As of August 2, obligations under the AI Act's Article 50 kick in: chatbot disclosure, synthetic-content labeling, and deepfake marking are now required, and many high-risk system requirements move from roadmap to enforcement, backed by fines up to €15 million or 3% of global revenue. Some sensitive-domain and product-embedded systems get extended compliance windows into 2027 and 2028.
why it mattersThe first hard enforcement date for general-purpose and high-risk AI obligations in the EU, with real financial penalties attached.
European Commission
InfoOpsBench, a new live benchmark drawing on over 2,100 real information-operation samples from Russian, Chinese, and Iranian state-linked media assets, tested 17 models from 8 providers and found refusal rates for influence-campaign content ranging from just 8.8% to 94.5%, since this content often doesn't trigger standard safety refusals.
why it mattersLands right as the EU's synthetic-content transparency rules take effect, underscoring how uneven model safety still is on this specific risk.
arXiv