diffusion models2 articles
The Bigger the AI Model, the Harder It Is to Blame It for Anything
MIT researchers found that as AI diffusion models grow larger and are trained on more data, it becomes increasingly difficult to attribute their outputs to specific training inputs — a phenomenon they call "attribution decay." Even removing specific images entirely from training data doesn't prevent large models from reproducing similar content or styles. The findings complicate AI regulation and copyright litigation, while also raising questions about fair use and whether large-model outputs could be considered novel, creative works in their own right.
Generative AI is reshaping catastrophe modelling, but physics violations and commercial incentives could undermine the whole thing
Insurers are increasingly using generative AI, particularly diffusion models, to produce more precise catastrophe risk assessments by synthetically generating vast numbers of weather scenarios and sharpening spatial resolution beyond what traditional physics-based models can achieve. However, the technology carries risks, including AI "hallucinations" that can produce physically implausible events. Even where the science improves, there is a commercial tension, as insurers may favour models that produce lower loss estimates to justify writing more business rather than those that most accurately reflect risk.