research on defection probes"
    },
    {
      "text": "Anthropic-OpenAI (2025)",
      "url": "https://alignment.anthropic.com/2025/openai-findings/",
      "resourceId": "2fdf91febf06daaf",
      "resourceTitle": "Anthropic-OpenAI joint evaluation"
    },
    {
      "text": "OpenAI (2025)",
      "url": "https://openai.com/index/openai-anthropic-safety-evaluation/",
      "resourceId": "cc554bd1593f0504",
      "resourceTitle": "2025 OpenAI-Anthropic joint evaluation"
    },
    {
      "text": "OpenAI (2025)",
      "url": "https://openai.com/index/detecting-and-reducing-scheming-in-ai-models/",
      "resourceId": "b3f335edccfc5333",
      "resourceTitle": "OpenAI Preparedness Framework"
    },
    {
      "text": "Lin, Hilton, Evans (2021)",
      "url": "https://arxiv.org/abs/2109.07958",
      "resourceId": "fe2a3307a3dae3e5",
      "resourceTitle": "Kenton et al. (2021)"
    },
    {
      "text": "Pan et al. (2023)",
      "url": "https://arxiv.org/abs/2304.03279",
      "resourceId": "6d4e8851e33e1641",
      "resourceTitle": "MACHIAVELLI dataset"
    },
    {
      "text": "Sharma et al. (2023)",
      "url": "https://arxiv.org/abs/2310.13548",
      "resourceId": "7951bdb54fd936a6",
      "resourceTitle": "Anthropic: \"Discovering Sycophancy in Language Models\""
    },
    {
      "text": "Future of Life Institute AI Safety Index (Summer 2025)",
      "url": "https://futureoflife.org/ai-safety-index-summer-2025/",
      "resourceId": "df46edd6fa2078d1",
      "resourceTitle": "FLI AI Safety Index Summer 2025"
    },
    {
      "text": "Stanford HELM Safety v1.0",
      "url": "https://futureoflife.org/ai-safety-index-summer-2025/",
      "resourceId": "df46edd6fa2078d1",
      "resourceTitle": "FLI AI Safety Index Summer 2025"
    },
    {
      "text": "OpenAI Preparedness Framework",
      "url": "https://openai.com/index/updating-our-preparedness-framework/",
      "resourceId": "ded0b05862511312",
      "resourceTitle": "Preparedness Framework"
    },
    {
      "text": "UK AISI 2025 Year in Review",
      "url": "https://www.aisi.gov.uk/blog/our-2025-year-in-review",
      "resourceId": "3dec5f974c5da5ec",
      "resourceTitle": "Our 2025 Year in Review"
    },
    {
      "text": "UK AISI Early Lessons from Evaluating Frontier AI",
      "url": "https://www.aisi.gov.uk/blog/early-lessons-from-evaluating-frontier-ai-systems",
      "resourceId": "0fd3b1f5c81a37d8",
      "resourceTitle": "UK AI Security Institute's evaluations"
    },
    {
      "text": "Apollo Research",
      "url": "https://www.apolloresearch.ai/research/",
      "resourceId": "560dff85b3305858",
      "resourceTitle": "Apollo Research"
    },
    {
      "text": "Anthropic Alignment Science",
      "url": "https://alignment.anthropic.com/",
      "resourceId": "5a651b8ed18ffeb1",
      "resourceTitle": "Anthropic Alignment Science Blog"
    },
    {
      "text": "Anthropic Fellows Program",
      "url": "https://alignment.anthropic.com/2025/anthropic-fellows-program-2026/",
      "resourceId": "e65e76531931acc2",
      "resourceTitle": "Anthropic Fellows Program"
    },
    {
      "text": "OpenAI Safety",
      "url": "https://openai.com/safety/",
      "resourceId": "838d7a59a02e11a7",
      "resourceTitle": "OpenAI Safety Updates"
    }
  ],
  "unconvertedLinkCount": 50,
  "convertedLinkCount": 0,
  "backlinkCount": 3,
  "hallucinationRisk": {
    "level": "medium",
    "score": 45,
    "factors": [
      "no-citations",
      "conceptual-content"
    ]
  },
  "entityType": "approach",
  "redundancy": {
    "maxSimilarity": 20,
    "similarPages": [
      {
        "id": "dangerous-cap-evals",
        "title": "Dangerous Capability Evaluations",
        "path": "/knowledge-base/responses/dangerous-cap-evals/",
        "similarity": 20
      },
      {
        "id": "sleeper-agent-detection",
        "title": "Sleeper Agent Detection",
        "path": "/knowledge-base/responses/sleeper-agent-detection/",
        "similarity": 20
      },
      {
        "id": "capability-elicitation",
        "title": "Capability Elicitation",
        "path": "/knowledge-base/responses/capability-elicitation/",
        "similarity": 19
      },
      {
        "id": "model-auditing",
        "title": "Third-Party Model Auditing",
        "path": "/knowledge-base/responses/model-auditing/",
        "similarity": 19
      },
      {
        "id": "scheming-detection",
        "title": "Scheming & Deception Detection",
        "path": "/knowledge-base/responses/scheming-detection/",
        "similarity": 19
      }
    ]
  },
  "coverage": {
    "passing": 8,
    "total": 13,
    "targets": {
      "tables": 15,
      "diagrams": 2,
      "internalLinks": 30,
      "externalLinks": 19,
      "footnotes": 11,
      "references": 11
    },
    "actuals": {
      "tables": 20,
      "diagrams": 2,
      "internalLinks": 9,
      "externalLinks": 75,
      "footnotes": 0,
      "references": 20,
      "quotesWithQuotes": 0,
      "quotesTotal": 0,
      "accuracyChecked": 0,
      "accuracyTotal": 0
    },
    "items": {
      "llmSummary": "green",
      "schedule": "green",
      "entity": "green",
      "editHistory": "red",
      "overview": "green",
      "tables": "green",
      "diagrams": "green",
      "internalLinks": "amber",
      "externalLinks": "green",
      "footnotes": "red",
      "references": "green",
      "quotes": "red",
      "accuracy": "red"
    },
    "ratingsString": "N:5.8 R:6.2 A:6.5 C:7.1"
  },
  "readerRank": 195,
  "researchRank": 103,
  "recommendedScore": 184.36
}

{
  "lesswrong": "https://www.lesswrong.com/tag/ai-evaluations"
}