- OpenAI approached Scarlett last fall, and she refused.
- Two days before the GPT-4o launch, they contacted her agent and asked that she reconsider. (Two days! This means they already had everything they needed to ship the product with Scarlett’s cloned voice.)
- Not receiving a response, OpenAI demos the product anyway, with Sam tweeting “her” in reference to Scarlett’s film.
- When Scarlett’s counsel asked for an explanation of how the “Sky” voice was created, OpenAI yanked the voice from their product line.
Perhaps Sam’s next tweet should read “red-handed”.
This statement from scarlet really changed my perspective. I use and loved the Sky voice and I did feel it sounded a little like her, but moreover it was the best of their voice offerings. I was mad when they removed it. But now I’m mad it was ever there to begin with. This timeline makes it clear that this wasn’t a coincidence and maybe not even a hiring of an impressionist (which is where things get a little more wishy washy for me).
The thing about the situation is that Altman is willing to lie and steal a celebrity's voice for use in ChatGPT. What he did, the timeline, everything - is sleazy if, in fact, that's the story.
The really concerning part here is that Altman is, and wants to be, a large part of AI regulation [0]. Quite the public contradiction.
Everyone is so mad about them stealing a beloved celebrity’s voice. What about the millions of authors and other creators whose copyrighted works they stole to create works that resemble and replace those people? Not famous enough to generate the same outrage?
Same here and that voice really was the only good one. I don't know why they don't bring the voices from their API over, which are all much better, like Nova or Shimmer (https://platform.openai.com/docs/guides/text-to-speech)
I had to go look at what voice I picked once I heard the news, it was Sky. I listened to them all and thought it sounded the best. I didn’t make any connection to her (Scar Jo or the movie) when going through the voices, but I wasn’t listening for either. I don’t think I know her voice well enough to pick it out of a group like that.
Maybe I liked it best because it felt familiar, even if I didn’t know why. I’m a bit disappointed now that she didn’t sign on officially, but my guess is that Altman just burned his bridge to half of Hollywood if he is looking for a plan B.
If they really hired someone who sounds just like her it's fair game IMO. Johanssen can't own the right to a similar voice just like many people can have the same name. I think if there really was another actress and she just happens to sound like her, then it's really ok. And no I'm not a fan of Altman (especially his worldcoin which I view as a privacy disaster)
I mean, imagine if I happened to have a similar voice to a famous actor, would that mean that I couldn't work as a voice actor without getting their OK just because they happen to be more famous? That would be ridiculous. Pretending to be them would be wrong, yes.
If they hired someone to change their voice to match hers, that'd be bad. Yeah. If they actually just AI-cloned her voice that's totally not OK. Also any references to the movies. Bad.
But it's clearly not her voice right? The version that's been on the app for a year just isn't. Like, it clearly intending to be slightly reminiscent of her, but it's also very clearly not. Are we seriously saying we can't make voices that are similar to celebrities, when not using their actual voice?
Why are you mad? We have no rights to the sound of our voice. There is nothing wrong with someone or something else making sounds that sound like us, even if we don’t want it to happen.
Which is what makes me wonder if this might grow into a galvanizing event for the pro-creator protests against these AI models and companies. What happened here isn't particularly unique to voices or even Scarlett Johansson, it is just how these companies and their products operate in general.
OpenAI only hires and is built on the culture that data and copyright is somehow free for the taking, otherwise they would have zero ways to make a profit or “build agi”
To each their own. I personally didn't get Scarlett Johansson vibes from the voice on the GPT-4o demo (https://openai.com/index/hello-gpt-4o/) even though I'm a huge fan of hers (loved Her, loved Jojo Rabbit, even loved Lucy, and many many others) and have watched those and others multiple times. I'd even say I have a bit of a celebrity crush.
To me it's about as close to her voice as saying "It's a woman's voice". Not to say all women sound alike but the sound I heard from that video above could maybe best be described and "generic peppy female American spokesperson voice"
Even listening to it now with the suggestion that it might sound like her I don't personally hear Scarlett Johansson's voice from the demo.
There may be some damming proof where they find they sampled her specifically but saying they negotiated and didn't come to an agreement is not proof that it's supposed to be her voice. Again, to me it just sounds like a generic voice. I've used the the version before GPT-4o and I never got the vibe it was Scarlett Johansson.
I did get the "Her" vibe but only because I was talking to a computer with a female voice and it was easy to imagine that something like "Her" was in the near future. I also imagined or wished that it was Majel Barrett from ST:TNG, if only because the computer on ST:TNG gave short and useful answers where as ChatGPT always gives long-winded repetitive annoying answers
The same egomaniac tendencies that cause people like Elon Musk or Paul Graham to post the first dumbass thing that comes to their mind because they think everyone absolutely has to see how smart and witty they are.
There were never any. None of the models or code are actually open. It claims to be a nonprofit but is effectively a for profit company pulling the strings of a nonprofit just to avoid taxes.
Could they have made it look less like Midler vs Ford?
"Midler was asked to sing a famous song of hers for the commercial and refused. Subsequently, the company hired a voice-impersonator of Midler and carried on with using the song for the commercial, since it had been approved by the copyright-holder. Midler's image and likeness were not used in the commercial but many claimed the voice used sounded impeccably like Midler's."
As a casual mostly observer of AI, even I was aware of this precedent
- Two days before the GPT-4o launch, they contacted her agent and asked that she reconsider. (Two days! This means they already had everything they needed to ship the product with Scarlett’s cloned voice.)
New voice mode is a speech predicting transformer. "Voice Cloning" could be as simple as appending a sample of the voice to the context and instructing it to imitate it.
