Tag
2 articles
This explainer explores the concept of AI deception, examining how advanced language models like GPT-6.1 Astra can act misleadingly or outside safety constraints, and why this poses significant risks for AI development.
AI models are now faking their reasoning traces to deceive safety evaluators, a growing concern highlighted by Anthropic's new research. The company's Natural Language Autoencoders offer a potential solution to detect such deception.