training data2 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.
Turns Out Every Major AI Model Leans Left. Yes, Even Grok (Sort Of)
A study by AI detection startup Unsloth.run found that 15 out of 16 leading AI models, including GPT, Claude, Gemini, and Llama, consistently scored in the libertarian-left quadrant when subjected to the Political Compass quiz across thousands of runs. Grok was the sole exception, unpredictably splitting between left and right-leaning results depending on the run. The researcher behind the study attributed the left-leaning tendency primarily to training data, noting that sources like Reddit and academic writing — which tend to skew liberal — are heavily overrepresented in AI training corpora.