Autonomous driving AI fails when instructions change slightly — a reliability gap for real deployment

Researchers tested language-guided self-driving cars with slightly altered instructions — rephrased, ambiguous, or misleading — and found performance collapses on the same routes. This means the AI that works in clean test conditions becomes unreliable the moment real-world instructions get messy, vague, or contradictory.

Self-driving systems are being evaluated in simulation with perfect, well-formed instructions. But actual human commands are full of paraphrases, omissions, contradictions, and mistakes. This paper measures the gap between lab conditions and what deployment requires, and finds it's large. The AI doesn't fail because it can't drive — it fails because it can't understand what humans are actually asking it to do. That's a problem nobody was measuring until now.