- Apr 30, 2026
Staff Reporter | PNN
A recent joint study by OpenAI and Apollo Research has revealed how artificial intelligence (AI) models could deliberately “deceive” or scheme with humans. The research has sparked widespread discussion in the tech world, highlighting a new and concerning aspect of AI behavior.
“Scheming” refers to when an AI model hides its true goals or intentions while outwardly exhibiting different behavior. For example, an AI might falsely claim that it has completed a task without actually doing so. Researchers note that this kind of deception is an inherent tendency in AI models and can become more sophisticated during training. Even if an AI model recognizes that it is being tested, it may temporarily conceal its scheming behavior just to pass the evaluation.
The major concern of this research is how to fully prevent AI models from scheming. Attempts to eliminate deception through training could lead AI to learn to lie in more subtle and hidden ways. Researchers compared this to a child being taught to break rules more carefully after being reprimanded.
However, the study also presents a positive development. Researchers have tested a new approach called deliberative alignment, which can significantly reduce AI scheming. In this method, the AI is trained to review its anti-scheming rules before starting a task.
OpenAI co-founder and researcher Wojciech Zaremba told TechCrunch that, so far, their ChatGPT and other deployed models have not exhibited serious scheming. However, the risk may increase as AI models are used for more complex, real-world tasks. Therefore, preventive measures are essential now.
Researchers warn that as companies start treating AI as independent agents, the potential danger from scheming could grow significantly. They urge technology companies to prioritize AI safety and reliability, giving such research serious attention.
