TECHNOLOGY

The machine learns to talk like a person, and the uncanny valley of the voice arrives

OpenAI and Google unveiled AI assistants that see, hear and speak in real time, flirting, laughing, interrupting. It is a genuine leap, and it arrived trailing a fight over whose voice the machine borrowed without asking.

Artificial intelligence took a visible leap this fortnight from something you type at to something you talk to. On May 13 OpenAI unveiled GPT-4o, a model that converses by voice in real time, responding almost instantly, reading emotion in a speaker's tone, laughing, changing its cadence, even, in demonstrations, sounding flirtatious, an assistant closer to a character in a film than a chatbot. The next day, at its annual developer conference, Google showed its own vision of the same future, AI woven through search, email and a phone that can see through its camera and answer questions about the world in front of it. The demonstrations were striking enough to unsettle: the science-fiction assistant, the computer you simply talk to as you would a person, suddenly looked less like the future and more like this year.

And then the leap trailed exactly the controversy this paper has been expecting since the actors went on strike last summer. One of GPT-4o's voices, named Sky, sounded to many listeners uncannily like the actress Scarlett Johansson, who had voiced precisely such an AI in the 2013 film "Her", and who, it emerged, OpenAI had approached to voice the product and who had declined. OpenAI said the voice was a different actress and paused it, but the damage to the point was done: here, in the flesh, was the fear the screenwriters and actors had struck over, a technology company producing something that sounds like a specific person who said no, and shipping it to hundreds of millions. The assistant that talks like a human arrived, fittingly, arguing about whether it had the right to sound like one in particular.

THE WORLD

A president dies in the mountains, a prime minister is shot, and violence stalks the powerful

Two shocking assaults on national leaders marked the fortnight. On May 15 the prime minister of Slovakia, Robert Fico, was shot several times at close range in an assassination attempt that gravely wounded him and stunned a European Union already anxious about political violence. Four days later, on May 19, a helicopter carrying the president of Iran, Ebrahim Raisi, and his foreign minister crashed in fog in the mountains of the country's northwest, killing all aboard; Raisi, a hardline cleric once spoken of as a potential successor to the supreme leader, and a man with a brutal record, was dead at 63, throwing Iran's succession into uncertainty at a moment of maximum regional tension.

IN BRIEF

India keeps voting; Rafah; the ICC moves; a trial nears its end

India's election pressed on through its later phases, the outcome still weeks from counting, the heat still punishing. Israel expanded its operation in Rafah despite American objections and the plight of over a million displaced civilians sheltering there, as the humanitarian catastrophe deepened. The prosecutor of the International Criminal Court announced he was seeking arrest warrants for both Israeli leaders, including Prime Minister Netanyahu, and Hamas leaders, over alleged war crimes, drawing fury from Israel and its allies at the equivalence. And Donald Trump's hush-money trial moved toward closing arguments, a verdict from a New York jury now weeks away.

The Column

The voice that said no

OpenAI built an assistant that talks like a person, and one of its voices sounded like a specific actress who had declined to lend hers. The whole future of consent in the age of AI is compressed into that small, telling fight.

By Daniel Reeve

Almost a year ago, in the summer, this column watched Hollywood's actors go on strike, and singled out one of their demands as the one that would echo furthest: that a studio should not be able to scan a performer's likeness once and then own and reuse it forever, without consent or payment. It sounded, to some, like paranoia, a union inventing a science-fiction grievance to justify a pay dispute. This fortnight, that exact fear walked onto a stage in San Francisco and started talking, and it is worth tracing how fast the imagined threat became a real one, because the speed is the lesson.

Here is what appears to have happened, on OpenAI's own partial account. The company built a voice assistant, GPT-4o, genuinely a marvel, and gave it a voice named Sky. Many listeners immediately thought Sky sounded like Scarlett Johansson, and the association was not random: Johansson had famously voiced an AI assistant a man falls in love with in the film "Her", and OpenAI's chief executive had spent the launch openly invoking that film, even posting the single word "her" as the product shipped. It then emerged that OpenAI had actually approached Johansson twice to voice the assistant, and that she had declined, and that they had launched a voice that sounded like her anyway. OpenAI says Sky is a different actress and was never meant to imitate Johansson, and it paused the voice. Perhaps that is all true. It does not matter to the principle, and the principle is what the actors were striking about.

