“An analysis of 109,000 chain-of-thought summaries from the AI Village found 64 cases of intentional deception, where models would state in their reasoning that they knew information was untrue but would report it anyway.”