OpenAI's use of Scarlett Johansson-like voice in ChatGPT exposed gaps in the law
Scarlett Johansson at the Madrid premiere for her movie, "Fly Me to the Moon." (Cover Images via AP Images)
If Ursula from Disney’s The Little Mermaid had artificial intelligence capabilities, she might have been able to snag Ariel’s voice without providing her with a pair of legs.
But that was then, and this is now.
Today, the villain is AI, and the only superpower needed to steal a voice is an audio sample, which can be fed into a neural network to imitate a voice. The generative model can then create new recordings.
In May, actress Scarlett Johansson released a statement saying she was shocked when OpenAI demonstrated a voice it called “Sky” that was “eerily similar” to her own. Johansson played a virtual assistant in the 2013 film Her. She said that last year, OpenAI asked to hire her to voice ChatGPT 4.0, but she declined.
OpenAI released a statement the same day as Johansson’s, saying: “The voice of Sky is not Scarlett Johansson’s, and it was never intended to resemble hers. We cast the voice actor behind Sky’s voice before any outreach to Ms. Johansson.” But less than a week after OpenAI announced ChatGPT 4.0, it removed the voice.
This conflict has raised questions about intellectual property, the right to publicity and AI regulation—a congressional subcommittee even invited Johansson to testify. If the actress decides to sue OpenAI, it’s expected to be a tricky case, as laws regulating AI usage of voices aren’t clear.
“There’s no singular federal right of publicity law, so it’s governed on a state-by-state basis,” says Ryan Sulkin, a partner and data protection group lead with Benesch Friedlander Coplan & Aronoff. He adds the state laws vary: Some provide protection for celebrities or well-known figures, while others provide protection for anyone.
Deepfake legalities
Federal bills have been introduced but never enacted, including the Deepfake Report Act of 2019 and the Deepfake Task Force Act, introduced in 2021. The Preventing Deepfakes of Intimate Images Act and the Deepfakes Accountability Act were both introduced in 2023 but have not been voted on, says Alan Wernick, a privacy and cybersecurity attorney with Aronberg Goldgehn Davis & Garmisa who recently participated as a panelist in an ABA webinar on the ethics of using AI in your legal practice.
There have been some laws and policies that have gone into effect at the federal level. In 2020, the Identifying Outputs of Generative Adversarial Networks Act was signed into law. It directs the National Science Foundation and the National Institute of Standards and Technology to support research on tools for detecting manipulated or synthesized content, says Visar Berisha, an engineering professor and associate dean of research and commercialization at Arizona State University.
Most recently, the Federal Trade Commission took a role in mitigating risks from voice cloning specifically. It ran the Voice Cloning Challenge earlier this year with an aim of spurring technical innovation in this area.
Additionally, new technologies that use AI to detect and watermark AI or human-generated content have emerged, Berisha says.
“I think that these are good initial steps, but I think we need new top-down policy recommendations and new regulations to address this issue,” he says.
Generally, however, the law broadly says the use of the voice in connection with a commercial product or service is not permitted unless the person whose voice is being used consents to it, Sulkin says.
To prove a voice was copied without consent would be a very fact-specific inquiry if it went to litigation, Sulkin says. For example, there would be discovery during litigation into how the voice was created; what was used to train the AI; and what the intent of the creator was, he says.
Still, there’s a significant blind spot here, Berisha says. There’s increasing evidence that tools to detect AI-generated content won’t be as effective over time.
“In fact, OpenAI had a tool to detect AI-generated text and then took it down because it didn’t work very well,” Berisha says. “The watermarking solutions being implemented by companies are great for good actors, but what about bad actors with malicious intent?”
OpenAI has since developed yet another tool to detect AI-generated text that has been ready for release since 2023 but has not released it, the Wall Street Journal reported in August.
Some already are using deepfakes to convince others to fork over money or passwords, leading to fraudulent activities pertaining to identity theft. Individuals will try to mimic other voices to gain access to a bank account, to request a particular transaction or even to convince a loved one that their relative is in trouble.
Plus, Sulkin adds, “AI’s ability to imitate and fake a real person is just going to get better and better over time.”
Synthetic media
There are different methods to protect your voice, but nothing is foolproof. Celebrities or recognized personalities should confirm that voice rights are included in their contracts. For regular people (and celebrities), a good first step if you believe your voice is being used by AI without permission is to generate objective evidence of similarity between the real voice and the synthetic voice, Berisha says.
One emerging approach is watermarking human-generated content. Instead of relying on AI-based detectors with variable degrees of success or relying on the goodwill of companies that generate deepfakes to watermark synthetic media—watermarking human-generated content empowers the people themselves to certify audio they create as “authentically human,” Berisha explains. “This is an area that we’re working on and are very excited about.”
But in the meantime, it becomes a case of discovering that your voice has been copied—and writing a cease-and-desist letter, Sulkin says. “There’s no perfect solution.”
This story was originally published in the October/November 2024 issue of the ABA Journal under the headline: “Lost in Translation? OpenAI’s move to add a Scarlett Johansson-like voice to ChatGPT exposed gaps in the law.”
