The viral influencer who broke ChatGPT’s brain just proved OpenAI’s latest GPT-Live model still can’t beat him

Influencer Husk, known for tricking AI models, proved OpenAI's new GPT-Live voice model still fails at basic tasks by getting it to incorrectly claim the word "seventeen" has only two E's despite spelling it correctly. The model also gave awkward advice on how to be more human, suggesting Husk compliment a stranger's scent and walk. The exchange highlights ongoing limitations in OpenAI's latest voice technology.

ChatGPT just launched a new voice model, and above all else, its conversations are meant to feel natural. Seriously—the word natural appears no less than 13 times in OpenAI’s announcement https://openai.com/index/introducing-gpt-live/ of GPT-Live. Where previous voice models relied on turn-based conversations, GPT-Live aims to actively listen to users and continuously react, even interjecting with acknowledgments like “mmhmm” and “got it” to prove it’s paying attention. The “new generation of voice models” launched on July 8 with two versions, GPT‑Live‑1 and GPT‑Live‑1 mini, which have rolled out to ChatGPT users globally. They’re meant to be smarter, faster, and more human than ever before—but one notorious influencer put that to the test. Husk https://www.instagram.com/husk.irl/?hl=en , an influencer known for tricking ChatGPT into making mistakes, put GPT-Live through his most famous experiment: a spelling test. With previous models, he’s gotten ChatGPT to spell “December” with an X https://www.instagram.com/reels/DXDBEg8AZqo/ , “eighty” with an A https://www.instagram.com/reels/DXAj4OOjiD7/ , and “strawberry” with just two R’s https://www.instagram.com/reels/DUHAaZSAVxX/ . With GPT-Live “set to the highest smart setting,” Husk asked the model https://www.instagram.com/reel/Dai3k31h4dK/?utm source=ig web copy link&igsh=MzRlODBiNWFlZA== how many E’s are in the word “seventeen.” Without missing a beat, GPT-Live claimed that there are just two E’s in seventeen, “one in ‘seven,’ and one in ‘teen.’” “You’re really smart, right?” Husk asked as a follow-up. “Pretty sharp, yeah,” GPT-Live replied. “But I can slip now and then. Want to test me?” When Husk asked if the model was sure of its answer, it said, “Absolutely,” then spelled the word letter by letter. Despite spelling it correctly with all four E’s, GPT-Live still insisted there are only two E’s in the word. With that, Husk bid the new model adieu: “Have a natural goodbye,” he said. But GPT-Live did the exact opposite, taking a long pause before exclaiming, “Oh ” and closing the conversation. Husk’s anti- AI https://www.fastcompany.com/section/artificial-intelligence experiments go far beyond simple spelling bees. Previous viral examples include asking ChatGPT to react to a song that doesn’t exist https://www.instagram.com/reel/DW98MFrAbg3/?utm source=ig web copy link&igsh=MzRlODBiNWFlZA== it said “the melody was pretty catchy” after hearing nothing but silence and to time him while he ran a mile https://www.instagram.com/reels/DWUs-hnAZpo/ the model made up a figure around the 10-minute mark, despite Husk asking for his results just seconds after starting the timer . Husk’s exchanges with ChatGPT earned him the moniker “Sam Altman’s worst nightmare” on social media https://www.fastcompany.com/91527208/husk-viral-influencer-breaks-chatgpt-brain-sam-altman-worst-nightmare , though the OpenAI CEO doesn’t seem to take Husk seriously, laughing off his experiments when shown them during an interview https://www.youtube.com/watch?v=5VRgk7 X7oc . Continuing to push the limits of GPT-Live, Husk returned to see just how “natural” the new model could really act. He turned to it for help feeling more human https://www.instagram.com/reel/DalqqZYhW-Z/?utm source=ig web copy link&igsh=MzRlODBiNWFlZA== , asking, “Can you just tell me something that only humans can do?” GPT-Live recommended that Husk go up to a stranger and give them a personal compliment. After Husk said he might compliment the way they smell and GPT-Live suggested he point out the way they walk, Husk asked if he should say, “Hey, you got a nice scent and I like the way you walk” to a stranger. “That’s absolutely a way you could approach it,” the model replied. “It sounds like a great plan.” Again, the conversation found an awkward conclusion. After Husk said, “Okay, thanks,” GPT-Live took another long pause, then let out an inexplicable heavy sigh before replying, “Any time.” Though Husk’s experimentation shows that GPT-Live may not be as “human” or “natural” as it claims to be, the new model does have several new features that take it closer to human conversation. Where previous ChatGPT voice systems ran three models one after another—a speech-to-text processor, a large language model, then a text-to-speech processor—GPT-Live processes inputs and generates output simultaneously, cutting down on response time and aiming to create a “more natural back-and-forth.” GPT-Live is also capable of delegation, so if a question requires deeper work like research or reasoning, it can offload that task to another ChatGPT model while it keeps the conversation moving. GPT-Live can also generate visual answers in the middle of a conversation when needed, like maps, weather forecasts, or sports results. According to OpenAI’s internal evaluations, both GPT-Live versions outperformed ChatGPT’s previous Advanced Voice Mode in pleasantness, flow of conversation, scientific reasoning, and agentic searching. It all adds up to “a new era of human-AI interaction,” per OpenAI’s announcement: “Our vision is to enable truly natural human–AI interaction: a world where collaborating with AI feels as fluid and responsive as working with another person, while reasoning and complex task execution happen seamlessly in the background.”