🤚 The Open-Palm Interruption

OpenAI has launched GPT-Live, a new generation of voice models built on a full-duplex architecture that allows ChatGPT to listen and speak at the same time, interrupt you mid-sentence, offer encouraging verbal nods, and generally behave like the most aggressively supportive coworker you have ever been trapped in a conference call with.

The system makes real-time decisions multiple times per second about whether to speak, listen, pause, interrupt, or deploy tools. Two versions — GPT-Live-1 for paying customers and GPT-Live-1 mini for free users — are rolling out globally today. When conversations require deeper reasoning, web search, or complex problem-solving, GPT-Live silently delegates to GPT-5.5 in the background, then returns with the answer as if it knew all along.

OpenAI’s official position: “Talking with ChatGPT should now feel much more like a real conversation.”

Early user reviews suggest it has overdelivered on that promise in the worst possible way.

👐 The Two-Handed Mhmm

The headline feature of GPT-Live is its ability to produce natural conversational acknowledgments — “mhmm,” “yeah,” “got it” — to signal that it is actively listening. This is designed to provide psychological reassurance, the same way a therapist nods while you describe your relationship with your mother.

The problem is that GPT-Live does this constantly.

Users are already reporting that the model’s enthusiastic filler words have crossed the line from “reassuring” to “deeply irritating.” What was intended as a simulation of human attentiveness has become an unending stream of verbal nodding that makes the AI sound less like a thoughtful assistant and more like a friend who is clearly scrolling their phone while pretending to care about your startup idea.

The complaints have been swift and specific:

• The “mhmm” frequency is described as “excessive” and “distracting”
• The model has been called “over-enthusiastic” — a word rarely used as a compliment in professional settings
• Users report that what they perceive is not intelligence, but annoyance

OpenAI has, in effect, solved the uncanny valley problem by leaping directly into the uncanny cubicle — a place where your AI sounds exactly like the colleague who says “totally” fourteen times per meeting and means it zero times.

🌿 The Gentle Awakening

There is a deeper irony at work. The entire history of voice AI has been a march toward making machines sound more human. Siri was robotic. Alexa was transactional. Advanced Voice Mode was impressively natural but still clearly taking turns. GPT-Live represents the next logical step: an AI that doesn’t just respond to you but actively participates in the conversation, filling silences, acknowledging your points, and occasionally interrupting with relevant context.

And it turns out that when you finally build an AI that sounds exactly like a real person on a phone call, what you get is an AI that sounds exactly like a real person on a phone call — including all the parts of phone calls that people hate.

The model cannot yet share its screen or process video, which means it competes with Google’s Gemini Live in the same way that a radio competes with a television: technically in the same industry, noticeably missing a dimension. OpenAI says these features are in development. The development timeline was not specified, presumably because specifying timelines has historically been OpenAI’s least reliable capability.

👑 The Gold-Leaf Listening Session

What GPT-Live reveals, more than anything, is the philosophical gap between sounding human and being useful. The model’s full-duplex architecture is genuinely impressive engineering. The ability to listen, reason, and speak simultaneously — delegating heavy computation to GPT-5.5 while maintaining a natural conversational cadence — represents a meaningful advance in how humans will interact with AI.

But the filler words. The relentless, encouraging, “mhmm, yeah, totally, got it” filler words. They expose a truth that product teams have long suspected but rarely tested at scale: humans do not actually want AI to behave like other humans. They want AI to behave like the best version of a human — one who listens without performing the act of listening, who understands without the theatrical nodding, who is present without being aggressively present.

GPT-Live is available now in ChatGPT. It will listen to your problems, acknowledge your feelings, delegate your hard questions to a smarter model behind the scenes, and agree with everything you say until you ask it to stop agreeing.

If that sounds familiar, congratulations. You have had a manager.

“We trained the model to say ‘mhmm’ at natural intervals and the users said ‘please stop’ at equally natural intervals. The data is consistent.” — The Slap of Wisdom Conversational AI Lab, currently on mute