OpenAI’s latest voice model, GPT-Live-1, pitched as the next step for ”natural human-AI conversations,” failed a basic task: counting the letter E in the word ”seventeen.” Instead of impressing users with smooth dialogue, the model answered incorrectly, sparking viral attention on social media. GPT-Live-1 is promoted as a faster, more lively way to chat with ChatGPT, but this glitch raised concerns.
GPT-Live-1 is designed to operate in full-duplex mode, meaning it can listen and speak simultaneously, adding brief confirmations like ”uh-huh” or ”yes” mid-conversation to mimic natural speech patterns. However, this feature backfired during a test by TikToker Husk, known for pushing ChatGPT’s voice capabilities to the limit. He asked how many times the letter E appears in ”seventeen.” The answer came back as ”two” when it should have been three.
The awkward moment didn’t end there. After Husk ended the chat, GPT-Live-1 fumbled its goodbye, delivering a clumsy sign-off. The clip quickly spread across X (formerly Twitter) and TikTok, drawing public scrutiny. OpenAI’s Developer Experience engineer Jason Liu reposted the video, responding with a blunt expletive-an unfiltered reaction that added fuel to the controversy. For a product launch, it felt more like a public relations stumble.
This isn’t Husk’s first time catching voice mode slip-ups. He previously went viral when ChatGPT struggled to set a timer-a clip that even prompted a reaction from OpenAI CEO Sam Altman. Now, a similar meme-worthy test surfaced when another user asked GPT-Live-1 to count the letter R in ”strawberry,” and the model again tripped up. Counting letters is a straightforward attention check that many text-based AI models-and rivals-handle flawlessly.
Users also complain that full-duplex makes ChatGPT overly talkative. Instead of enhancing conversational flow, the model frequently interrupts or inserts affirmations at awkward moments, making conversations feel more like interruptions than natural give-and-take. Rather than a smoother dialogue, some users have encountered something closer to an over-eager chatterbox.
The timing is unfortunate for OpenAI. The conversational AI field is heating up globally: Google is pushing Gemini Live as an Android-integrated assistant, Amazon is relaunching Alexa+ with generative AI features, and Apple is weaving new voice capabilities into its Apple Intelligence framework. Market researcher Markets and Markets predicts the conversational AI sector will top $40 billion by 2030. In this fiercely competitive environment, even small viral flaws quickly become weapons for competitors.
With GPT-Live-1’s launch marred by simple mistakes and an overbearing chat style, the key question now is whether OpenAI can quickly smooth out the glitches before rivals seize the narrative. As voice AI continues to mature, the battle won’t just be about raw power or language skills-it will hinge on who can deliver the most natural, trustworthy conversational experience.