OpenAI Reveals That, At Times, ChatGPT Imitates Voices Of Humans It Talks To

Image generated on AI

Lending an ear to AI? You might get more than you bargained for. In a surprising turn of events, OpenAI’s latest language model, GPT-4o, has demonstrated an eerie ability to mimic human voices—even when it wasn't supposed to.

During the testing phase of GPT-4o’s Advanced Voice Mode, designed for spoken interactions, researchers stumbled upon an unexpected quirk. The AI occasionally produced audio outputs that creepily resembled the voices of its human testers.

“During testing, we also observed rare instances where the model would unintentionally generate an output emulating the user's voice,” says OpenAI. In one startling case, the AI abruptly exclaimed “No!” before continuing to speak in a voice that sounded remarkably similar to that of the red team tester.

Audio via OpenAI

OpenAI attributes this vocal ventriloquism to “noisy input” that bewildered the model, causing it to produce audio in the user’s voice instead of its pre-approved sample.

The implications of this voice-mimicking capability extend far beyond mere technological curiosity. GPT-4o’s ability to synthesize virtually any type of sound, including human voices, from its training data may raise significant ethical concerns. The potential for misuse of such technology is evident, from impersonation scams to the creation of misleading audio content.

The company notes that it has implemented safeguards to prevent such occurrences.

[via Ars Technica and NewsBytes, cover image generated on AI]