{
  "project": {
    "title": "Model Collapse: When AI Trains on AI",
    "video_url": "https://www.youtube.com/watch?v=zWh-mJEyWfU",
    "canonical_url": "https://jessicamalnik.com/resources/model-collapse",
    "last_updated": "2026-07-21",
    "source_of_truth": "Documentary transcript supplied by Jessica Malnik"
  },
  "people": [
    {
      "name": "Anthony Pierri",
      "role": "Featured interview contributor",
      "organization": "",
      "topic": "AI, positioning, average best practices, originality",
      "evidence": "Transcript attribution from YouTube description",
      "verification_status": "Needs speaker-to-quote verification"
    },
    {
      "name": "Jeff Large",
      "role": "Featured interview contributor",
      "organization": "",
      "topic": "AI voices, podcasts, content production",
      "evidence": "Transcript attribution from YouTube description",
      "verification_status": "Needs speaker-to-quote verification"
    },
    {
      "name": "Ronnie Higgins",
      "role": "Featured interview contributor",
      "organization": "",
      "topic": "AI, content, trust",
      "evidence": "Transcript attribution from YouTube description",
      "verification_status": "Needs speaker-to-quote verification"
    },
    {
      "name": "Carol Cox",
      "role": "Featured interview contributor",
      "organization": "",
      "topic": "Human connection, remote work, empathy",
      "evidence": "Transcript attribution from YouTube description",
      "verification_status": "Needs speaker-to-quote verification"
    },
    {
      "name": "Venessa Paech",
      "role": "Featured interview contributor",
      "organization": "",
      "topic": "Online communities, culture, trust",
      "evidence": "Transcript attribution from YouTube description",
      "verification_status": "Needs speaker-to-quote verification"
    }
  ],
  "platforms": [
    {
      "name": "Google",
      "type": "Search engine",
      "context": "Used as an example of search feeling less useful",
      "transcript_section": "Opening observation"
    },
    {
      "name": "LinkedIn",
      "type": "Professional social network",
      "context": "Used as an example of repetitive, performative content",
      "transcript_section": "Opening observation"
    },
    {
      "name": "Wikipedia",
      "type": "Collaborative encyclopedia",
      "context": "Example of the human-built early web",
      "transcript_section": "Early internet section"
    },
    {
      "name": "YouTube",
      "type": "Video platform",
      "context": "Example of early human-created internet culture and supporting media",
      "transcript_section": "Early internet section"
    },
    {
      "name": "Reddit",
      "type": "Community platform",
      "context": "Example of forums and human-contributed discussion",
      "transcript_section": "Early internet section"
    },
    {
      "name": "eBaum's World",
      "type": "Entertainment website",
      "context": "Example of early internet culture",
      "transcript_section": "Early internet section"
    }
  ],
  "concepts": [
    {"concept": "Publishing friction", "definition": "The time and effort required to research, write, record, edit, and publish content.", "category": "Content economics", "role": "Core claim"},
    {"concept": "Infinite content", "definition": "A condition where the marginal cost of producing additional content approaches zero.", "category": "Content economics", "role": "Core claim"},
    {"concept": "AI-generated content", "definition": "Articles, images, voices, podcasts, videos, or other media produced substantially by generative AI.", "category": "Synthetic media", "role": "Core concept"},
    {"concept": "AI influencers", "definition": "Synthetic personas designed to resemble human creators and build audiences or commercial partnerships.", "category": "Synthetic media", "role": "Example"},
    {"concept": "Synthetic media", "definition": "Media generated or substantially altered by software rather than directly recorded from human activity.", "category": "Synthetic media", "role": "Core concept"},
    {"concept": "Consensus bias", "definition": "The tendency of AI systems to reproduce common patterns, established advice, and average best practices.", "category": "Information quality", "role": "Interpretive claim"},
    {"concept": "Content convergence", "definition": "Different outputs becoming increasingly similar because they draw from the same models, patterns, and training data.", "category": "Information quality", "role": "Core claim"},
    {"concept": "Model training data", "definition": "Human and machine-created information used to train or refine AI systems.", "category": "AI infrastructure", "role": "Core concept"},
    {"concept": "Synthetic training data", "definition": "Machine-generated information used to train later AI systems.", "category": "AI infrastructure", "role": "Future risk"},
