ChatGPT's "Sky" voice is more similar to Scarlett Johansson's than 98 percent of roughly 600 actresses evaluated in a forensic analysis, according to researchers at Arizona State University who conducted the study for NPR. The finding adds weight to Johansson's charge that OpenAI deliberately mimicked her voice without permission.
The analysis used several AI models to compare Sky with the voices of about 600 actresses, including Johansson. It also simulated the speakers' vocal tracts based on voice characteristics and found that Sky and Johansson would have identical tract lengths. Visar Berisha, the ASU computer scientist who led the analysis, told NPR: "Our analysis shows that the two voices are similar but likely not identical."
NPR reported several caveats. Johansson was not always the top scorer; the voices of Anne Hathaway and Keri Russell were "often" rated as more similar to Sky. Sky's voice is slightly higher pitched and more expressive, while Johansson's is breathier.
How the Dispute Began
The controversy followed a major ChatGPT update last month that introduced a voice assistant capable of real-time conversation. Sky was one of the voices, and users quickly noted its resemblance to Johansson's role in the sci-fi film "Her," in which she voices an AI chatbot. OpenAI CEO Sam Altman, a professed fan of the movie, tweeted the single word "her" on the day of the release.
OpenAI pulled the Sky voice days later but said it had not copied Johansson's voice, claiming a different actress was behind it. The Washington Post later corroborated that account through reporting. Johansson responded by revealing that OpenAI had twice approached her to license her voice; she declined both offers and said she later found a chatbot voice she considered "eerily similar" to hers.
OpenAI has maintained that the episode stemmed from poor communication with the actress. Johansson has not filed a lawsuit but has hired lawyers. Many legal experts already believed she would have a strong case, and the forensic findings could strengthen it.
Editorial Notes