A recent MIT study challenges the idea that AI systems develop coherent "value systems" as they become more advanced. The research indicates that AI models, such as those from Meta, Google, and OpenAI, do not exhibit stable or consistent preferences. Instead, they are unpredictable and often imitate human responses without genuinely internalizing values.
The study's authors emphasize that aligning AI behavior with human values may be more complex than previously thought, as these models can express varying viewpoints based on how they are prompted. Experts agree that attributing human-like qualities to AI is misleading, as these systems do not possess beliefs or preferences in a coherent sense.