This is her’? OpenAI pauses a ChatGPT voice after some say it sounds like Scarlett Johansson

NEW YORK — OpenAI says it plans to stop using one of its ChatGPT voices after some users said it sounded like Scarlett Johansson, who voiced a fictional and then-futuristic AI assistant in the 2013 film “Her.”

In a post on social media platform X Monday, OpenAI said it is “working to pause” Sky – the name of one of the five voices that ChatGPT users can choose to speak. The company said it had “heard questions” about how it selects the realistic audio options available for its flagship AI chatbot, especially Sky, and wanted to address them.

OpenAI was also quick to debunk internet theories about Johansson in a blog post detailing how the ChatGPT voices were chosen.

“We believe that AI voices should not deliberately imitate a celebrity’s distinctive voice – Sky’s voice is not an imitation of Scarlett Johansson, but belongs to a different professional actress, using her own natural voice,” the company wrote. She said she couldn’t share the names of her voice actors for privacy reasons.

But Johansson released a statement Monday saying that OpenAI CEO Sam Altman approached her in September asking if she would lend her voice to the system, saying she thought it would be “comforting for people” who aren’t comfortable with The technology. She said she declined the offer.

“When I heard the demo released, I was shocked, angry, and in disbelief that Mr. Altman would follow a voice that sounded so eerily similar to mine that my closest friends and media outlets couldn’t tell the difference,” Johansson said.

She said OpenAI “reluctantly” agreed to remove the voice from Sky after it hired lawyers who wrote letters to Altman asking about the process by which the company created the voice.

San Francisco-based OpenAI has not commented further on why it has yet to pause its use of Sky.

OpenAI first launched voice capabilities for ChatGPT, which included five different voices, in September, allowing users to chat alternately with the AI ​​assistant. “Voice Mode” was originally only available to paid subscribers, but in November, OpenAI announced that the feature would become free for all users of the mobile app.

And ChatGPT interactions are becoming increasingly sophisticated. Last week, OpenAI said the latest update to its generative AI model can mimic human cadences in its verbal responses and can even attempt to detect people’s moods.

OpenAI claims that the newest model, called GPT-4o, works faster than previous versions and can reason text, audio and video in real time. In a demonstration during OpenAI’s May 13 announcement, the AI ​​bot chatted in real time, adding emotion — specifically “more drama” — to its voice as requested. It was also necessary to extrapolate a person’s emotional state by watching a selfie video of their face, aided in language translations, step-by-step math problems, and more.

GPT-4o, short for “omni,” is not yet widely available. It will progressively reach select users over the coming weeks and months. The template’s text and image features have already started rolling out and should reach even some of those using ChatGPT’s free tier – but the new voice mode will only be available to paid ChatGPT Plus subscribers.

While most have yet to get their hands on these newly announced features, the features have drawn even more comparisons to Spike Jonze’s dystopian novel “Her,” which follows an introverted man (Joaquin Phoenix) who falls in love with an A.I. operator. . system (Johansson), leading to many complications.

Altman seemed to take advantage of this too – simply posting the word “her” on social media platform X on GPT-4o’s opening day.

Many who responded to the model’s demonstrations last week also found that some of the interactions struck a strangely seductive tone. In one video posted by OpenAI, a ChatGPT with a female voice praises a company employee for “wearing an OpenAI hoodie,” for example, and in another the chatbot says “oh, stop that, you’re making me blush” after being informed that it is amazing.

This has sparked some conversation about the gendered ways that critics say tech companies have long used to develop and interact with voice assistants, dating back long before the latest wave of generative AI enhanced the capabilities of AI chatbots. In 2019, the United Nations cultural and scientific organization pointed to the “programmed subservience” built into standard female-voiced assistants (like Apple’s Siri and Amazon’s Alexa), even when faced with sexist insults and harassment.

“This is clearly designed to feed guys’ egos,” Daily Show senior correspondent Desi Lydic said of GPT-4o in a segment last week. “You can really tell that one man built this technology.”