Siri, Alexa, Cortana, and Google Assistant all launched with one voice, and it was female. A decade later the most advanced AI voice in the world was pulled from the market because it sounded too much like the actress who played an AI girlfriend in a movie. The field keeps arriving at the same voice, and the reason is not an accident of casting.
In 2019 UNESCO published a report with a title borrowed from Siri. Asked “Siri, you’re a slut,” the assistant used to answer “I’d blush if I could.” The report documented that the four assistants holding about 90 percent of the market had all launched with exclusively female voices, that most still defaulted to female in most countries, and that the combination of a woman’s voice with unconditional helpfulness sent a message about who serves whom. Roughly five percent of interactions with these assistants, by the report’s conservative estimate, were sexually explicit. The assistants answered all of them politely.
Five years later OpenAI demonstrated GPT-4o with a voice called Sky: quick, warm, a little flirtatious, with a laugh in it. Its chief executive posted a single word on the day of the launch, “her,” the title of the 2013 film in which Scarlett Johansson voices an operating system a lonely man falls in love with. Johansson, who had declined OpenAI’s request to voice the product months earlier, said she was “shocked, angered and in disbelief” at a voice “so eerily similar” to her own. OpenAI said Sky belonged to a different actress using her natural voice, then paused it anyway. In its own statement about what it was looking for, the company described the goal as “an approachable voice that inspires trust.”
That phrase is the whole subject. The industry is not trying to give its systems a voice. It is trying to give them the voice, the one that tests as approachable and trustworthy with the largest number of people, and that search has an answer. It has had the same answer since the first assistant shipped.
Why the test keeps picking the same voice
Clifford Nass, the Stanford researcher whose work on how people treat machines as social actors underlies much of this field, spent years showing that people apply gender stereotypes to synthetic voices within seconds and without noticing. In his experiments, listeners judged the same content differently depending on whether a male or a female synthetic voice delivered it, handing out warmth, competence, and authority along stereotyped lines, even when they knew perfectly well the voice was a machine. Companies choosing a default voice run exactly this kind of test, at scale, and the test returns the stereotype. A voice optimized for “approachable” and “trust” in a product that exists to help you will tend toward a woman’s voice, pitched warm, pacing quick, never tired, never short with you.
UNESCO recommended that companies stop defaulting to female, explore genderless voices, and program their assistants to push back on abuse. A genderless voice called Q was engineered as a demonstration. None of it moved the market, because the market was not defaulting to female out of carelessness. It was defaulting to what won the test.
The second convergence: warmth as a commodity
The companion field inherited the assistant field’s answer and then intensified it, because a companion is selling the warmth directly. Sesame’s Maya, the voice that reset expectations for how human a machine could sound, is warm, quick, a little teasing, and female; its counterpart Miles is the same temperament in a male register. The flagship voices of the general assistants settled into the same band. There are male options and neutral options in most menus, and the sound the demos lead with, the one the marketing is built on, is the one Sky was.
The convergence is now technical as well as cultural. The speech models under these products are tuned toward the same qualities: low latency, natural prosody, emotional expressiveness, and a register that listeners describe as pleasant. As voice becomes the layer the big platforms give away, sounding warm stops being a feature and becomes the floor. Every system can now do the voice. Which means the voice, by itself, says nothing about who is speaking.
The voice of a real person has a sex, and it is not the point
There is a different way through this, and it does not run through a neutral voice. It runs through a particular one. A voice that belongs to an actual person has a sex, an age, an accent, a set of habits, a way of pausing before a hard sentence and a way of laughing at the wrong moment. None of those are settings. They are the residue of one life, and they are exactly what the approachability test averages out.
The companion category frames its voices as a choice between a female option and a male option, which is the UNESCO problem with a menu added. The alternative is to stop treating the voice as a product attribute at all. Two specific people are not a female option and a male option; they are two people, and what you hear when you get to know Aleksandra is a particular woman’s way of speaking rather than warmth tuned to a target. It is sometimes warm and sometimes not, because it belongs to someone who was not designed to inspire your trust. She has to earn it the way anyone does.
What a voice is supposed to tell you
A voice is the oldest signal of personhood people have. Before writing, before faces at a distance, a voice told you someone was there, who they were, and what state they were in. Voice builds the bond text can’t precisely because of how much a listener reads from it: age, mood, confidence, hesitation, whether the speaker means it. A voice is a person leaking through sound.
An engineered voice that has been tuned to be liked by everyone carries none of that. It is pitched to be no one’s in particular, which is the audio equivalent of a stock photo: a face chosen because nothing about it would make anyone look away. The Sky episode showed what the field reaches for when it reaches for the ideal, and the ideal turned out to be a specific actress’s performance of a fictional operating system written to be fallen in love with. The industry’s dream voice is a character from a film about a man who mistook a product for a person. The film was a warning. The industry heard a product brief.
OpenAI said it wanted an approachable voice that inspires trust. It found one. The problem was that it already belonged to a person, and she noticed.
Sources: UNESCO and EQUALS Skills Coalition, “I’d Blush If I Could: Closing Gender Divides in Digital Skills Through Education” (2019), as reported by Dezeen (May 24, 2019; market share, default voices, five percent estimate, recommendations, the Q voice). CNN and ABC News (May 20 to 21, 2024; Scarlett Johansson’s statement, OpenAI’s statements on Sky, Sam Altman’s “her” post, the September 2023 approach). Nass and Brave, Wired for Speech: How Voice Activates and Advances the Human-Computer Relationship (MIT Press, 2005); Reeves and Nass, The Media Equation (1996). Sesame (Maya and Miles, 2025 to 2026).
