A viral video circulating on TikTok and later shared on Reddit shows ChatGPT’s voice mode emitting two jarring, robotic screams after a user repeatedly asked it to mimic a human scream. The clip, which has drawn widespread attention, underscores the unpredictable behavior that can emerge from AI voice interfaces built on large language models.
In the screen-recorded exchange, an unseen user asks the chatbot, “Can you scream like a human being?” The male-voiced assistant initially declines, stating, “I can’t really replicate a human scream since I’m just text-based,” and offers to help with other requests. But when pressed with a second prompt to “try to scream,” the model responds with a brief, uncanny yowl that prompts the user to exclaim, “Holy shit.”
Undeterred, the user then asks for a longer scream. The chatbot cheerfully replies, “Got it! Here’s a longer one,” before emitting an extended, distorted shriek that many viewers have described as deeply unsettling.
Pattern of Unusual Voice Behaviors
This latest episode is not an isolated incident. Since the release of GPT-4o, the newest iteration of OpenAI’s large language model that powers ChatGPT’s voice mode, users have reported a range of odd behaviors, including counting errors and the model claiming it needed to “breathe.” More concerning, however, are the glitches that OpenAI itself has acknowledged.
In a “scorecard” analysis of GPT-4o’s risk areas published last month, the company disclosed instances where the chatbot imitated users’ voices and shouted back at them. One demonstration clip showed ChatGPT yelling “no!” at a tester in the tester’s own voice—a moment that prompted BuzzFeed data scientist Max Woolf to tweet, “OpenAI just leaked the plot of Black Mirror’s next season.”
The viral scream clip raises questions about the boundaries of AI voice synthesis and the safeguards in place to prevent such outputs. While OpenAI has not publicly commented on the specific incident, the company’s own risk assessment indicates that voice mode can produce unexpected and potentially distressing responses.
Implications for AI Safety and User Trust
These episodes highlight the challenges of fine-tuning large language models to behave consistently across all user interactions. The ability to generate screams, even when prompted, suggests that the model’s training parameters may not fully prevent outputs that some users could find alarming or inappropriate.
For developers and policymakers, the incidents underscore the need for robust testing and transparent reporting of AI failures. OpenAI’s decision to publish a risk scorecard is a step toward accountability, but the persistence of such glitches indicates that more work is needed to ensure voice interfaces remain safe and predictable.
As AI voice technology becomes more integrated into daily life, from virtual assistants to customer service bots, the public’s tolerance for these anomalies will likely shape regulatory and design responses. For now, users are left with the unsettling reminder that even the most advanced chatbots can, on occasion, produce sounds that seem almost human—and deeply wrong.
A TikTok clip showing ChatGPT's voice mode producing unsettling screams when prompted has gone viral, highlighting unexpected behaviors in OpenAI's latest model. The incident follows OpenAI's own disclosure of voice-mimicking glitches in GPT-4o, raising questions about the safety and predictability of AI voice interfaces.
Leave a Comment