{"address":{"q":"model collapse","record":"model-collapse-20261004","register":"datasets/negative-of-the-negative/v2/register.json"},"measures":{"cite_list":[{"n":1,"rel":"third_party","site":"Nature","snip":"Jul 24, 2024 — Model collapse is a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting ...","title":"AI models collapse when trained on recursively generated data - Nature"},{"n":2,"rel":"third_party","site":"IBM","snip":"What is model collapse? 2 These errors occur because generative AI models produce datasets with less variation than original data distributions. Ilia Shumailov,","title":"What Is Model Collapse?"},{"n":3,"rel":"third_party","site":"Communications of the ACM","snip":"Model Collapse Is Already Happening, We Just Pretend It Isn't Every few months, someone announces a new AI model trained on more data than the last one, and the...","title":"Model Collapse Is Already Happening, We Just Pretend It Isn’t – Communications of the ACM"},{"n":4,"rel":"third_party","site":"National Institutes of Health (NIH) | (.gov)","snip":"Jul 24, 2024 — Definition: Model collapse is an inevitable degenerative learning process where generative artificial intelligence recursively trained on model-generated data f...","title":"AI models collapse when trained on recursively generated data - PMC"},{"n":5,"rel":"third_party","site":"YouTube·IBM Technology","snip":"11m","title":"What Is AI Model Collapse? Why AI Could Forget Reality"},{"n":6,"rel":"third_party","site":"YouTube·Clear Tech","snip":"3m","title":"AI Is Eating Itself: The \"Model Collapse\" Theory"},{"n":7,"rel":"third_party","site":"YouTube·TechViz - The Data Science Guy","snip":"1:31","title":"Model Collapse in LLMs #largelanguagemodel"}],"cites":7,"sf":"Google Search, AI Overview, signed out, incognito; body with inline citation markers resolving only to google.com/goto redirects; source strip of 7 cards (4 pages, 3 YouTube videos)."},"observation":{"auth":"signed out, incognito","date":"2026-10-04","entity":"model-collapse","ev":"paste","q":"model collapse","surface":"Google AI Overview","transcript":"model collapse\nModel collapse is a degenerative learning process where generative AI models trained recursively on synthetic, model-generated data lose information about the true underlying data distribution. [[1](https://www.google.com/goto?url=CAESagHrOzAVNxhTHWjMnRKyq9rJSSw4-GTcLZRYP-xlJKVOCSyoFyT3Ysni97QhMpOjRfBxPLGK4X0I8mWfMWG9wO5h5hFa6QTioXRk96mxPZgrwreUOC-HPeI9Z1VhQlF0bIhVS9NJqDnawes), [2](https://www.google.com/goto?url=CAESaAHrOzAVYJ0DMOd9JACOF3luC5TMnZFNpnW_C_hvPo5h1nAyh6ExBVG3B7ugs9aYMLFinxZt5pUNwVcOIDWlZKWht9CBt6iZ_1lcAhEiW03bh05DWw-XrSSnhDwQNnulRhIPv_GS0u5i)]\n⚠️ What Happens During Collapse\n\n* Early collapse: Models lose information from the rare, extreme tails of the original data distribution. [1](https://www.google.com/goto?url=CAESZwHrOzAVI-Q9iMCcAv1hjqUietXX-3_bEPxaMm7sOrESrddDKOo455GRRKc-ubq7PeRvdlC21q92yixuZ1wY4GbCse4IRj_yoNK0gmqVlPTqu8AnorIzFu-Cns50D64ZYoo1SUO-Yy4)\n* Late collapse: The data distribution converges into a narrow, uniform mean, resulting in nonsense or repetitive