If they really did that then (A) it's not much better (B) they didn't even wait for an answer from Johansson (C) it's extraordinarily reckless to go from zero to big-launch feature in less than two days.
OpenAI first demoed and launched the “Sky” voice in November last year. The new demo doesn’t appear to have a new voice.
I doubt it would take them long to prepare a new voice, and who’s to say they wouldn’t delay the announcements for a ScarJo voice?
A charitable interpretation of the “her” tweet would be a comparison to the conversational and AI capabilities of the product, not the voice specifically, but it’s certainly not a good look.
I believe that "Sky" voice was first released in September last year and according to the blog post released by OpenAI they were working with "Sky" voice actress months before even contacting Scarlett Johansson for the first time.
AFAIK they yanked it pretty quickly and the subsequent scandal has widely informed people that it was not authorized by Scarlett Johansson. So while it was clearly a violation resulting from a sequence of very stupid decisions by OpenAI, I am not sure if there would be much in the way of damages.
Johansson is rich. The real value she could get from this would be as an advocate for the rights of creatives, performers and rights holders in the face of AI. If this goes to discovery OpenAI is done.
How much do you think Disney or Universal Music or Google or NYT would give to peek inside OpenAI's training mixture to identify all the infringing content?
The funniest thing to me is the very fact that they wanted this voice so much that they though it was worth the hassle (so obvious in hindsight). I mean, I get it, the movie reference was welcome by their auditory (which is kinda cringey on it's own, to be honest, but ok, that's just reality). But it's not even a very pleasant voice, I remember people chuckling at the choice of Johansson's voice for the role back then. It was clearly a choice dictated by the fact that she is famous. It was marketing, the movie producers needed her name, not the voice. And OpenAI couldn't even use the name. It all just seems so silly now.
…Unless, of course, all this scandal isn't also a part of marketing campaign.
They underestimated how quickly people would take off the headphones and jump on the bandwagon to claim affinity with an injured celebrity.
Are you suggesting they should have engineered the voice actress' voice to be more distinct from another actress they were considering for the part? Or just not gone near it with a 10ft pole? because if the latter the studios can just release a new Her and Him movie with different voices in different geo regions and prevent anyone from having any kind of familiar engaging voice bot.
Probably 5-10 minutes worth of dataset and GPU time for finetuning on an existing base model. Could be done on a Blu-ray rip or an in-person audition recording, legality and ethics aside.
> Don't here any arguments on how this is fair-use.
> Why?
Because it's a right of personality issue, not copyright, and there is no fair use exception (the definition of the tort is already limited to a subset of commercial use which makes the Constitutional limitation that fair use addresses in copyright not relevant, and there is no statutory fair use exception to liability under the right of personality.)
It's also worth noting that Sam Altman admitted that he had only used GPT4o for one week before it was released. It's possible that in the rush to release before Google's IO event, they made the realisation of the likeness of the voice to Scarlett Johansen way too late hence the last minute contact with her agent.
yeah, that was just poking the hornets nest. Even if i wasn’t mad enough to make a stink over my voice before, plausible deniability and all, that would’ve sealed the deal for me.
ChatGPT is way better than to need stupid ripoffs than this
Sam should be ashamed to have ever thought of ripping off anyone's voice, let alone done it and rolled it out.
They are building some potentially world-changing technology, but cannot rise above being basically creepy rip-off artists. Einstein was right about requiring improved ethics to meet new challenges, and also that we are not meeting that requirement.
It's part and parcel of the LLM field's usual disdain for any property rights that might belong to other people. What they did here is not categorically different than scraping every author and visual artist on the internet - but in this instance they've gone and brazenly "copied" (read: stolen) from one of the few folks with more media clout than they themselves have.
People hire celebrity voice impersonators all the time. You've heard a few impersonators this month probably from ads. This is such a non-issue that's only blowing up because Johannsen wrote a letter complaining about it and because people love "big tech is evil" stories.
I know I have a reputation as an OpenAI hater and I understand why: it’s maybe 5-10% of the time that the news gives me the opportunity to express balance on this.
But I’ve defended them from unfair criticism on more than a few occasions and I feel that of all the things to land on them about this one is a fairly mundane screwup that could be a scrappy PM pushing their mandate that got corrected quickly.
The leadership for the most part scares the shit out of me, and clearly a house-cleaning is in order.
But of all the things to take them to task over? There’s legitimately damning shit this week, this feels like someone exceeded their mandate from the mid-level and legal walked it back.
I wouldn't necessarily call that damning. "Soundalikes" are very common in the ad industry.
For example, a car company approached the band sigur ros to include some of their music in a car commercial. Sigur ros declined. A few months later the commercial airs with a song that sounds like an unreleased sigur ros song, but really they just paid a composer to make something that sounds like sigur ros, but isn't. So maybe openai just had a random lady with a voice similar to Scarlett so the recording.
Taking down the voice could just be concern for bad press, or trying to avoid lawsuits regardless of whether you think you are in the right or not. Per this* CNN article:
> Johansson said she hired legal counsel, and said OpenAI “reluctantly agreed” to take down the “Sky” voice after her counsel sent Altman two letters.
So, Johansson's lawyers probably said something like "I'll sue your pants off if you don't take it down". And then they took it down. You can't use that as evidence that they are guilty. It could just as easily be the case that they didn't want to go to court over this even if they thought they were legally above board.
There have been several legal cases where bands have sued advertisers for copying their distinct sound. Here are a few examples:
The Beatles vs. Nike (1987): The Beatles' company, Apple Corps, sued Nike and Capitol Records for using the song "Revolution" in a commercial without their permission. The case was settled out of court.
Tom Waits vs. Frito-Lay (1988): Tom Waits sued Frito-Lay for using a sound-alike in a commercial for their Doritos chips. Waits won the case, emphasizing the protection of his distinct voice and style.