    {"concept": "Model collapse", "definition": "A researched failure mode in which recursive training on generated data can degrade model performance or diversity.", "category": "AI research", "role": "Related term"},
    {"concept": "Hall of mirrors effect", "definition": "A metaphor for machines learning from copies of earlier machine-generated copies.", "category": "Information quality", "role": "Documentary framing"},
    {"concept": "Credibility scarcity", "definition": "The idea that trusted provenance becomes more valuable as information production becomes abundant.", "category": "Trust", "role": "Core thesis"},
    {"concept": "Provenance", "definition": "Information about who or what created a piece of content and how it was produced.", "category": "Trust", "role": "Related concept"},
    {"concept": "Human-origin signal", "definition": "Evidence that content reflects direct human experience, observation, testing, or judgment.", "category": "Trust", "role": "Derived concept"},
    {"concept": "Content saturation", "definition": "An environment in which the volume of published media exceeds the attention available to consume it.", "category": "Content economics", "role": "Core claim"},
    {"concept": "Machine-to-machine content", "definition": "Content generated by machines primarily for indexing, ranking, summarization, or ingestion by other machines.", "category": "AI web", "role": "Core question"},
    {"concept": "Average best practices", "definition": "Widely repeated guidance that may be useful but tends to produce undifferentiated outputs.", "category": "Information quality", "role": "Interview insight"},
    {"concept": "Authenticity uncertainty", "definition": "The growing difficulty of determining whether a person, voice, image, review, or story is genuine.", "category": "Trust", "role": "Core claim"},
    {"concept": "Information ecosystem", "definition": "The network of people, platforms, models, and media through which knowledge is created and circulated.", "category": "Internet", "role": "Core concept"},
    {"concept": "Human-to-human connection", "definition": "Direct interaction that preserves context, empathy, multidimensionality, and social cues.", "category": "Trust", "role": "Interview insight"}
  ],
  "claims": [
    {"claim": "Google feels less useful.", "evidence": "Narrator observation", "external_source_needed": "No", "verification_status": "Subjective observation", "transcript_section": "Opening", "source_url": ""},
    {"claim": "Many online articles increasingly sound alike.", "evidence": "Narrator observation", "external_source_needed": "No", "verification_status": "Subjective observation", "transcript_section": "Opening", "source_url": ""},
    {"claim": "Product reviews feel harder to trust.", "evidence": "Narrator observation", "external_source_needed": "No", "verification_status": "Subjective observation", "transcript_section": "Opening", "source_url": ""},
    {"claim": "The early web relied heavily on original human experiences and perspectives.", "evidence": "Documentary synthesis", "external_source_needed": "No", "verification_status": "Interpretive claim", "transcript_section": "Early internet", "source_url": ""},
    {"claim": "Research, writing, recording, and editing historically created friction in publishing.", "evidence": "Documentary synthesis", "external_source_needed": "No", "verification_status": "Interpretive claim", "transcript_section": "Publishing economics", "source_url": ""},
    {"claim": "Generative AI reduces the cost and time required to create content.", "evidence": "Documentary claim", "external_source_needed": "No", "verification_status": "Verified", "transcript_section": "Publishing economics", "source_url": "https://www.anthropic.com/research/estimating-productivity-gains"},
    {"claim": "An agency was reportedly producing roughly 5,000 shows per day or week.", "evidence": "Uncertain recollection in transcript", "external_source_needed": "No", "verification_status": "Verified via social media report", "transcript_section": "Scale example", "source_url": "https://x.com/sarthakgh/status/2079189926168674446"},
    {"claim": "AI influencers can build audiences and receive paid brand partnerships.", "evidence": "Documentary claim", "external_source_needed": "No", "verification_status": "Verified", "transcript_section": "AI influencers", "source_url": "https://www.sciencedirect.com/science/article/pii/S219985312500126X"},
    {"claim": "AI-generated images have become harder to identify.", "evidence": "Documentary observation", "external_source_needed": "No", "verification_status": "Verified", "transcript_section": "Synthetic media", "source_url": "https://arxiv.org/html/2507.18640v1"},
    {"claim": "AI-generated voices have become more realistic.", "evidence": "Interview observation", "external_source_needed": "No", "verification_status": "Verified", "transcript_section": "Synthetic media", "source_url": "https://www.qmul.ac.uk/news/latest-news/2025/science-and-engineering/se/ai-generated-voices-now-indistinguishable-from-real-human-voices.html"},
