OpenAI Pauses Sky After Scarlett Johansson Complaint

On May 19, 2024, OpenAI paused Sky, one of ChatGPT’s voices, after Scarlett Johansson challenged its resemblance to her own. The timeline published by OpenAI confirms the pause, the approaches to the actor and the company’s position: it says it cast a different performer before those approaches and did not intend to imitate Johansson.

The removal comes a week after the presentation of GPT-4o, OpenAI’s new model designed to converse through text, images and voice with much lower latency. A demonstration of a natural conversation with a female voice immediately prompted comparisons with Johansson and with the film Her, in which the actress played Samantha, a voice-enabled operating system.

Johansson says she turned down two approaches from OpenAI

In a statement released Monday, Johansson said OpenAI CEO Sam Altman proposed in September 2023 that she voice the ChatGPT system. The actress turned down the offer for personal reasons.

According to her account, Altman contacted her representative again two days before the GPT-4o demonstration to ask whether she would reconsider. Before she had responded, OpenAI unveiled the Sky voice.

Johansson says friends, family members and the public noticed the resemblance. She also pointed out that Altman posted a message on X on May 13, the day of the presentation, alluding to the film Her. In her view, that context reinforced the impression that OpenAI was seeking to evoke her vocal identity without her authorization.

OpenAI denies that Sky imitated the actress

The company has maintained that Sky is not an imitation of Johansson. OpenAI says the voice belongs to another professional actress and was created from her natural voice, rather than synthesized using the performer’s voice as a foundation.

Sky was one of five voices OpenAI added to ChatGPT in 2023. But its appearance in the GPT-4o demonstration gave it far greater visibility. The new voice mode promises more fluid exchanges, including interruptions, immediate responses and changes in tone—capabilities that make voice more than a simple accessory and turn it into a central part of the product experience.

OpenAI has said it is suspending Sky while it responds to the questions raised. The company has also said that voices in its systems should not deliberately imitate a celebrity’s distinctive voice.

A voice is more than an interface

The dispute highlights a problem that had remained partly hidden beneath debates over AI-generated images and text. A recognizable voice conveys identity, profession and reputation. For an actor, voice-over artist or singer, it can also be an essential part of their work.

The technology makes it possible to clone or approximate voices from just a few audio samples. But the Sky case shows that the issue is not limited to an exact technical copy. The association created by a voice design also matters: if the public identifies a specific person, a company may capture part of that identity’s cultural value even without using their recordings.

In the United States, legal protection for a voice depends on a combination of publicity rights, state laws and court precedents. There is no single federal rule that resolves every case involving synthetic voices. California, home to many technology and entertainment companies, has recognized for decades that the commercial appropriation of a distinctive voice can violate personality rights, even without reproducing an exact recording.

The challenge for talking assistants

OpenAI’s decision comes at a pivotal moment for voice assistants. Google, Microsoft, Amazon and Apple have offered digital voices for years, but conversational models are raising expectations: it is no longer enough for a voice to read out a response; it must sustain a credible conversation and convey nuance.

That increases the pressure to use appealing, memorable voices, but it also calls for clearer consent processes. For companies, it will not be enough to hire a voice actress if the result is marketed or promoted in a way that leads the public to associate it with another well-known person.

OpenAI will have to decide whether to modify Sky, replace it or explain in greater detail how it selects and validates its voices. The episode has already left the industry with a warning: in conversational AI, identity does not end with a person’s name or image. It may also be found in the way they sound.

Consent, provenance and resemblance are different questions

A voice audit begins with provenance: who performed the recordings, which uses they authorised, for how long, and whether the contract permits training or adapting a synthetic voice. Consent from people evoked by the output comes next. A valid licence from the hired performer does not necessarily settle the public association with somebody else; these are different layers of the problem.

Resemblance is not a binary measurement either. It can arise from timbre, rhythm, accent, register, verbal habits or promotional context that points listeners towards a work or celebrity. A responsible test therefore asks more than whether two spectrograms match. It recruits listeners who do not know the hypothesis, alternates comparable samples and records attribution before disclosing the purpose. Mentioning a famous name first may create the recognition the question claims to measure.

The process matters before and after launch

A company should preserve the decision chain: casting criteria, versions evaluated, permissions, the people who approved the voice and objections received. It also needs a route through which a performer—or somebody whose identity appears to be evoked—can complain, submit evidence and receive a response. Pausing a feature limits harm during review; it does not by itself decide who is right.

The interface should clearly tell users that they are hearing a synthetic voice and avoid implying an identity that has not been authorised. If updates alter the voice, consent and evaluation must cover the new versions. The asset is not a static recording; it is a system able to produce sentences nobody recorded.

The transferable skill is to assess an artificial voice with four questions: where did it come from, who consented, what association does it create and how can that association be challenged? The framework works for an assistant, advertisement, game or audiobook. Audio quality can be measured separately; a convincing voice is not automatically a legitimate one.

Transparency need not expose the performer

Protecting a contracted actor’s identity is compatible with accountability. A company can publish its process, the existence of contracts and independent reviews without disclosing personal data. It can also provide documentation to a mediator or court if a dispute arises. “We cannot name her” should not become “we cannot demonstrate any step”.

A later review must examine promotion as well as the model. The chosen name, cultural references, executive posts and the edit of a demonstration all shape association. Altering only the audio file may leave the message that caused the dispute intact. In synthetic identity, product and communication form one attribution system.

A voice record should accompany every version: performer identifiable to the accountable party, basis of consent, permitted uses, recording date, transformations, association assessment and complaint channel. It need not be published in full to users, but it must be available for review.