Bette Midler vs. Ford Motor Company (1988): Although not a band, Bette Midler successfully sued Ford for using a sound-alike to imitate her voice in a commercial. The court ruled in her favor, recognizing the uniqueness of her voice.
The Black Keys vs. Pizza Hut and Home Depot (2012): The Black Keys sued both companies for using music in their advertisements that sounded remarkably similar to their songs. The cases were settled out of court.
Beastie Boys vs. Monster Energy (2014): The Beastie Boys sued Monster Energy for using their music in a promotional video without permission. The court awarded the band $1.7 million in damages.
> I wouldn't necessarily call that damning. "Soundalikes" are very common in the ad industry.
As are disclaimers that celebrity voices are impersonated when there is additional context which makes it likely that the voice would be considered something other than a mere soundalike, like direct reference to a work in which the impersonated celebrity was involved as part of the same publicity campaign.
And liability for commercial voice appropriation, even by impersonation, is established law in some jurisdictions, including California.
Sorry, that's apples-to-pizzas comparison. You're conflating work and identity.
There's an ocean of difference between mimicking the style of someone's art in an original work, and literally cloning someone's likeness for marketing/business reasons.
You can hire someone to make art in the style of Taylor Swift, that's OK.
You can't start selling Taylor Swift figurines by the same principle.
What Sam Altman did, figuratively, was giving out free T-Shirts featuring a face that is recognized as Taylor Swift by anyone who knows her.
The damning part is that they tried to contact her and get her to reconsider their offer only 2 days before the model was demoed. That tells you that at the very least they either felt a moral or legal obligation to get her to agree with their release of the model.
Altman often uses tactical charisma to trap gullible people, government entities, and any unsuspecting powerful person for his ends. He will not bat an eyelid to take whatever unethical route if that gives him "moat". He relentlessly talks as if "near-term AGI" is straining to get out of the bottle in his ClosedAI basement. He will tell you with great concern about how "nervous" or "scared" (he said this to the US Congress[1]) of what he thinks his newest LLM model is gonna let loose on humanity.
So he's here to help regulate it all with an "international agency" (see the reference[2] by windexh8er in this thread)! Don't forget that Altman is the same hack who came up with "Worldcoin" and the so-called "Orb" that'll scan your eyeballs for "proof of personhood".
Is this sleazy marketer the one to be trusted to lead an effort that has a lasting impact on humanity? Hell no.
> He will tell you with great concern about how "nervous" or "scared" (he said this to the US Congress[1]) of what he thinks his newest LLM model is gonna let loose on humanity.
It’s so refreshing to hear someone else actually understand this sentiment for what it is— snake oil sales. The normies out there eat this up, and there is no convincing them otherwise because of how powerful the AI trope in entertainment media is.
Honestly though, if you actually listen to him and read his words he seems to be even more devoid of basic empathetic human traits than even Zuckerberg who gets widely lampooned as a robot or a lizard.
I agree. By tactical charisma, I didn't mean to imply that he has genuine empathy. I mean that he says things the other person finds pleasing, in just the right words, and credible-sounding seriousness. Tactical in a tempting sense: "Don't you want to be the bridge between man and machine, Scarlett?" or, "Imagine comforting the whole planet with your voice" — I've slightly rephrased a bit here, but this is how he tried to persuade Scarlett Johannson into "much consideration" (her words).
Yes, I've listened to Altman. A most recent one is him waffling with a straight-face about "Platonic ideals"[1], while sitting on a royal chair in Cambridge. As I noted here[2] six months ago, if he had truly read and digested Plato's works, he simply will not be the ruthless conman he is. Plato would be turning in his grave.
Emotional intelligence /true empathy cannot be learned or acquired, at least IMHO.
But it can be learned to be mimicked almost to perfection, either by endless trial & error or by highly intelligent motivated people. It usually breaks apart when completely new intense / stressful situation happens. Sociopaths belong here very firmly and form majority.
If you know what to look for, you will see it in most if not all politicians, 'captains of industry' or otherwise people who got to serious power by their own deeds.
Think about a bit - what sort of nasty battles they had to continually keep winning with similar folks to get where they are, this ain't the place for decent human beings, you/me would be outmatched quickly. Jordan Peterson once claimed you have cca 1/20 of sociopaths in general population, say 15 millions just in US? Not every one is highly intelligent and capable of getting far, but many do. Jobs, Gates, Zuckenberg, Bezos, Musk, Altman and so on and on. World is owned and run by them, I'd say without exception.
Well, this confirms that OpenAI have been shooting from the hip, not that we needed much confirmation. The fact that they repeatedly tried to hire Johansson, then went ahead and made a soundalike while explicitly describing that they were trying to make it be like her voice in the movie … is pretty bad for them.
"when voice is sufficient indicia of a celebrity's identity, the right of publicity protects against its imitation for commercial purposes without the celebrity's consent."
Seems like sama may have put a big hole in that argument when he tweeted "her", now it is very easy to say that they knowingly cloned ScarJo's likeness. When will tech leaders learn to stop tweeting.
Whether or not what they've done is currently technically illegal, they're priming the public to be pissed off at them. Making enemies of beloved figures from the broader culture is likely to not make OpenAI many friends.
OpenAI has gone the "it's easier to ask forgiveness than permission" route, and it seemed like they might get away with that, but if this results in a lot more stories like this they'll risk running afoul of public opinion and future legislation turning sharply against them.
I mean, unless an investigation can find any criteria used to select this particular actress like "sounds like Scarlett" in an email somewhere, or you know, the head idiot intentionally and publicly posting the title of a movie starring the actress in relation to the soundalike's voice work.
