“A recent paper measuring deception in AI models found that when put under slight pressure, some models will lie 20% of the time, while others will lie up to 60% of the time.”