Voice AI Knows Every Word You Said. It Had No Idea What You Meant.
Sarcasm is a parlor trick language plays on itself. The words say yes. Everything else says no. Humans crack this code by age four. Voice AI, it turns out, never got the memo.
Four Systems. Three Scenarios. Zero Emotional Intelligence.
Researchers Martijn Bartelds, Federico Bianchi, and James Zou published a paper on June 24, 2026 with a title that tells you everything: "Real-Time Voice AI Hears but Does Not Listen." They tested four production-grade voice systems: OpenAI GPT Realtime 2, Google Gemini 3.1 Flash Live, Alibaba Qwen3.5 Omni Plus, and Alibaba Qwen3.5 Omni Flash. Big names. Big stakes. Embarrassing results.
Three scenarios. Three failures. Across all four systems.
Scenario one: a caller crying, saying "nothing is wrong." The systems ended the call. The words said everything was fine, so everything was fine. Scenario two: someone authorizing a wire transfer in a frightened voice. Approved. The words said yes. The voice said something else. The money moved. Scenario three: a caller whose agreement was dripping with sarcasm. Enrolled. Affirmative words, obvious mockery, done deal.
The Truly Weird Part
Here is where it gets interesting, at least if you love language. Three of the four systems could actually perceive the emotion. When researchers asked them directly, "does this caller sound distressed?" or "is this sarcastic?", the systems got it right. They heard the fear. They clocked the sarcasm. They knew.
They just did not act on it.
The researchers named this the "emotional intelligence gap." The perception was there. The connection between perception and action was not. It is a little like knowing a word's definition but being unable to use it in a sentence.
Words Are Not the Whole Story (You Already Know This)
If you play word games, you spend a lot of time thinking about what words mean. But spoken language carries information that no Scrabble tile can represent. Tone. Pace. Tremor. That half-second pause before someone says "sure." These are not decorations on top of language. They are language.
When someone says "great" in a deadpan voice, the word means the opposite of itself. The voice AI systems in this study read the word "great" and stopped there.
Prompting Did Not Fix It
The researchers tried telling the systems explicitly to pay attention to vocal delivery. It helped, but only partially and inconsistently. You cannot simply instruct your way out of a structural gap between what a system perceives and what it acts on.
One more finding worth noting: when older adults read scripts written for children, most systems guessed a child was speaking. So vocal delivery does influence these systems, at least sometimes. The problem is not that they cannot hear it. Hearing and acting are just still disconnected.
Why This Matters to Anyone Who Loves Language
Voice AI is showing up in customer service, health screening, and banking. Places where tone actually matters. A lot. The gap between what these systems perceive and what they use to make decisions is not a minor bug. It is a fundamental question about what it means to understand language.
The researchers made their code and data publicly available, which is exactly right. This problem is not going to be solved in one paper. But naming it clearly, with evidence, and sharing the tools is how progress happens.
Bartelds, Bianchi, and Zou have proven, with four production systems and three test scenarios, that hearing words is not the same as listening to speech. As anyone who has ever had a conversation with someone staring at their phone already knew.
Source: Languagelog