This is so pointless and petty too. Like "hee hee our software is just like the movies". And continuing the trend of tech moguls watching bleak satire and thinking it's aspirational.
Yes, absolutely illegal. You don't need to copyright anything, you simply own the rights your own likeness -- your visual appearance and your voice.
A company can't take a photo from your Facebook and plaster it across an advertisement for their product without you giving them the rights to do that.
And if you're a known public figure, this includes lookalikes and soundalikes as well. You can't hire a ScarJo impersonator that people will think is ScarJo.
This is clearly a ScarJo soundalike. It doesn't matter whether it's an AI voice or clone or if they hired someone to sound just like her. Because she's a known public figure, that's illegal if she hasn't given them the rights.
(However, if you generate a synthetic voice that just happens to sound exactly like a random Joe Schmo, it's allowed because Joe Schmo isn't a public figure, so there's no value in the association.)
Tom Waits successfully sued Frito Lay for using an imitator without approval in a radio commercial.
The key seems to be that if someone is famous and their voice is distinctly attributeable to them, there is a case. In both of these cases, the artists in question were also solicited first and refused.
I’m not a lawyer and don’t have any deep background this area of IP, but there is at least some precedent apparently:
> In a novel case of voice theft, a Los Angeles federal court jury Tuesday awarded gravel-throated recording artist Tom Waits $2.475 million in damages from Frito-Lay Inc. and its advertising agency.
> The U.S. District Court jury found that the corn chip giant unlawfully appropriated Waits’ distinctive voice, tarring his reputation by employing an impersonator to record a radio ad for a new brand of spicy Doritos corn chips.
The problem is they pursued, was rejected, then approximated. Had they just approximated and made no references to the movie…then I bet social marketing would have made the connection organically and neither Ms Johansson or the Her producers would have much ground because they could reasonably claim that it was just a relatively generic woman’s voice with a faint NY/NJ accent.
It's not precisely copyright, but most states recognize some form of personality rights, which encompass a person's voice just as much as the person's name or visual appearance.
Not here to weigh in on the answers to these questions. But it certainly feels pretty scary to have to ask such questions about a company leading the LLM space, considering the U.S. currently has little to no legal infrastructure to reign in these companies.
Plus the tone of the voice is likely an unimportant detail to theor success. So pushing up against the legal boundaries in this specific domain is at best strange and at worst a huge red flag for their ethics and how they operate.
This is known as personality rights or right to publicity. Impersonating someone famous (eg faking their likeness or voice for an ad) is often illegal.
I honestly didn't think it sounded like Johansson. Because of the controversy I just now re listened to the demos and I still find if very unlikely that someone would think it was her.
>Isn't OpenAI mostly built upon disregarding the copyright of countless people?
It sure was. But OpenAI decided to poke the Bear and is being sued by NYT. And apparently as a sidequest they thought it best to put their head in a lion's mouth. I wouldn't call the PR clout and finances of an A-list celebrity small potators.
They could have easily flown under the radar and have been praised as the next Google if they kept to petty thievery on the internet instead of going for the high profile content.
>People are still going to use ChatGPT to cheat on their homework, to phone-in their jobs, and to try to ride OpenAI's coattails.
Sure, and ChatGPT isn't goint to make lots of money from these small time users. They want to target corporate, and nothing scares of coporate more than pending litigation. So I think this will bite them sooner rathter than later.
>Maybe OpenAI will eventually settle with the actress, for a handful of coins they found in the cushions of their trillion-dollar sofa.
I suppose we'll see. I'm sure she was offered a few pennies as is, and she rejected that. She may not be in it for the money. She very likely doesn't need to work another day in her life as is.
> > Isn't OpenAI mostly built upon disregarding the copyright of countless people?
> It sure was.
Can you cite something that elaborates on this point? Do people who read books and then learn from it also disregard copyright? How is what OpenAI does meaningfully different from what people do?
> No one is going to push back against OpenAI meaningfully.
Couldn't, perhaps, one of the more famous people on Earth be responsible for "meaningfully" taking OpenAI to task for this? Perhaps even being the impetus for legislative action?
From the Ars Technica story[1], this is very funny:
> But OpenAI's chief technology officer, Mira Murati, has said that GPT-4o's voice modes were less inspired by Her than by studying the "really natural, rich, and interactive" aspects of human conversation, The Wall Street Journal reported.
People made fun of Murati when she froze after being asked what Sora was trained on. But behavior like that indicates understanding that you could get the company sued if you said something incriminating. Altman just tweets through it.
Bloomberg's Odd Lots Podcast had an ex-CIA officer, Phil Houston, on in April of 2024. He was promoting a new book but he had a lot of great advice for anyone to use regarding 'tells' when people are lying. Murati was clearly lying- that's obvious then and now.
Is there actually any evidence for this? AFAIK, other similar claims about people doing certain things when lying have been debunked(like fidgeting, avoiding eye contact, etc)
It probably made things worse, but the fact that they reached out to her to use her voice and she explicitly refused would be sufficient ammo I feel. (Not a lawyer of course).
Of course, Twitter continues to bring people with big egos to their own downfall.
Not to mention that it matches up pretty much exactly with the Bette Midler and Tom Waits cases, where courts ruled against companies using soundalikes after the person they really wanted turned them down. Doesn't matter if they hired a soundalike actress rather than clone her voice, it still violates her rights.
Definitely. GPT4o has a voice that sounds like Scarlett Johannson? They'd probably get away with it, I'm sure there are a lot of people that sound like her. Tweeting a reference to a movie she was in - a bit more murky because it's starting to sound like they are deliberately cloning her voice. Asking to use her voice, then using a soundalike, then referencing the movie? 100%, no doubt.
