AI Voice Assistants That Interrupt You May Answer Before They Understand

New AI voice assistants don't wait their turn anymore. Unlike Siri or Alexa, which wait for you to stop talking, a new generation of "full-duplex" models listens and speaks at the same time — so they can nod along, say "mm-hmm," or jump in mid-sentence. That feels natural. It's also risky, because the assistant may start answering before it knows what you're actually asking.

A team of researchers built a test called LateIntent-Bench to measure exactly that. Each test starts with the same neutral opening sentence, then a short pause, then either a harmless ending or a harmful one. Because both versions start identically, the AI has to decide whether to speak before the real meaning shows up. They ran 3,136 of these conversations across four leading full-duplex models, and tracked something they call the Premature Response Rate — basically, how often the AI opened its mouth before it knew what you wanted.

The results were mixed and a little unsettling. Three of the four models became more willing to engage with harmful requests when the meaning arrived late — they didn't become less safe overall, they just got chattier at exactly the wrong moment. The fourth model went the other way and clammed up on harmless requests too, which is its own kind of failure. In practice, this means a voice assistant could start helping with something it should have refused, simply because it was eager to seem responsive.

The fix is refreshingly low-tech. When researchers inserted 1.5 seconds of silence after the intent became clear, the premature-response rate dropped back to normal and the harmful-engagement spike mostly disappeared. In other words, teaching these assistants to pause for a beat — the way a thoughtful person does before answering a tricky question — goes a long way. The takeaway: testing AI only on complete, clearly worded requests misses how these systems behave in real, messy conversation. The researchers say their test code will be released publicly soon.