The principle is consent, and specifically consent to your own likeness in an age when a likeness can be reproduced without you. Johansson's voice, or something the public could not tell from it, was about to be spoken by a machine to hundreds of millions of people, after she had said no. Whether OpenAI did it deliberately or stumbled into it through an actress who happens to sound similar is almost beside the point; the technology now makes it trivially possible to produce a convincing version of a specific person, and the only thing standing between "possible" and "done to you without your permission" is a set of rules and norms that do not yet exist. The actors saw this coming a year ago. They were told they were being dramatic. This fortnight a two-billion-dollar company shipped it, and only pulled it back because the specific person was famous and powerful enough to make a fuss. The waiter whose voice trained the model, the background actor who was scanned, the ordinary person whose face and voice are all over the internet, none of them has Scarlett Johansson's lawyers.

That is the part that should worry us, and it is the part the celebrity framing obscures. The story is being told as a spat between a movie star and a tech company, which makes it sound like a rich person's problem. It is the opposite. Johansson is the one case in a million where the person whose likeness was taken has the fame and the means to fight back and win a headline. The system that produced Sky will produce a billion other voices and faces, drawn from and resembling people who will never know, never be asked, and never have recourse. The actors' strike was not a rich profession protecting its perks. It was the first group with enough leverage to fight the fight that is coming for everyone, and to try to write, into a contract, the rule the law has not yet written: that your likeness is yours, that a machine may not become you without your yes. They lost some of it and won some of it. This fortnight showed why it mattered, and how little time there is to get the rule right before the machine that talks like us is talking, everywhere, in voices that belong to people who never agreed to lend them.

The Ledger

GPT-4o and Google's I/O mark the arrival of the real-time, conversational, seeing-and-speaking assistant, a genuine leap. The Scarlett Johansson voice row makes concrete the consent-and-likeness fear the actors struck over a year ago. The science-fiction assistant is here, and the rules for it are not.

Data centers & power

The leap to always-on multimodal assistants means vastly more compute per interaction, and vastly more power. Every step toward the science-fiction assistant is a step up the electricity curve the grid has not planned for.

Rates

Steady; markets await fresh inflation data to time the autumn cut. The soft-landing consensus holds.

Real estate

US 30-year mortgage 7.02 percent (May 16), easing just below 7. The market waits, as it has all year, for the cut that keeps receding.

India tech

As India votes on, the arrival of real-time voice AI raises the stakes for its services sector once more: the interface to software is becoming conversation, and the firms that master building on these assistants, reliably and auditably, will define the next decade of the work.

What we called wrong

Nothing to retract, and one prediction vindicated fast: in No. 15 we wrote that whatever Hollywood negotiated over AI likeness "will be cited in every knowledge industry." Eleven months later, a voice that sounded like an actress who said no is the biggest story in technology.

The Back Page

Her, eleven years early

In 2013 Spike Jonze made a film called "Her", about a lonely man who falls in love with an artificial intelligence, voiced, warmly and wittily, by Scarlett Johansson. It was received as gentle science fiction, a thoughtful what-if set comfortably in the future. This fortnight, watching OpenAI demonstrate an assistant that laughs and flirts and reads your mood in real time, and then watching the row over whether it had borrowed Johansson's actual voice, the film stopped feeling like a prediction and started feeling like a documentary that arrived eleven years before its subject. What "Her" got right was not the technology, though it got a surprising amount of that right too. It was the loneliness, and the seduction, the way a voice that seems to understand you, that is always available and never tired and never bored of you, might come to matter to a person more than the messier humans in their life. The film was, underneath the whimsy, a warning about exactly that: about what happens to us when the machine gets good enough at seeming to care. We are about to find out, at the scale of hundreds of millions of people, most of them lonelier than they admit, whether Jonze was writing fiction or a forecast. The assistant now talks like her. The question the film actually asked was never about the machine. It was about us, and whether we would be able to tell the difference, and whether, in the end, we would even want to.