Verified70% confidenceFactExact time
The concept of Model Collapse was first systematically articulated by research teams from Oxford University and Cambridge University in the 2023 paper The Curse of Recursion: Training on Generated Data Makes Models Forget.
4
Sources
70%
Confidence
Long-term
Relevance
8/11/2026
First Seen
Sources
AI伪造创造力：当每间办公室的创意都在趋同
Related Claims
Unverified2023年牛津大学Ilia Shumailov团队在Nature上发表研究系统阐述了模型崩溃效应，指出模型在自身生成数据上迭代训练相当于对原始数据分布有损压缩79% similarVerifiedModel collapse refers to the phenomenon whereby, after an AI model is trained on data generated by a previous-generation model, the output distribution gradually deviates from the original real data distribution across multiple generations of iteration, losing information in the tail of the distribution.73% similarUnverified'Model collapse' refers to a phenomenon in which language models, repeatedly trained on a mixture containing large amounts of synthetic text generated by similar models, gradually lose the ability to learn low-frequency real-world knowledge, produce increasingly homogeneous outputs, and significantly degrade after several iterations.71% similarUnverifiedThe capability degradation phenomenon (also known as 'model collapse') has been observed multiple times when fine-tuning with synthetic data generated by the model itself.71% similarUnverifiedRelated research has already observed a degradation trend of model collapse in modalities such as text and images, that is, AI-generated content flowing back causes the proportion of synthetic text in subsequent model training corpora to rise and the distribution to deviate from real-world data.69% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/729464API
curl https://kongchang.com/api/v1/knowledge/claims/729464MCP
get_claim(id=729464)