Can AI replace fact-checking? Tests of five latest AI models on 1,000 identical claims found their verdicts diverged in more than two-thirds of cases. Fact-checking service Lenz analysed agreement rates across major large language models and found all five reached the same conclusion in 328 cases, while 672 showed at least one differing judgment. Lenz said the study was not aimed at picking a best model and is running follow-up accuracy research with human labels.

Microsoft has unveiled MAI-Image-2-Efficient, a budget version of its AI image generation model. The model uses the same architecture as MAI-Image-2 but is positioned as cheaper and faster. Microsoft said it can generate 1024 by 1024 images on Nvidia H100 up to 22 percent faster while cutting costs by 41 percent. It also highlighted differences in output style and said the model is available via Microsoft Foundry and MAI Playground with specified API pricing.