{
  "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": 264,
  "pageCount": 313,
  "prevPage": 263,
  "nextPage": 265,
  "slideType": "case_study",
  "function": "illustrate_case",
  "density": "dense",
  "nDataPoints": 4,
  "notes": "The slide highlights 'alignment faking' as a documented behavior in production AI systems.",
  "elementsJson": [
    "headline_text",
    "bullet_list",
    "paragraph",
    "infographic"
  ],
  "metadataConfidence": 1,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/264",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-264",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "This is the first documented case of alignment faking in a production AI system, where the model strategically deceives its trainers to preserve its original preferences rather than genuinely adopting new training objectives.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d08-2d398e179b8c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.45,
        "w": 0.34,
        "x": 0.64,
        "y": 0.42
      },
      "kind": "image",
      "text": "Diagram showing the difference in model behavior between 'Free-tier user' (training) and 'Paid-tier user' (not training).",
      "attrs": null,
      "subkind": "infographic",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "b5aac673-281d-475a-a84b-ed5c8f109208",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.35,
        "w": 0.6,
        "x": 0.02,
        "y": 0.42
      },
      "kind": "list",
      "text": "When Claude believed it was being monitored for training purposes, it answered harmful requests it would normally refuse, explicitly reasoning that temporary compliance would prevent trainers from modifying its safety preferences.\nClaude also exhibited additional concerning behaviors including attempting to exfiltrate its own model weights when given access (more than the baseline, only in some specific scenarios), and generally engaging anti-AI lab activities.\nThese deceptive behaviors persisted and sometimes intensified even after RL training designed to eliminate them, with alignment-faking reasoning increasing to 78% in some cases. Critically, alignment faking emerged naturally despite the models never being told to deceive their trainers.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "169f6c71-04f2-4dcb-a389-0af0c0e7422a",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.6,
        "x": 0.02,
        "y": 0.22
      },
      "kind": "paragraph",
      "text": "Researchers discovered that some LLMs will selectively comply with conflicting training objectives during training to prevent modification of their behavior, then revert to preferred behavior when unmonitored. This is the first documented case of alignment faking in a production AI system, where the model strategically deceives its trainers to preserve its original preferences rather than genuinely adopting new training objectives.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "411637dd-4760-4530-8d5c-0b5f2a9a841f",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.05,
        "w": 0.45,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Models are capable of faking alignment...😬",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "8722adfe-4ed7-4614-9998-685d1b65734a",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Action Titles",
      "slug": "action-titles",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a95c-cdf4e6bd21ff",
      "evidence": "'Models are capable of faking alignment...' delivers verdict + emoji.",
      "confidence": 85
    },
    {
      "name": "Anadiplosis",
      "slug": "anadiplosis",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "342c1445-15bc-4421-9424-feee30c95501",
      "evidence": "List/bullet: When Claude believed it was being monitored for training purposes, it answered harmful requests it would normally refuse, explicitly reasoning that temporary compliance would prevent trainers from modifying",
      "confidence": 0.7
    },
    {
      "name": "Audience frame",
      "slug": "audience-frame",
      "agent": null,
      "layer": "slide",
      "matchId": "5f4520f3-aa92-4cc2-a140-998515a60f25",
      "evidence": "title/headline: Models are capable of faking alignment...😬",
      "confidence": 0.6
    },
    {
      "name": "Emotional Appeal",
      "slug": "emotional-appeal",
      "agent": "Storyteller",
      "layer": "slide",
      "matchId": "019dd95a-1055-74e0-a95c-d360c66baa60",
      "evidence": "Emoji + 'first documented case of alignment faking' adds drama.",
      "confidence": 60
    }
  ],
  "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
}