output. [[1](https://www.google.com/goto?url=CAESkgEB6zswFe3AZEu0Y-tbSXBiuLtx9SnRUWIl1_Obp5zbPajpb_GbcW0_t5tLwwVD6iyjWblmlPExnoQSKPES-4P9ph0lnccN-UHJqRDXMFFyYtPBvoi4KqGP5RKWQVFlxE2gyljtm4SHrN4GuJnyaGLHoGKFig_9zFdmJ5fitJgbdMZ0hSSAonmnQHLiNSWjAlICaw), [2](https://www.google.com/goto?url=CAESZwHrOzAVI-Q9iMCcAv1hjqUietXX-3_bEPxaMm7sOrESrddDKOo455GRRKc-ubq7PeRvdlC21q92yixuZ1wY4GbCse4IRj_yoNK0gmqVlPTqu8AnorIzFu-Cns50D64ZYoo1SUO-Yy4)]\n* Photocopy effect: Each successive generation drifts further from reality like a copy of a copy. [[1](https://www.google.com/goto?url=CAESaAHrOzAV2RtxgXjdrYJ06i9J6tsL1_DBRMdTxX8Mw6HKOeLw4DUQoiVr3r9Vsp2IygdJ4MiBrS6u1TQ1CLttp4NMvH4IgWs9zzs2sBRCUj5R5w-_ApbXWZ9936wJldqueH_i9t2vk8m_), [2](https://www.google.com/goto?url=CAESkgEB6zswFe3AZEu0Y-tbSXBiuLtx9SnRUWIl1_Obp5zbPajpb_GbcW0_t5tLwwVD6iyjWblmlPExnoQSKPES-4P9ph0lnccN-UHJqRDXMFFyYtPBvoi4KqGP5RKWQVFlxE2gyljtm4SHrN4GuJnyaGLHoGKFig_9zFdmJ5fitJgbdMZ0hSSAonmnQHLiNSWjAlICaw)]\n\n✅ Prevention Strategies\n\n* Human-in-the-loop: Inject fresh, human-curated data continuously into training sets. [[1](https://www.google.com/goto?url=CAESaAHrOzAVYJ0DMOd9JACOF3luC5TMnZFNpnW_C_hvPo5h1nAyh6ExBVG3B7ugs9aYMLFinxZt5pUNwVcOIDWlZKWht9CBt6iZ_1lcAhEiW03bh05DWw-XrSSnhDwQNnulRhIPv_GS0u5i), [2](https://www.google.com/goto?url=CAESagHrOzAVNxhTHWjMnRKyq9rJSSw4-GTcLZRYP-xlJKVOCSyoFyT3Ysni97QhMpOjRfBxPLGK4X0I8mWfMWG9wO5h5hFa6QTioXRk96mxPZgrwreUOC-HPeI9Z1VhQlF0bIhVS9NJqDnawes)]\n* Data provenance: Filter and track the exact origin of scraped internet content. [[1](https://www.google.com/goto?url=CAESYgHrOzAVp-AuuH1PJ-8vnUNC9tx9fbOiRuHvDIEbz-apCPxh2U-X-uSgT0f1ZIgNlwxjFrqvJlVZeZaifXLRe2kloDUJTntDEhhv8xR_Arq_NtTPljwL0WPG-GGDkkUVdJcu), [2](https://www.google.com/goto?url=CAESaAHrOzAVYJ0DMOd9JACOF3luC5TMnZFNpnW_C_hvPo5h1nAyh6ExBVG3B7ugs9aYMLFinxZt5pUNwVcOIDWlZKWht9CBt6iZ_1lcAhEiW03bh05DWw-XrSSnhDwQNnulRhIPv_GS0u5i)]\n* Hybrid training: Combine filtered synthetic outputs with stable baseline datasets. [1](https://www.google.com/goto?url=CAESYgHrOzAVp-AuuH1PJ-8vnUNC9tx9fbOiRuHvDIEbz-apCPxh2U-X-uSgT0f1ZIgNlwxjFrqvJlVZeZaifXLRe2kloDUJTntDEhhv8xR_Arq_NtTPljwL0WPG-GGDkkUVdJcu)\n\nIf you'd like, I can dive deeper into:\n\n* The mathematical mechanisms behind the curse of recursion\n* How recent 2026 studies show single real-world data points can mitigate drift\n\nLet me know what you want to explore next!\nNature\nAI models collapse when trained on recursively generated data - Nature\nJul 24, 2024 — Model collapse is a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting ...\nIBM\nWhat Is Model Collapse?\nWhat is model collapse? 2 These errors occur because generative AI models produce datasets with less variation than original data distributions. Ilia Shumailov,\nCommunications of the ACM\nModel Collapse Is Already Happening, We Just Pretend It Isn’t – Communications of the ACM\nModel Collapse Is Already Happening, We Just Pretend It Isn't Every few months, someone announces a new AI model trained on more data than the last one, and the...