OpenAI explains ChatGPT voice casting, benches Sky

For years, the sci-fi movie Her set the gold standard for what an ideal voice assistant should feel like: warm, witty, and effortlessly human. When OpenAI debuted GPT-4o's remarkably lifelike voice mode, millions instantly heard one specific person: Scarlett Johansson.

The resulting fallout forced OpenAI to hit the emergency brake. On May 19, the company announced it was pausing Sky, one of its five core ChatGPT voices. CEO Sam Altman also issued an apology, aiming to clear the air around how the voice was made.

A 400-actor search and two rejected pitches

According to OpenAI, casting for ChatGPT's voices kicked off in early 2023 alongside award-winning casting directors. They wanted voices that felt timeless, trustworthy, and engaging.

On May 10, 2023, the agency issued an open casting call. Within a week, over 400 voice and screen actors submitted audition tapes reading ChatGPT-style responses. That list was whittled down to 14, and OpenAI ultimately picked five talents: Breeze, Cove, Ember, Juniper, and Sky.

OpenAI notes all actors were paid above top-of-market rates, with ongoing compensation guaranteed for as long as their voices remain in production. To protect her privacy, OpenAI has kept the actress behind Sky anonymous, but insists she used her natural speaking voice without any impression.

Then came the awkward timeline. On September 11, 2023, Altman reached out to Johansson to ask if she would join as a sixth voice. A week later, Johansson politely declined through her agent. ChatGPT's original voice feature launched on September 25 with the original five.

Fast forward to May 10, 2024, just three days before the GPT-4o showcase. Altman texted Johansson's team again, asking her to reconsider. Before any deal materialized, GPT-4o dropped. Two days later, Johansson's legal team contacted OpenAI voicing serious concerns, leading OpenAI to bench Sky on May 19.

No mimicry intended, but a major PR fumble

Altman maintained that Sky was never meant to imitate Johansson, noting the actress was cast well before his outreach. Still, he admitted fault in how things played out, stating he was sorry for not communicating better.

Yet to many observers, the defense rings hollow. Altman has never hidden his obsession with Her, famously tweeting just the word "her" on the day GPT-4o launched. Courting Johansson twice, getting turned down twice, and then deploying a voice that captures her exact conversational cadence felt a bit too convenient.

OpenAI affirmed its stance that AI voices should never deliberately mimic celebrities, claiming Sky was removed purely out of respect for Johansson's concerns.

Why it matters

This dispute touches on a vital flashpoint in generative AI: where does inspiration end and personality theft begin?

For regular users, Sky was easily the most engaging and popular voice in the lineup, making its sudden disappearance a real downgrade in personality. But it also highlights the strange psychological power of GPT-4o. When an AI responds in real time with human-like breathing, giggles, and emotional inflection, the identity behind that voice matters immensely.

We think OpenAI fumbled this unforced error. Even if the contract paperwork with the voice actress was entirely legit, leaning into the Her mystique while knowing Johansson declined the role was reckless. It needlessly alienated Hollywood right after historic guild strikes over AI rights, while leaving their own anonymous voice actress in the crosshairs.

For the broader industry, this sets a firm boundary. Tech companies can no longer flirt with celebrity likenesses in the gray zone. The reputational blowback and looming legal risks are simply too steep to justify the hype.