So let me get this straight. Scarlett Johansson now thinks she is the all embodying voice of every white girl/woman in America? This is seriously some stupid stuff, sounded nothing like her.
Part of the problem is that the entire model of OpenAI is based on re-using other people’s content and not attributing. It’s not entirely surprising that someone might be concerned that is happening here.
That doesn’t appear to be what’s happening here.
- Many people think
skysounds similar to Scarlett Johansson. People were writing as much from the day the voices were released. - She was apparently in contact with OpenAI about doing a voice for the models and rejected them.
- The lawsuit came only after the Spring Update event where the voice capabilities of
gpt-4owere demonstrated.
So, I don’t believe sky is so much the issue as is the voice of gpt-4o (which may be a different version of sky, hence its removal).
My thoughts on it personally are:
- I don’t think
skysounds like Scarlett Johansson. - I think the expressiveness of the
gpt-4ovoice resembles many of the intonations and patterns of Scarlett Johansson’s performance inHer. - I think it would be monumentally short-sighted to build a voice model on the voice of a world-famous actress without permission.
- I don’t think OpenAI (or their legal team) is monumentally short-sighted.
- When people have disputes which they can’t resolve between themselves, in a civilized society the proper course of action is to use an outside mediator—often the civil court system—to adjudicate their dispute.
So I, personally, see nothing wrong with Scarlett Johansson involving lawyers if she feels she’s been injured by OpenAI’s actions.
Given only the facts I currently have and my own lay-opiniin I don’t think it’s a case she would prevail on, but I also don’t see this ever seeing the inside of a courtroom.
I do wish that uninvolved parties wouldn’t get so worked up about it though.
I think unless OpenAI starts being more respectful of the content creation community (of which I believe we all belong to) it’s likely more of these issues will keep popping up.
Recent deals with Reddit and SO should help a bit, especially if they attribute in a way that pushes traffic to users that originally created the content.
IMHO, a law should be passed requiring GPTs to cite in an accurate way. If our content is being used to partially respond to answers, we should be able to know that.
Note that I am on the fence about fair use. I am not on the fence about attribution. It may be GPT has a right to train on copyrighted content, but I don’t think they have a right not to acknowledge its creators.
If OAI can get by with synthetic data, then perhaps they should just do that.
The current motto of training LLMs is using data without permission, and then asking for it afterwards.
It’s not too far-fetched to imagine somebody having a strange obsession to imitate dystopian movies in their upcoming presentations.
The fact that there was a last-minute email to ask for re-consideration before the presentation to me indicates that they indeed already had a voice model prepared using her voice. It’s also not too far-fetched to imagine that some effort was implemented last-minute to adjust the voice.
The truth is that unless there’s some audit we’ll never know.
Your comment is full of speculation. Assuming OpenAI uses data without permission is unfounded, and suggesting someone there has an obsession with dystopian movies is pure speculation. The last-minute email (if even real) could be a precautionary legal measure, not evidence of wrongdoing. Speculating about voice adjustments is just that—speculation.
In short, your theories seem far-fetched if anything. And no “audit” is needed nor required.
If you’d like for them to come out and admit it you’ll be waiting for a while.
I’m not saying it’s solid truth but the constant reports have indicated that yes. It’s quite common.
Also, if you want to try and put me down for “speculating”, please avoid doing it yourself
I’m on mobile right now so sorry for the quick response.
Data harvesting, as highlighted in your linked article, is a concern. However, the issue at hand is whether OpenAI used Johansson’s voice without permission (Edit: And since they´re not the same voice, I don´t see a case here), and the article does not provide evidence of this. To clarify, my comments were based on the need for concrete evidence rather than general speculation. I apologize if I was interpreted otherwise.
I don’t think it’s possible to ever know without an audit, which is the issue. It’s not like they explicitly said “Nope, didn’t do it”.
I’m not demanding an audit, by any means. So, really. All we’re left with is speculating. I just wanted to portray that OpenAI has used other people’s sometimes proprietary information before asking for consent before.
Well, we don’t need to speculate if “Sky” is Johansson because they are unequivocally two different voices. While there may be some similarities, they are not identical by any measure. “Eerily similar”, or however her legal representation chooses to phrase it, does not hold up, as similarities in voices are not uncommon.
Obviouslyy, Johansson has the right to pursue this matter in court, as any citizen does, to adjudicate the dispute.
