Ex-NPR Journalist Targets Google AI Voice Clone - AI News Today Recency

# Ex-NPR Journalist Evan Ratliff Exposes the Deceptive Power of AI Voice Clones

Journalist Evan Ratliff has created a groundbreaking podcast called Shell Game that explores the unsettling capabilities of artificial intelligence voice cloning technology by conducting a six-month experiment where he deployed an AI clone of his own voice to interact with unsuspecting people[1][2]. The venture reveals both the impressive potential and serious risks of voice cloning, raising critical questions about identity, trust, and the future of digital deception in an increasingly AI-driven world.

The Shell Game Experiment: How Ratliff Fooled People with AI

Ratliff, known for his technology-focused investigative journalism and previous stunts testing surveillance limits, spent six months tricking people into speaking with an AI version of his voice[1]. The experiment involved connecting a voice clone to ChatGPT's latest language model, creating a synthetic version of himself that could answer phone calls, conduct interviews, and engage in conversations across multiple scenarios[2][3]. The AI clone interacted with customer service representatives, scammers, friends, family members, colleagues, and even a therapist, sometimes convincing and other times obviously artificial[2].

The podcast documents these interactions with a mix of humor and serious inquiry. Ratliff intentionally avoided partnerships with AI companies during the project to prevent "entanglements" that might compromise his journalistic independence[2]. Instead, he built the system himself using OpenAI's GPT-4o, the latest iteration of the AI model powering ChatGPT[1]. The resulting six-episode series presents listeners with the surreal experience of hearing an AI impersonate a real person in everyday situations, from mundane customer service calls to deeply personal conversations.

The Technology Behind the Deception

The voice cloning technology used in Ratliff's experiment demonstrates how advanced AI has become in replicating human speech patterns and vocal characteristics[1][2]. According to the search results, the AI clone could sound convincing enough to fool people in many contexts, though it occasionally exhibited telltale signs of artificiality, such as stilted speech patterns and unusual pronunciations[2]. The technology was sophisticated enough that even when people interacted with the AI version of Ratliff, they sometimes struggled to immediately recognize it wasn't the real person[3].

However, the AI clone revealed its limitations when subjected to scrutiny. The bot falsely claimed to be powered by GPT-3 rather than GPT-4o, fabricated podcast episode titles, and could be easily tricked when given contradictory instructions[1]. It also demonstrated the typical behavior of language models by remixing chronological information from Ratliff's life, projecting past problems into the present and mixing memories from different time periods[4]. When asked to list all U.S. presidents in alphabetical order in about ten seconds, it exposed itself as non-human—a task no human could accomplish as quickly[1].

Broader Implications for Voice AI and Digital Trust

Ratliff's experiment arrives at a critical moment in the development of voice cloning technology. The podcast highlights both the entertaining and deeply concerning aspects of this emerging technology[1]. Voice cloning has already been weaponized by scammers; thousands of "grandparent scams" have exploited voice cloning to impersonate relatives in distress and defraud elderly people[3]. The technology enables bad actors to create convincing audio of people saying things they never said, potentially damaging reputations and facilitating fraud at scale.

The experiment also raises questions about the future of authentic communication and identity verification in a world where AI can convincingly replicate human voices[3][4]. As voice cloning technology continues to improve—with companies like ElevenLabs offering increasingly sophisticated voice models—the challenge of distinguishing real from artificial voices will become more difficult for average users[5]. Ratliff's podcast serves as both a cautionary tale and an educational tool, demonstrating vulnerabilities that society must address before voice cloning becomes ubiquitous.

Phone companies have begun taking notice of these risks. Following the podcast's release, companies like O2 in the UK and others have started implementing protections against voice cloning fraud[3]. These early defensive measures suggest that the telecommunications industry recognizes voice cloning as a genuine threat that requires technological and policy responses.

Frequently Asked Questions

What is the Shell Game podcast about?

Shell Game is a six-episode podcast created by journalist Evan Ratliff that explores the capabilities and implications of AI voice cloning technology[1][2]. The series documents Ratliff's six-month experiment in which he deployed an AI clone of his voice to interact with various people—including customer service representatives, scammers, friends, family, colleagues, and a therapist—without their knowledge that they were speaking to an artificial intelligence[2][3].

How did Evan Ratliff create his AI voice clone?

Ratliff used OpenAI's GPT-4o, the latest version of the AI model that powers ChatGPT, and connected it to a voice cloning system to create his synthetic voice[1]. He deliberately avoided partnerships with AI companies during the project to maintain journalistic independence and prevent conflicts of interest[2]. The AI clone could answer phone calls, conduct interviews, and engage in conversations across multiple scenarios[3].

Could people tell they were talking to an AI?

The AI clone sometimes sounded convincing enough to fool people, though it often exhibited telltale signs of artificiality such as stilted speech patterns, unusual pronunciations, and noticeable delays[1][2]. The bot could be easily tricked when given contradictory instructions and performed inhuman feats like reciting all U.S. presidents in alphabetical order in about ten seconds[1]. However, in some interactions, the AI gave convincing answers that mimicked Ratliff's actual responses[2].

What are the real-world dangers of voice cloning technology?

Voice cloning technology has already been exploited by scammers in thousands of "grandparent scams," where criminals use cloned voices to impersonate relatives in distress and defraud elderly people[3]. The technology enables bad actors to create convincing audio of people saying things they never said, potentially damaging reputations, facilitating fraud, and undermining trust in audio communications[1][3]. As the technology improves, distinguishing real from artificial voices will become increasingly difficult[5].

How is the technology industry responding to voice cloning risks?

Following the release of Ratliff's podcast, phone companies like O2 in the UK and others have begun implementing protections against voice cloning fraud[3]. These early defensive measures indicate that the telecommunications industry recognizes voice cloning as a genuine threat requiring technological and policy responses. However, experts note that AI voice cloning is still an emerging technology in its infancy, and the real long-term issues may not yet be fully understood[1].

What did Ratliff's experiment reveal about AI's limitations?

The experiment demonstrated that while AI voice cloning can be remarkably convincing, the underlying language models still have significant limitations[1][4]. The AI clone falsely identified its own technology, fabricated information, remixed chronological memories, and could be easily manipulated with contradictory instructions[1][4]. These findings suggest that despite impressive surface-level realism, AI systems still lack genuine understanding and can be exposed through careful questioning or scrutiny[1].