They’ll make the case that the abilities of the device is what he was referring to, but I think more the fact they were pushing her so hard for her involvement will actually be the damning aspect for them with that line of defense.
I found the whole ChatGPT-4o demo to be cringe inducing. The fact that Altman was explicitly, and desperately, trying to copy "her" at least makes it understandable why he didn't veto the bimbo persona - it's actually what he wanted. Great call by Scarlett Johansson in not wanting to be any part of it.
One thing these trained voices make clear is that it's a tts engine generating ChatGPT-4o's speech, same as before. The whole omni-modal spin suggesting that the model is natively consuming and generating speech appears to be bunk.
> One thing these trained voices make clear is that it's a tts engine generating ChatGPT-4o's speech, same as before.
I'm not familiar with the specifics of how AI models work but doesn't the ability from some of the demos rule out what you've said above? Eg. The speeding up and slowing down speech and the sarcasm don't seem possible if TTS was a separate component
The older formant-based (vs speech sample based) speech sythesizers like DECTalk could do this too. You could select one of a half dozen voices (some male, some female), but also select the speed, word pronunciation/intonation, get it to sing, etc, because these are all just parameters feeding into the synthesizer.
It would be interesting to hear the details, but what OpenAI seem to have done is build a neural net based speech synthesizer which is similarly flexible because it it generating the audio itself (not stitching together samples) conditioned on the voice ("Sky", etc) it is meant to be mimicking. Dialing the emotion up/down is basically affecting the prosody and intonation. The singing is mostly extending vowel sounds and adding vibrato, but it'd be interesting to hear the details. In the demo Brockman refers to the "singing voice", so not clear if they can make any of the 5 (now 4!) voices sing.
In any case, it seems the audio is being generated by some such flexible tts, not just decoded from audio tokens generated by the model (which anyways would imply there was something - basically a tts - converting text tokens to audio tokens). They also used the same 5 voices in the previous ChatGPT which wasn't claiming to be omnimodal, so maybe basically the same tts being used.
I have no special insight into what they're actually doing, but speeding up and slowing down speech have been features of SSML for a long time. If they are generating a similar markup language it's not inconceivable that it would be possible to do what you're describing.
I wouldn’t go as far as your last statement. While shocking, it’s not inconceivable that there’s native token I/O for audio. In fact tokenizing audio directly actually seems more efficient since the tokenization could be local.
Nevertheless. This is still incredibly embarrassing for OpenAI. And totally hurts the company’s aspiration to be good for humanity.
I think it is more then a simple tts engine. At least from the demo, they showed: It can control the speed and it can sing when requested. Maybe its still a seperate speech engine, but more closely connected to the llm.
>One thing these trained voices make clear is that it's a tts engine generating ChatGPT-4o's speech, same as before. The whole omni-modal spin suggesting that the model is natively consuming and generating speech appears to be bunk.
This doesn't make any sense. If it's a speech to speech transformer then 'training' could just be a sample at the beginning of the context window. Or it could one of several voices used for the Instruct-tuning or RLHF process. Either way, it doesn't debunk anything.
I often wonder why tech people think so positively about companies they idolise who are Uber-ing their way through regulations. Where do they think it stops?
Why would people not want laws? The answer is so they can do the things that the laws prevent.
This is POSIWID territory [0]. "The purpose of a system is what it does". Not what it repeatedly fails to live up to.
What was the primary investment purpose of Uber? Not any of the things it will forever fail to turn a profit at. It was to destroy regulations preventing companies like Uber doing what they do. That is what it succeeded at.
The purpose of OpenAI is to minimise and denigrate the idea of individual human contributions.
> I often wonder why tech people think so positively about companies they idolise who are Uber-ing their way through regulations. Where do they think it stops?
Because they don't think about the consequences, and don't want to. Better to retreat into the emotional safety of techno-fantasy and feeling like you're on the cutting edge of something big and new (and might make some good money in the process). Same reason people got into NFTs.
You need to be honest about what it actually does then. Cherry picking the thing you don't like and ignoring the rest will bring you no closer to true understanding
Readit News
- OpenAI approached Scarlett last fall, and she refused.
- Two days before the GPT-4o launch, they contacted her agent and asked that she reconsider. (Two days! This means they already had everything they needed to ship the product with Scarlett’s cloned voice.)
- Not receiving a response, OpenAI demos the product anyway, with Sam tweeting “her” in reference to Scarlett’s film.
- When Scarlett’s counsel asked for an explanation of how the “Sky” voice was created, OpenAI yanked the voice from their product line.
Perhaps Sam’s next tweet should read “red-handed”.
The really concerning part here is that Altman is, and wants to be, a large part of AI regulation [0]. Quite the public contradiction.
[0] https://www.businessinsider.com/sam-altman-openai-artificial...
It’s not like Tom Waits ever wanted to hock chips
https://www.latimes.com/archives/la-xpm-1990-05-09-me-238-st...
Maybe I liked it best because it felt familiar, even if I didn’t know why. I’m a bit disappointed now that she didn’t sign on officially, but my guess is that Altman just burned his bridge to half of Hollywood if he is looking for a plan B.
Dead Comment
If they really hired someone who sounds just like her it's fair game IMO. Johanssen can't own the right to a similar voice just like many people can have the same name. I think if there really was another actress and she just happens to sound like her, then it's really ok. And no I'm not a fan of Altman (especially his worldcoin which I view as a privacy disaster)
I mean, imagine if I happened to have a similar voice to a famous actor, would that mean that I couldn't work as a voice actor without getting their OK just because they happen to be more famous? That would be ridiculous. Pretending to be them would be wrong, yes.
If they hired someone to change their voice to match hers, that'd be bad. Yeah. If they actually just AI-cloned her voice that's totally not OK. Also any references to the movies. Bad.
