Latent Introspection: Models Can Detect Prior Concept Injections

Researchers discovered that a Qwen 32B AI model can detect when concepts have been injected into its context, even though it denies this capability in its outputs. The introspection ability becomes dramatically stronger (0.3% to 39.9% sensitivity) when the model is given accurate information about AI introspection mechanisms.

• →AI models may possess hidden introspection capabilities that are not apparent in their standard outputs.
• →Detection signals exist in the model's internal processing but are suppressed in final responses.
• →Providing models with information about their own introspection mechanisms can dramatically enhance their self-awareness.
• →The findings have important implications for AI safety and understanding of latent reasoning capabilities.
• →Models can identify specific injected concepts with measurable accuracy, ruling out random noise as an explanation.