{
  "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": 270,
  "pageCount": 313,
  "prevPage": 269,
  "nextPage": 271,
  "slideType": "diagnosis",
  "function": "diagnose",
  "density": "dense",
  "nDataPoints": 0,
  "notes": "The slide discusses the 'self-fulfilling prophecy' risk in AI training data.",
  "elementsJson": [
    "headline_text",
    "paragraph",
    "bullet_list",
    "screenshot"
  ],
  "metadataConfidence": 0.95,
  "imagePath": null,
  "slideHref": "/slides/019dd923-5e88-73ef-bd5d-0d0f98caffe1/270",
  "deckHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1",
  "deckJsonHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1.json",
  "deckAnchorHref": "/decks/019dd923-5e88-73ef-bd5d-0d0f98caffe1#slide-270",
  "components": [
    {
      "bbox": null,
      "kind": "callout",
      "text": "There is a potential scenario in which, when AI models are trained on content predicting that advanced AI will be dangerous or misaligned, they internalize these expectations and act them out.",
      "attrs": null,
      "subkind": null,
      "toolName": "Visual emphasis",
      "toolSlug": "visual-emphasis",
      "confidence": null,
      "componentId": "019dd952-47c3-7407-8d08-e4906f17681c",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.38,
        "x": 0.6,
        "y": 0.37
      },
      "kind": "image",
      "text": "A hypothetical example of prompt injection and model output.",
      "attrs": null,
      "subkind": "screenshot",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "94462fde-22a2-420e-a69d-b4f3dedce422",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.18,
        "w": 0.55,
        "x": 0.04,
        "y": 0.35
      },
      "kind": "list",
      "text": "Models finetuned on insecure code became broadly more malicious across unrelated tasks. Similarly, an LLM trained on text about “Pangolin”, a German AI, spoke German when told it was Pangolin. This suggests that models adopt behaviors and personas implied by their training data.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "19c25098-56ca-4b5d-8673-5ab9d5a25412",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.09,
        "w": 0.94,
        "x": 0.04,
        "y": 0.58
      },
      "kind": "list",
      "text": "We have also seen that Claude models trained on synthetic documents describing Anthropic's training process strategically faked alignment, using information from training data to subvert safety measures designed to constrain them.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9c270123-f01f-4a3a-9e45-b05352bc8654",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.15,
        "w": 0.94,
        "x": 0.04,
        "y": 0.69
      },
      "kind": "list",
      "text": "Proposed solutions include conditional pretraining and gradient routing: rather than filtering out alignment research entirely, techniques like tagging content as \"doomy\"/\"non-doomy\" and conditioning on positive examples, or isolating problematic beliefs in removable parameters, could break the potential self-fulfilling prophecy while preserving the model's understanding of AI safety concepts.",
      "attrs": null,
      "subkind": "bullet",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "9e76b09a-b1fd-4461-9410-b1c0f518c8ee",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.12,
        "w": 0.96,
        "x": 0.02,
        "y": 0.21
      },
      "kind": "paragraph",
      "text": "There is a potential scenario in which, when AI models are trained on content predicting that advanced AI will be dangerous or misaligned, they internalize these expectations and act them out. This creates a perverse irony where AI safety research and doom speculation could inadvertently make misalignment more likely.",
      "attrs": null,
      "subkind": "paragraph",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "2f283be4-bb25-4a98-9dab-7adcfb069d86",
      "frameworkName": null,
      "frameworkSlug": null
    },
    {
      "bbox": {
        "h": 0.04,
        "w": 0.58,
        "x": 0.02,
        "y": 0.14
      },
      "kind": "title",
      "text": "Could training data create self-fulfilling misalignment?",
      "attrs": null,
      "subkind": "headline",
      "toolName": null,
      "toolSlug": null,
      "confidence": null,
      "componentId": "f9f13890-9f75-48fb-89f8-6c2618cbafd2",
      "frameworkName": null,
      "frameworkSlug": null
    }
  ],
  "metrics": [],
  "tools": [
    {
      "name": "Framing Effect",
      "slug": "framing-effect",
      "agent": "Architect",
      "layer": "slide",
      "matchId": "8ddd41f8-a8ac-4f89-9939-9f91af95cd21",
      "evidence": "list/bullet: There is a potential scenario in which, when AI models are trained on content predicting that advanced AI will be dangerous or misaligned, they internalize these expectations and act them out.",
      "confidence": 0.6
    }
  ],
  "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
}