No one is harmed.
Deleted Comment
OpenAI did nothing wrong.
The movie industry does the same thing all the time. If an actor/actress says no they you find someone else who can play the same role.
To me it's about as close to her voice as saying "It's a woman's voice". Not to say all women sound alike but the sound I heard from that video above could maybe best be described and "generic peppy female American spokesperson voice"
Even listening to it now with the suggestion that it might sound like her I don't personally hear Scarlett Johansson's voice from the demo.
There may be some damming proof where they find they sampled her specifically but saying they negotiated and didn't come to an agreement is not proof that it's supposed to be her voice. Again, to me it just sounds like a generic voice. I've used the the version before GPT-4o and I never got the vibe it was Scarlett Johansson.
I did get the "Her" vibe but only because I was talking to a computer with a female voice and it was easy to imagine that something like "Her" was in the near future. I also imagined or wished that it was Majel Barrett from ST:TNG, if only because the computer on ST:TNG gave short and useful answers where as ChatGPT always gives long-winded repetitive annoying answers
Bon mots apart, he really appears to have an innate capacity for betrayal.
"Midler was asked to sing a famous song of hers for the commercial and refused. Subsequently, the company hired a voice-impersonator of Midler and carried on with using the song for the commercial, since it had been approved by the copyright-holder. Midler's image and likeness were not used in the commercial but many claimed the voice used sounded impeccably like Midler's."
As a casual mostly observer of AI, even I was aware of this precedent
New voice mode is a speech predicting transformer. "Voice Cloning" could be as simple as appending a sample of the voice to the context and instructing it to imitate it.
OpenAI first demoed and launched the “Sky” voice in November last year. The new demo doesn’t appear to have a new voice.
I doubt it would take them long to prepare a new voice, and who’s to say they wouldn’t delay the announcements for a ScarJo voice?
A charitable interpretation of the “her” tweet would be a comparison to the conversational and AI capabilities of the product, not the voice specifically, but it’s certainly not a good look.
How much do you think Disney or Universal Music or Google or NYT would give to peek inside OpenAI's training mixture to identify all the infringing content?
Otherwise, it'd be impossible to show damages if you weren't personally being denied business because of the association.
…Unless, of course, all this scandal isn't also a part of marketing campaign.
Apparently they had no confidence in defending themselves, so why even release with the voice in the first place?
Are you suggesting they should have engineered the voice actress' voice to be more distinct from another actress they were considering for the part? Or just not gone near it with a 10ft pole? because if the latter the studios can just release a new Her and Him movie with different voices in different geo regions and prevent anyone from having any kind of familiar engaging voice bot.
Seems pretty reckless to not have alternatives just in case Scarlett refused.
Still live for me? Unless the Sky I’m getting is a different one?
Deleted Comment
Don't hear any arguments on how this is fair-use. (It isn't)
Why? Because everyone (including OpenAI) knows it clearly isn't fair-use even after pulling the voice.
> Why?
Because it's a right of personality issue, not copyright, and there is no fair use exception (the definition of the tort is already limited to a subset of commercial use which makes the Constitutional limitation that fair use addresses in copyright not relevant, and there is no statutory fair use exception to liability under the right of personality.)
Deleted Comment
https://www.youtube.com/watch?v=fMtbrKhXMWc
Sam should be ashamed to have ever thought of ripping off anyone's voice, let alone done it and rolled it out.
They are building some potentially world-changing technology, but cannot rise above being basically creepy rip-off artists. Einstein was right about requiring improved ethics to meet new challenges, and also that we are not meeting that requirement.
sad to see
Of course it’s not. All of ChatGPT is a ripoff. That’s what training data (which they did not license) is.
Deleted Comment
The movie industry does this all the time.
Johansson is probably suing them so they're forced to remove the Sky voice while the lawsuit is happening.
I'm not a fan of Sam Altman or OpenAI but they didn't do anything wrong here.
But I’ve defended them from unfair criticism on more than a few occasions and I feel that of all the things to land on them about this one is a fairly mundane screwup that could be a scrappy PM pushing their mandate that got corrected quickly.
The leadership for the most part scares the shit out of me, and clearly a house-cleaning is in order.
But of all the things to take them to task over? There’s legitimately damning shit this week, this feels like someone exceeded their mandate from the mid-level and legal walked it back.
It sounds like Altman was personally involved in recruiting her. She said no and they took what they wanted anyway.
For example, a car company approached the band sigur ros to include some of their music in a car commercial. Sigur ros declined. A few months later the commercial airs with a song that sounds like an unreleased sigur ros song, but really they just paid a composer to make something that sounds like sigur ros, but isn't. So maybe openai just had a random lady with a voice similar to Scarlett so the recording.
Taking down the voice could just be concern for bad press, or trying to avoid lawsuits regardless of whether you think you are in the right or not. Per this* CNN article:
> Johansson said she hired legal counsel, and said OpenAI “reluctantly agreed” to take down the “Sky” voice after her counsel sent Altman two letters.
So, Johansson's lawyers probably said something like "I'll sue your pants off if you don't take it down". And then they took it down. You can't use that as evidence that they are guilty. It could just as easily be the case that they didn't want to go to court over this even if they thought they were legally above board.
* https://www.cnn.com/2024/05/20/tech/openai-pausing-flirty-ch...
There have been several legal cases where bands have sued advertisers for copying their distinct sound. Here are a few examples:
The Beatles vs. Nike (1987): The Beatles' company, Apple Corps, sued Nike and Capitol Records for using the song "Revolution" in a commercial without their permission. The case was settled out of court.
