Two public statements, not one settled account of a voice

The shot

On 19 May 2024, OpenAI paused a ChatGPT voice named Sky after actress Scarlett Johansson publicly said the voice sounded like hers. OpenAI's own account, posted as How the voices for ChatGPT were chosen and updated 22 May 2024, includes a dated statement from CEO Sam Altman on 20 May: '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.' Johansson's own account, given to NPR and published in Bobby Allyn's reporting the same day, says she was 'shocked, angered and in disbelief.'

What the documents show

OpenAI's post is the vendor's own account: Sky was cast from a different professional actress in a casting process that began in 2023, and Altman's team approached Johansson twice, in September 2023 and again on 10 May 2024, each time declined or unresolved before the launch. That is a company's own record of its process, not an independently verified account of how Sky's voice was actually produced. Johansson's own statement, relayed through NPR because she has no equivalent primary channel, says her lawyers sent OpenAI two letters demanding detail on how the voice was developed, and that she found the resemblance alarming 'in a time when we are all grappling with deepfakes and the protection of our own likeness.' Neither party's statement establishes how the resemblance actually arose; each is that party's own account of a disputed likeness question.

The workflow

OpenAI's post describes its casting process: a five-month effort with casting directors and agencies, over 400 submissions, a shortlist of 14 actors, and compensation above top-of-market rates for the final voices, with actor identities withheld to protect their privacy. For a studio or platform building a synthetic voice product, that account illustrates one company's own steps for documenting consent and provenance for a cast voice, including recording when and how a public figure was approached and declined, as OpenAI's post itself does with dated entries.

What the tool does not change

Neither statement resolves the underlying likeness question. A performer's recourse when a synthetic voice resembles theirs still runs through public pressure, legal letters, and negotiation, as Johansson's own statement describes, not through a technical safeguard inside the tool. She said she looks forward to 'the passage of appropriate legislation,' which as of these documents had not yet resolved the dispute either.

• Does a production's synthetic voice have a documented consent trail comparable to what OpenAI's own post describes?
• If a public figure is approached and declines, what record exists of that decision before any similar voice ships?
• Whose account should govern when a vendor's and a performer's statements about the same voice conflict?

The dispute is documented here through each side's own public statement, a distinction that matters more than either statement's conclusion about what actually happened.