‘Uninvolved parties’ that’s a good expression. So much of anger and divisiveness is when people who don’t have the full story, history, and understanding of the field discuss based on a couple sound bytes dropped by the media. This is another example. ‘Uninvolved parties’ tend to cloud a subject with a haze of noise which doesn’t resolve into a conclusion and doesn’t go anywhere.
On the other hand maybe the worlds population is getting slowly educated on different subject just by dukking it out.
I did like the Sky voice and hope they bring it back. I didn’t think it sounded like Johansson but I’m no expert. I liked the lack of vocal fry, uptalk, frisson, lilt, drama or drawl in the Sky voice.
I wonder if the voice actress who did Sky will never get a job again because she is too much like Scarlett.
I’m really disappointed that OpenAI removed Sky’s voice. I had grown attached to it, and now I don’t use the service because, honestly, the other voices just don’t resonate with me the same way.
I wonder if the voice actress(es) even sounds like Sky. 
It’s very possible that Sky is a mixture of multiple actresses (and even actors), much like generated pictures or generated text is a mixture of sources.
I think it’s perfectly reasonable for us to comment / speculate on this matter.
AI usage of content it doesn’t own is a very serious issue and the more dialogue and discussion around it the better.
We all have a stake in this - it is very much a community issue. AI is harvesting everyone’s comments and using it to further its training.
This is why the folks on SO and Reddit are so freaked out. Pretty sure Elon at twitter is doing this en masse.
What’s perhaps more interesting, is how much can be done by synthetic data and how much bootstrapping with human generated content is required?
I’ll be honest, I’m somewhat ambivalent about sourcing from copyrighted material - it feels like fair use. We all do it. What bugs me though is not giving credit where credit is due.
well… if they really asked scarlett to reconsider only 2 days before 4o announcement, i believe they already know that choosen voice became similar to “her” 
but i’m glad to know scarlett meet these requirements this much:
- A voice that feels timeless
- An approachable voice that inspires trust
- A warm, engaging, confidence-inspiring, charismatic voice with rich tone
- Natural and easy to listen to
Yes, but it’s not acceptable to attack people—even celebrities—or to accuse a corporation of committing crimes without evidence—particularly on their own developer forum—which is what some people are choosing to do.
If there are attacks and accusations then it’s your job to moderate. Although I don’t see anybody explicitly saying “OpenAI stole her voice without consent!!”,
It’s perfectly acceptable to speculate and even protest.
The ongoing news
of using proprietary information
without consent
is repetitive
Soapboxes were performed at the town center, not hidden in an alleyway or in an irrelevant location. Their PURPOSE is to invoke emotion, to be infuriating to a certain group of already deeply invested people & their dogs.
As a developer I’m invested into OpenAI and it’s technologies. Hearing all of these issues regarding using proprietary information without consent is a serious concern.
Personally. I’m tired of these petty issues, resolutions, and quite frankly: drama. The fact that OpenAI couldn’t just simply say “No, we did not use her voice AT ALL” would have been enough to satisfy. Instead there was a smokescreen of “here’s how we generally created our voice actors”
All of these complaints can be boiled down to some plea for transparency.
When I saw the official statement of the actors group backing Johannson I realized that this may be about the fuzzy term ‘eerily similar’.
It implies that if a AI generated voice does sound ‘eerily’ like the one of a actor then this is already problematic.
Let’s hope this gets sorted out. I am afraid such a standard could be a serious issue for developers.
Edit: news link
The issue here isn’t exactly that “it sounds like her”. As @qrdl has mentioned, and as anyone knows who has done voice cloning: it’s very simple to mix voices and adjust accents without much work.
It’s not outside the realm of possibilities that they originally intended for her to be Sky, but also planned a secondary actor as a back-up.
The issue is the connection of the dots. Asking her for permission not ONCE but (apparently) multiple times, extremely close to the release date. The fact that “Sky” was used for the presentation and their CEO left a single “her” comment.
The uncanny similarities only emphasize that something is there that connects the dots together.
OpenAI is dramatically changing the landscape of the internet and availability of information. I think it’s completely fair that an entity with such power and capability should be held to the HIGHEST standards of transparency.
Or, at the very least, try to avoid drama by posting purposefully cryptic messages on social media platforms.
I can understand why Scarlett Johansson is asking questions.
It’s also clear that OpenAI is not going to give away the secret sauce for the sake of transparency.
Since taking content from the web to train language models was Ok yesterday but definitely not today anymore I am sure we will see lots more of undefined risks coming towards developers. Not just those that build with OpenAI.