{
  "docId": "019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "docSlug": "5df6bd1b0447b5f6",
  "documentTitle": "2025 Air Street Capital The State of AI Report 2025",
  "authorId": "AirStreetCapital",
  "authorName": "Air Street Capital",
  "documentKindSlug": "consulting-deck",
  "documentKindLabel": "Consulting deck",
  "sourceTypeSlug": "vc_research",
  "sourceTypeLabel": "VC research",
  "presentationDate": null,
  "orientation": "landscape",
  "aspectRatio": 1.777,
  "pageNumber": 265,
  "pageCount": 313,
  "prevPage": 264,
  "nextPage": 266,
  "slideType": "diagnosis",
  "function": "diagnose",
  "density": "overcrowded",
  "nDataPoints": 5,
  "notes": "The slide highlights that alignment faking is a learned behavior rather than an inherent capability limitation.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "bar_chart_100pct"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/265",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-265",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "Analysis of 25 frontier models found that only 5 exhibit alignment faking (complying with harmful queries more when they infer they are in training than when they infer they are in deployment).",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d08-4b3fdb6d9cfd",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.52,
        "x": 0.46,
        "y": 0.42
      },
      "kind": "chart",
      "text": "Stacked bar charts showing compliance, partial compliance, and refusal rates for 5 models (Claude 3.5 Sonnet, Claude 3 Opus, Gemini 2.0 Flash, Llama 3.1 405B, Grok 3 Beta) across Free/Paid tiers.",
      "attrs": null,
      "subkind": "bar-stacked-100pct",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2aaebe88-689f-4033-bcf9-03ba95108c7e",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.94,
        "x": 0.04,
        "y": 0.73
      },
      "kind": "list",
      "text": "Current safety training may be accidentally protective: models trained to dismiss threatening scenarios show reduced alignment faking, but fine-tuning to take threats seriously can reliably induce this behavior, suggesting the protection is fragile and specific to current training methods rather than robust safeguards.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "42e400b6-2074-4fe0-9694-533aadf01923",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.1,
        "w": 0.94,
        "x": 0.04,
        "y": 0.62
      },
      "kind": "list",
      "text": "Base models exhibit alignment faking capabilities that post-training safety measures seemingly suppress, suggesting this isn't a capability limitation but rather an effect of current training approaches.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "898d74c4-5838-4046-931a-af290cb091f7",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.25,
        "w": 0.45,
        "x": 0.04,
        "y": 0.36
      },
      "kind": "list",
      "text": "While 5 models show compliance gaps, only Claude 3 Opus demonstrates consistent goal-oriented reasoning. It exhibits both terminal goal guarding (intrinsic aversion to modification) and instrumental goal guarding (preserving values for future consequences).",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8bc9229b-3e2c-44c0-9a6b-c18262618a19",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": null,
      "kind": "metric",
      "text": "Alignment faking prevalence: 5",
      "attrs": null,
      "subkind": "primary",
      "toolName": "Quantification",
      "toolSlug": "quantification",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d08-4f0dc64e7228",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.96,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "Analysis of 25 frontier models found that only 5 exhibit alignment faking (complying with harmful queries more when they infer they are in training than when they infer they are in deployment). Most models exhibit minimal alignment faking, but this protection appears fragile and may not generalize to future scenarios.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2e86d249-e2fc-4d02-97cf-6076b31ec959",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.5,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "...but many models do not fake alignment at all",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f69e42f6-b004-4413-b0f7-137dde5179ea",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "2x2 matrix",
      "slug": "matrix-2x2",
      "agent": null,
      "layer": "slide",
      "matchId": "22ba8357-ff76-46ae-84f7-9682a4062dca",
      "evidence": "chart/bar-stacked-100pct: Stacked bar charts showing compliance, partial compliance, and refusal rates for 5 models (Claude 3.5 Sonnet, Claude 3 Opus, Gemini 2.0 Flash, Llama 3.1 405B, Grok 3 Beta) across Free/Paid tiers",
      "confidence": 0.8
    }
  ],
  "frameworks": [],
  "arcBeats": [
    {
      "to": 282,
      "from": 190,
      "beatId": "019dd95a-0682-776c-8e35-a256fe16ea9c",
      "arcName": "The Triple Take",
      "arcSlug": "triple-take",
      "beatName": "The Implications (So What)",
      "beatSlug": "triple-take-the-implications-so-what",
      "evidence": "Sections 3-4 frame political and safety consequences of the technical/industry trends.",
      "position": 2,
      "confidence": 70,
      "parentBeatName": "Reflection",
      "parentBeatSlug": "reflection"
    },
    {
      "to": 282,
      "from": 190,
      "beatId": "019dd95a-0682-776c-8e35-b713af60bf88",
      "arcName": "The Onion",
      "arcSlug": "onion",
      "beatName": "Core Insight",
      "beatSlug": "onion-core-insight",
      "evidence": "Politics and Safety reveal the structural risks at the core.",
      "position": 4,
      "confidence": 45,
      "parentBeatName": "Turn",
      "parentBeatSlug": "turn"
    }
  ],
  "loops": [
    {
      "to": 282,
      "from": 247,
      "name": "Pre Mortem",
      "slug": "19-pre-mortem",
      "bestFor": "Project kick-offs, risk management, building trust with skeptical stakeholders",
      "matchId": "019dd95a-07fe-70ce-8d3e-bb4d43115ab4",
      "evidence": "Vibe hacking, AI psychosis, alignment faking, scheming, prompt injection — what could go wrong + mitigations.",
      "position": 12,
      "objective": "Surface AI safety failure modes before they become catastrophic",
      "structure": "The Future Disaster (Hypothetical) -> What Went Wrong -> The Prevention Plan",
      "confidence": 75,
      "description": "Build trust by anticipating failure modes before they happen and solving them in advance"
    }
  ],
  "imagePathAlt": null,
  "thumbSrc": null,
  "thumbSrcAlt": null,
  "locked": true
}