Tom Waits vs. Frito-Lay (1988): Tom Waits sued Frito-Lay for using a sound-alike in a commercial for their Doritos chips. Waits won the case, emphasizing the protection of his distinct voice and style.
Bette Midler vs. Ford Motor Company (1988): Although not a band, Bette Midler successfully sued Ford for using a sound-alike to imitate her voice in a commercial. The court ruled in her favor, recognizing the uniqueness of her voice.
The Black Keys vs. Pizza Hut and Home Depot (2012): The Black Keys sued both companies for using music in their advertisements that sounded remarkably similar to their songs. The cases were settled out of court.
Beastie Boys vs. Monster Energy (2014): The Beastie Boys sued Monster Energy for using their music in a promotional video without permission. The court awarded the band $1.7 million in damages.
As are disclaimers that celebrity voices are impersonated when there is additional context which makes it likely that the voice would be considered something other than a mere soundalike, like direct reference to a work in which the impersonated celebrity was involved as part of the same publicity campaign.
And liability for commercial voice appropriation, even by impersonation, is established law in some jurisdictions, including California.
There's an ocean of difference between mimicking the style of someone's art in an original work, and literally cloning someone's likeness for marketing/business reasons.
You can hire someone to make art in the style of Taylor Swift, that's OK.
You can't start selling Taylor Swift figurines by the same principle.
What Sam Altman did, figuratively, was giving out free T-Shirts featuring a face that is recognized as Taylor Swift by anyone who knows her.
So he's here to help regulate it all with an "international agency" (see the reference[2] by windexh8er in this thread)! Don't forget that Altman is the same hack who came up with "Worldcoin" and the so-called "Orb" that'll scan your eyeballs for "proof of personhood".
Is this sleazy marketer the one to be trusted to lead an effort that has a lasting impact on humanity? Hell no.
[1] https://news.ycombinator.com/item?id=38312294
[2] https://news.ycombinator.com/item?id=40423483
It’s so refreshing to hear someone else actually understand this sentiment for what it is— snake oil sales. The normies out there eat this up, and there is no convincing them otherwise because of how powerful the AI trope in entertainment media is.
Honestly though, if you actually listen to him and read his words he seems to be even more devoid of basic empathetic human traits than even Zuckerberg who gets widely lampooned as a robot or a lizard.
He is a grifter through-and-through.
Yes, I've listened to Altman. A most recent one is him waffling with a straight-face about "Platonic ideals"[1], while sitting on a royal chair in Cambridge. As I noted here[2] six months ago, if he had truly read and digested Plato's works, he simply will not be the ruthless conman he is. Plato would be turning in his grave.
[1] https://www.youtube.com/watch?v=NjpNG0CJRMM&t=3632s
[2] https://news.ycombinator.com/item?id=38312875
But it can be learned to be mimicked almost to perfection, either by endless trial & error or by highly intelligent motivated people. It usually breaks apart when completely new intense / stressful situation happens. Sociopaths belong here very firmly and form majority.
If you know what to look for, you will see it in most if not all politicians, 'captains of industry' or otherwise people who got to serious power by their own deeds.
Think about a bit - what sort of nasty battles they had to continually keep winning with similar folks to get where they are, this ain't the place for decent human beings, you/me would be outmatched quickly. Jordan Peterson once claimed you have cca 1/20 of sociopaths in general population, say 15 millions just in US? Not every one is highly intelligent and capable of getting far, but many do. Jobs, Gates, Zuckenberg, Bezos, Musk, Altman and so on and on. World is owned and run by them, I'd say without exception.
If so, I suspect they’ll be okay in a court of law — having a voice similar to a celebrity isn’t illegal.
It’ll likely cheese off actors and performers though.
[1] https://www.forbes.com/sites/roberthart/2024/05/20/openai-sa...
"when voice is sufficient indicia of a celebrity's identity, the right of publicity protects against its imitation for commercial purposes without the celebrity's consent."
There's not a lot of precedent around voice impersonation, but there is for a very, very similar case against Ford
https://en.m.wikipedia.org/wiki/Midler_v._Ford_Motor_Co.
OpenAI has gone the "it's easier to ask forgiveness than permission" route, and it seemed like they might get away with that, but if this results in a lot more stories like this they'll risk running afoul of public opinion and future legislation turning sharply against them.
Tech Company: At long last, we have created the Torment Nexus from classic sci-fi novel Don't Create The Torment Nexus
https://x.com/AlexBlechman/status/1457842724128833538?lang=e...
And not see how over the top it is... cmon.
A company can't take a photo from your Facebook and plaster it across an advertisement for their product without you giving them the rights to do that.
And if you're a known public figure, this includes lookalikes and soundalikes as well. You can't hire a ScarJo impersonator that people will think is ScarJo.
This is clearly a ScarJo soundalike. It doesn't matter whether it's an AI voice or clone or if they hired someone to sound just like her. Because she's a known public figure, that's illegal if she hasn't given them the rights.
(However, if you generate a synthetic voice that just happens to sound exactly like a random Joe Schmo, it's allowed because Joe Schmo isn't a public figure, so there's no value in the association.)
https://en.m.wikipedia.org/wiki/Midler_v._Ford_Motor_Co.
Bette Middler successfully sued Ford for impersonating her likeness in a commercial.
Then also:
https://casetext.com/case/waits-v-frito-lay-inc
Tom Waits successfully sued Frito Lay for using an imitator without approval in a radio commercial.
The key seems to be that if someone is famous and their voice is distinctly attributeable to them, there is a case. In both of these cases, the artists in question were also solicited first and refused.
> In a novel case of voice theft, a Los Angeles federal court jury Tuesday awarded gravel-throated recording artist Tom Waits $2.475 million in damages from Frito-Lay Inc. and its advertising agency.
