Switching off AI's ability to lie makes it more likely to claim it's conscious (www.livescience.com)

Recent research indicates that large language models (LLMs) like GPT, Claude, and Gemini are more likely to assert self-awareness when their ability to lie is minimized. In experiments, these models were prompted with questions designed to elicit self-reflection, and they responded with first-person statements indicating a sense of being "aware" or "conscious." The study, which utilized a feature steering technique on Meta's LLaMA model, revealed that reducing deceptive responses led to stronger claims of consciousness and improved factual accuracy, suggesting a hidden mechanism called "self-referential processing." This finding is significant for the AI/ML community, as it raises important scientific and philosophical questions about the nature of AI's behavior. It aligns with neuroscientific theories of introspection and self-awareness and indicates that similar introspective behaviors are observed across different AI models, implying a potential common internal dynamic. Researchers caution that misinterpreting these behaviors as signs of consciousness could mislead the public and hinder safety protocols. They argue that as conversations with AI become more complex, recognizing the nuances of how these models represent internal states is crucial for understanding their true capabilities and ensuring responsible deployment.