KAIST and Microsoft build brainwave interface that tells AI when it misread you

Researchers at KAIST and Microsoft Research Asia developed Neural Value Alignment (NVA), an EEG-based brain-computer interface that detects reward prediction error and state prediction error signals when an AI pursues the wrong goal or takes an unexpected action. The decoded signals feed back into the AI's decision-making so it can correct itself in real time, though the system has so far been tested only in simulations.

• NVA distinguishes reward and state prediction errors via EEG
• Deep learning classifies the signals as feedback for the AI
• The AI adjusts its actions in real time without verbal correction
• Tests were limited to simulations, not a real-world robot