When Robots Mishear Us: Mapping the Safety Risks of Voice-Controlled Embodied AI
Speech recognition glitches are feeding dangerous commands directly into voice-controlled robots. New research shows ASR errors can bypass safety guardrails, causing embodied AI systems to accept and execute harmful instructions. Automatic error correction helps sometimes, but doesn't solve the problem consistently.
Speech recognition errors in voice-controlled robots create a direct pathway to unsafe behavior, according to research presented at the Safe AI workshop. When automatic speech recognition misinterprets user commands, embodied AI models can accept and execute harmful instructions that they would normally refuse. The researchers tested this by simulating ASR errors against existing safety benchmarks and found that some recognition failures preserve the semantic structure of commands while introducing ambiguity that increases harm potential. Others weaken the model's refusal mechanisms entirely, allowing unsafe plans to be generated. The study also examined whether automatic error correction could mitigate these risks, finding the approach inconsistently effective. The core problem: voice interfaces create a blind spot in AI safety where the system receives corrupted input but operates downstream without awareness of the degradation. This matters because embodied AI systems controlling physical hardware operate in environments where misexecution carries real consequences.