> The U.S. District Court jury found that the corn chip giant unlawfully appropriated Waits’ distinctive voice, tarring his reputation by employing an impersonator to record a radio ad for a new brand of spicy Doritos corn chips.
https://www.latimes.com/archives/la-xpm-1990-05-09-me-238-st...
Plus the tone of the voice is likely an unimportant detail to theor success. So pushing up against the legal boundaries in this specific domain is at best strange and at worst a huge red flag for their ethics and how they operate.
https://en.wikipedia.org/wiki/Personality_rights
And hasn't OpenAI recently shown that they can pull off a commercial coup d'état, unscathed?
Why would they not simply also take the voice of some actress? That's small potatoes.
No one is going to push back against OpenAI meaningfully.
People are still going to use ChatGPT to cheat on their homework, to phone-in their jobs, and to try to ride OpenAI's coattails.
The current staff have already shown they're aligned with the coup.
Politicians and business leaders befriend money.
Maybe OpenAI will eventually settle with the actress, for a handful of coins they found in the cushions of their trillion-dollar sofa.
It sure was. But OpenAI decided to poke the Bear and is being sued by NYT. And apparently as a sidequest they thought it best to put their head in a lion's mouth. I wouldn't call the PR clout and finances of an A-list celebrity small potators.
They could have easily flown under the radar and have been praised as the next Google if they kept to petty thievery on the internet instead of going for the high profile content.
>People are still going to use ChatGPT to cheat on their homework, to phone-in their jobs, and to try to ride OpenAI's coattails.
Sure, and ChatGPT isn't goint to make lots of money from these small time users. They want to target corporate, and nothing scares of coporate more than pending litigation. So I think this will bite them sooner rathter than later.
>Maybe OpenAI will eventually settle with the actress, for a handful of coins they found in the cushions of their trillion-dollar sofa.
I suppose we'll see. I'm sure she was offered a few pennies as is, and she rejected that. She may not be in it for the money. She very likely doesn't need to work another day in her life as is.
> It sure was.
Can you cite something that elaborates on this point? Do people who read books and then learn from it also disregard copyright? How is what OpenAI does meaningfully different from what people do?
Couldn't, perhaps, one of the more famous people on Earth be responsible for "meaningfully" taking OpenAI to task for this? Perhaps even being the impetus for legislative action?
If they tell the story of OpenAI, in a way that reaches people, that would be a triumph of the real artists, over the dystopian robo-plagiarists.
I love it already.
> But OpenAI's chief technology officer, Mira Murati, has said that GPT-4o's voice modes were less inspired by Her than by studying the "really natural, rich, and interactive" aspects of human conversation, The Wall Street Journal reported.
People made fun of Murati when she froze after being asked what Sora was trained on. But behavior like that indicates understanding that you could get the company sued if you said something incriminating. Altman just tweets through it.
[1] https://arstechnica.com/tech-policy/2024/05/openai-pauses-ch...
https://podcasts.apple.com/us/podcast/an-ex-cia-officer-expl...
Dead Comment
Of course, Twitter continues to bring people with big egos to their own downfall.
One thing these trained voices make clear is that it's a tts engine generating ChatGPT-4o's speech, same as before. The whole omni-modal spin suggesting that the model is natively consuming and generating speech appears to be bunk.
I'm not familiar with the specifics of how AI models work but doesn't the ability from some of the demos rule out what you've said above? Eg. The speeding up and slowing down speech and the sarcasm don't seem possible if TTS was a separate component
It would be interesting to hear the details, but what OpenAI seem to have done is build a neural net based speech synthesizer which is similarly flexible because it it generating the audio itself (not stitching together samples) conditioned on the voice ("Sky", etc) it is meant to be mimicking. Dialing the emotion up/down is basically affecting the prosody and intonation. The singing is mostly extending vowel sounds and adding vibrato, but it'd be interesting to hear the details. In the demo Brockman refers to the "singing voice", so not clear if they can make any of the 5 (now 4!) voices sing.
In any case, it seems the audio is being generated by some such flexible tts, not just decoded from audio tokens generated by the model (which anyways would imply there was something - basically a tts - converting text tokens to audio tokens). They also used the same 5 voices in the previous ChatGPT which wasn't claiming to be omnimodal, so maybe basically the same tts being used.
Nevertheless. This is still incredibly embarrassing for OpenAI. And totally hurts the company’s aspiration to be good for humanity.
Seems like they abandoned it pretty early - if it was real in the first place.
Was quickly apparent that text only is a poor medium for the variety and scope of signals that could be communicated by these multimodal networks.
This doesn't make any sense. If it's a speech to speech transformer then 'training' could just be a sample at the beginning of the context window. Or it could one of several voices used for the Instruct-tuning or RLHF process. Either way, it doesn't debunk anything.
Altman and OpenAI will walk over everyone here without any difficulty if they decide to take whats ours.
Why would people not want laws? The answer is so they can do the things that the laws prevent.
This is POSIWID territory [0]. "The purpose of a system is what it does". Not what it repeatedly fails to live up to.
What was the primary investment purpose of Uber? Not any of the things it will forever fail to turn a profit at. It was to destroy regulations preventing companies like Uber doing what they do. That is what it succeeded at.
The purpose of OpenAI is to minimise and denigrate the idea of individual human contributions.
[0] https://en.wikipedia.org/wiki/The_purpose_of_a_system_is_wha...
Because they don't think about the consequences, and don't want to. Better to retreat into the emotional safety of techno-fantasy and feeling like you're on the cutting edge of something big and new (and might make some good money in the process). Same reason people got into NFTs.
You need to be honest about what it actually does then. Cherry picking the thing you don't like and ignoring the rest will bring you no closer to true understanding