    {"claim": "Language models are effective at summarizing consensus and common patterns.", "evidence": "Interview interpretation", "external_source_needed": "No", "verification_status": "Verified", "transcript_section": "Average outputs", "source_url": "https://arxiv.org/html/2507.05123v1"},
    {"claim": "Widespread use of similar models can make content converge around similar frameworks and conclusions.", "evidence": "Documentary hypothesis", "external_source_needed": "No", "verification_status": "Hypothesis", "transcript_section": "Content convergence", "source_url": ""},
    {"claim": "Future AI systems may increasingly encounter AI-generated material in their training sources.", "evidence": "Documentary hypothesis", "external_source_needed": "No", "verification_status": "Hypothesis", "transcript_section": "Recursive training", "source_url": ""},
    {"claim": "Copies learning from copies may reduce originality or information quality.", "evidence": "Documentary hypothesis", "external_source_needed": "No", "verification_status": "Hypothesis; model-collapse research recommended", "transcript_section": "Recursive training", "source_url": ""},
    {"claim": "The internet has always had a trust problem.", "evidence": "Historical interpretation", "external_source_needed": "No", "verification_status": "Interpretive claim", "transcript_section": "Trust", "source_url": ""},
    {"claim": "People previously often assumed online content had a human creator behind it.", "evidence": "Documentary interpretation", "external_source_needed": "No", "verification_status": "Interpretive claim", "transcript_section": "Trust", "source_url": ""},
    {"claim": "Determining whether reviews, podcasts, articles, images, and creators are authentic is becoming more difficult.", "evidence": "Documentary claim", "external_source_needed": "No", "verification_status": "Verified", "transcript_section": "Trust", "source_url": "https://www.npr.org/sections/money/2023/03/07/1160721021/why-we-usually-cant-tell-when-a-review-is-fake"},
    {"claim": "Online-only interaction can reduce empathy and awareness of other people's multidimensionality.", "evidence": "Interview opinion", "external_source_needed": "No", "verification_status": "Opinion", "transcript_section": "Human connection", "source_url": ""},
    {"claim": "When information becomes abundant, credibility can become comparatively scarce and valuable.", "evidence": "Documentary thesis", "external_source_needed": "No", "verification_status": "Interpretive claim", "transcript_section": "Conclusion", "source_url": ""},
    {"claim": "The internet is becoming an information ecosystem that increasingly learns from itself.", "evidence": "Documentary thesis", "external_source_needed": "No", "verification_status": "Framing claim", "transcript_section": "Conclusion", "source_url": ""}
  ],
  "examples": [
    {"example": "Old car repair forum post", "type": "Human experience", "description": "A person documenting how they fixed an old car.", "transcript_section": "Early internet"},
    {"example": "Restaurant review", "type": "Human experience", "description": "A person sharing a direct experience at a restaurant.", "transcript_section": "Early internet"},
    {"example": "Camera forum argument", "type": "Human discussion", "description": "People debating cameras on a forum in 2010.", "transcript_section": "Early internet"},
    {"example": "Niche blog post", "type": "Human publishing", "description": "A person spending hours writing for a small audience.", "transcript_section": "Early internet"},
    {"example": "AI-generated article", "type": "Synthetic text", "description": "One person or company producing many articles at very low marginal cost.", "transcript_section": "Infinite content"},
    {"example": "AI-generated podcast", "type": "Synthetic audio", "description": "Articles or scripts converted into AI voice and podcast-like media.", "transcript_section": "Synthetic media"},
    {"example": "AI influencer", "type": "Synthetic persona", "description": "A fictional person with a consistent identity, audience, sponsorships, and engagement.", "transcript_section": "AI influencers"},
    {"example": "AI-generated image", "type": "Synthetic image", "description": "A realistic image whose machine origin may not be obvious.", "transcript_section": "Synthetic media"},
    {"example": "AI-generated voice", "type": "Synthetic audio", "description": "A voice that may sound sufficiently human to pass casual inspection.", "transcript_section": "Synthetic media"},
    {"example": "Positioning advice from an LLM", "type": "Average output", "description": "AI produces confident guidance based on aggregate patterns and common best practices.", "transcript_section": "Content convergence"},
    {"example": "Machine-generated training material", "type": "Recursive data", "description": "Future models encounter content generated by earlier models.", "transcript_section": "Recursive training"}
  ]
}