Unverified60% confidenceFactExact time
Research by institutions including Carnegie Mellon University has proven that large diffusion models exhibit a 'training data memorization' phenomenon, capable of highly faithfully reproducing copyrighted images from the training set under specific prompts.
2
Sources
60%
Confidence
Long-term
Relevance
7/19/2026
First Seen
Sources
Related Claims
VerifiedResearch by Google DeepMind and other institutions in 2023 confirmed that a model's memorization behavior is closely related to the number of repetitions of training data.77% similarUnverifiedThe core of the training-data copyright dispute is that generative models are usually trained on massive amounts of online video and images, many of which are copyrighted, and that model outputs may resemble training material in style or composition.75% similarUnverifiedResearch shows that under certain conditions (such as a particular piece of text appearing at very high frequency in training data), models may reproduce verbatim content from training data, and in 2023 researchers from Google DeepMind and the University of California, Berkeley successfully extracted original fragments of training data from ChatGPT.72% similarUnverifiedMainstream benchmark problems have been widely discussed on the internet, and models' training data may contain memorization of these problems, causing benchmark scores to overestimate the models' capabilities on truly unseen tasks71% similarUnverifiedUsing real pre-trained PLE performs significantly better than control groups using random memory and shuffled memory.71% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/558570API
curl https://kongchang.com/api/v1/knowledge/claims/558570MCP
get_claim(id=558570)