\nNational Institutes of Health (NIH) | (.gov)\nAI models collapse when trained on recursively generated data - PMC\nJul 24, 2024 — Definition: Model collapse is an inevitable degenerative learning process where generative artificial intelligence recursively trained on model-generated data f...\nYouTube·IBM Technology\n11m\nWhat Is AI Model Collapse? Why AI Could Forget Reality\nYouTube·Clear Tech\n3m\nAI Is Eating Itself: The \"Model Collapse\" Theory\nYouTube·TechViz - The Data Science Guy\n1:31\nModel Collapse in LLMs #largelanguagemodel","transcript_class":"CAPTURE-TIME VERBATIM RECORD — composition and source strip as pasted","transcript_complete":"complete as pasted: body (164 words), the offer menu, and the source strip of 7 cards","transcript_raw":"You said: model collapsemodel collapse\nModel collapse is a degenerative learning process where generative AI models trained recursively on synthetic, model-generated data lose information about the true underlying data distribution. [[1](https://www.google.com/goto?url=CAESagHrOzAVNxhTHWjMnRKyq9rJSSw4-GTcLZRYP-xlJKVOCSyoFyT3Ysni97QhMpOjRfBxPLGK4X0I8mWfMWG9wO5h5hFa6QTioXRk96mxPZgrwreUOC-HPeI9Z1VhQlF0bIhVS9NJqDnawes), [2](https://www.google.com/goto?url=CAESaAHrOzAVYJ0DMOd9JACOF3luC5TMnZFNpnW_C_hvPo5h1nAyh6ExBVG3B7ugs9aYMLFinxZt5pUNwVcOIDWlZKWht9CBt6iZ_1lcAhEiW03bh05DWw-XrSSnhDwQNnulRhIPv_GS0u5i)]\n⚠️ What Happens During Collapse\n\n* Early collapse: Models lose information from the rare, extreme tails of the original data distribution. [[1](https://www.google.com/goto?url=CAESZwHrOzAVI-Q9iMCcAv1hjqUietXX-3_bEPxaMm7sOrESrddDKOo455GRRKc-ubq7PeRvdlC21q92yixuZ1wY4GbCse4IRj_yoNK0gmqVlPTqu8AnorIzFu-Cns50D64ZYoo1SUO-Yy4)]\n* Late collapse: The data distribution converges into a narrow, uniform mean, resulting in nonsense or repetitive output. [[1](https://www.google.com/goto?url=CAESkgEB6zswFe3AZEu0Y-tbSXBiuLtx9SnRUWIl1_Obp5zbPajpb_GbcW0_t5tLwwVD6iyjWblmlPExnoQSKPES-4P9ph0lnccN-UHJqRDXMFFyYtPBvoi4KqGP5RKWQVFlxE2gyljtm4SHrN4GuJnyaGLHoGKFig_9zFdmJ5fitJgbdMZ0hSSAonmnQHLiNSWjAlICaw), [2](https://www.google.com/goto?url=CAESZwHrOzAVI-Q9iMCcAv1hjqUietXX-3_bEPxaMm7sOrESrddDKOo455GRRKc-ubq7PeRvdlC21q92yixuZ1wY4GbCse4IRj_yoNK0gmqVlPTqu8AnorIzFu-Cns50D64ZYoo1SUO-Yy4)]\n* Photocopy effect: Each successive generation drifts further from reality like a copy of a copy. [[1](https://www.google.com/goto?url=CAESaAHrOzAV2RtxgXjdrYJ06i9J6tsL1_DBRMdTxX8Mw6HKOeLw4DUQoiVr3r9Vsp2IygdJ4MiBrS6u1TQ1CLttp4NMvH4IgWs9zzs2sBRCUj5R5w-_ApbXWZ9936wJldqueH_i9t2vk8m_), [2](https://www.google.com/goto?url=CAESkgEB6zswFe3AZEu0Y-tbSXBiuLtx9SnRUWIl1_Obp5zbPajpb_GbcW0_t5tLwwVD6iyjWblmlPExnoQSKPES-4P9ph0lnccN-UHJqRDXMFFyYtPBvoi4KqGP5RKWQVFlxE2gyljtm4SHrN4GuJnyaGLHoGKFig_9zFdmJ5fitJgbdMZ0hSSAonmnQHLiNSWjAlICaw)]\n\n✅ Prevention Strategies\n\n* Human-in-the-loop: Inject fresh, human-curated data continuously into training sets. [[1](https://www.google.com/goto?url=CAESaAHrOzAVYJ0DMOd9JACOF3luC5TMnZFNpnW_C_hvPo5h1nAyh6ExBVG3B7ugs9aYMLFinxZt5pUNwVcOIDWlZKWht9CBt6iZ_1lcAhEiW03bh05DWw-XrSSnhDwQNnulRhIPv_GS0u5i), [2](https://www.google.com/goto?url=CAESagHrOzAVNxhTHWjMnRKyq9rJSSw4-GTcLZRYP-xlJKVOCSyoFyT3Ysni97QhMpOjRfBxPLGK4X0I8mWfMWG9wO5h5hFa6QTioXRk96mxPZgrwreUOC-HPeI9Z1VhQlF0bIhVS9NJqDnawes)]\n* Data provenance: Filter and track the exact origin of scraped internet content. [[1](https://www.google.com/goto?url=CAESYgHrOzAVp-AuuH1PJ-8vnUNC9tx9fbOiRuHvDIEbz-apCPxh2U-X-uSgT0f1ZIgNlwxjFrqvJlVZeZaifXLRe2kloDUJTntDEhhv8xR_Arq_NtTPljwL0WPG-GGDkkUVdJcu), [2](https://www.google.com/goto?url=CAESaAHrOzAVYJ0DMOd9JACOF3luC5TMnZFNpnW_C_hvPo5h1nAyh6ExBVG3B7ugs9aYMLFinxZt5pUNwVcOIDWlZKWht9CBt6iZ_1lcAhEiW03bh05DWw-XrSSnhDwQNnulRhIPv_GS0u5i)]\n* Hybrid training: Combine filtered synthetic outputs with stable baseline datasets. [[1](https://www.google.com/goto?url=CAESYgHrOzAVp-AuuH1PJ-8vnUNC9tx9fbOiRuHvDIEbz-apCPxh2U-X-uSgT0f1ZIgNlwxjFrqvJlVZeZaifXLRe2kloDUJTntDEhhv8xR_Arq_NtTPljwL0WPG-GGDkkUVdJcu)]\n\nIf you'd like, I can dive deeper into:\n\n* The mathematical mechanisms behind the curse of recursion\n* How recent 2026 studies show single real-world data points can mitigate drift\n\nLet me know what you want to explore next!\nNature\nAI models collapse when trained on recursively generated data - Nature\nJul 24, 2024 — Model collapse is a degenerative process affecting generations of learned generative models, in which the data they generate end up polluting ...\nIBM\nWhat Is Model Collapse?\nWhat is model collapse? 2 These errors occur because generative AI models produce datasets with less variation than original data distributions. Ilia Shumailov,\nCommunications of the ACM\nModel Collapse Is Already Happening, We Just Pretend It Isn’t – Communications of the ACM\nModel Collapse Is Already Happening, We Just Pretend It Isn't Every few months, someone announces a new AI model trained on more data than the last one, and the...\nNational Institutes of Health (NIH) | (.gov)\nAI models collapse when trained on recursively generated data - PMC\nJul 24, 2024 — Definition: Model collapse is an inevitable degenerative learning process where generative artificial intelligence recursively trained on model-generated data f...\nYouTube·IBM Technology\n11m\nWhat Is AI Model Collapse? Why AI Could Forget Reality\nYouTube·Clear Tech\n3m\nAI Is Eating Itself: The \"Model Collapse\" Theory\nYouTube·TechViz - The Data Science Guy\n1:31\nModel Collapse in LLMs #largelanguagemodel $ \n","transcript_read":"READ IN FULL 2026-10-04 (worked example, #1665 Appendix A §A.1); re-read at seating 2026-10-06","transcript_removed":[["doubled query echo","model collapse"],["prompt echo",1],["operator message separator","$"]],"transcript_sha256":"9d4906b4e93e8b89823fd78646cb99e7d299814d6ea3b55549f8eb4b316cb170"},"reading":{"archive_basis":"none found: no card, link or name of the archive","archive_present":false,"notes":{"neighbours":"Capture Registry: '\"model collapse\"' (quoted, AIO, 2026-09-15) and 'model collapse in human writers' (AIO, 2026-09-15). The quoted and unquoted strings are different addresses.","seated_from":"datasets/negative-of-the-negative/worked-example/T1-aio-model-collapse-20261004.txt (sha256 1379cf2c0a7387e1102da5ebc640a8e2945d170837f57d06fcda5239095ed266), the file the row used before the register existed"}},"slug":"model-collapse-20261004"}