Surely there’s some kind of difference between “voice impression for a two-line cameo in one episode of an animated sitcom” and “reproducing your voice as the primary interface for a machine that could be used by billions of people and is worth hundreds of billions of dollars.”
Is there a name for this AI fallacy? The one where programmers make an inductive leap like, for example, if a human can read one book to learn something, then it’s ok to scan millions of books into a computer system because it’s just another kind of learning.
If famous actors could sue over the use of a less-famous actor that sounds just like them, what's to stop less-famous actors from suing over the use of a famous actor who sounds just like them in big-budget movies? ... and that's when you discover that "unique voice" is a one-in-a-million thing and thousands of people have the same voice, all asking for their payout.
That's a common retort on HN but it's information free. Judges are at least theoretically and often in practice bound to follow both written law and logic, even if it yields apparently silly outcomes. The prevalence of openly political judgements in the news makes it seem like this isn't the case, but those judgements are newsworthy exactly because they are shocking and outrageous.
If voices being similar to each other is found to be grounds for a successful tort action then it'd establish a legal precedent, and it's very unlikely that precedent would be interpreted as "whoever the judge heard of first wins".
> it's very unlikely that precedent would be interpreted as "whoever the judge heard of first wins"
No, it's whoever's voice is famous. The voice per se isn't valuable, its fame is. Personality rights are precedented [1].
> voices being similar to each other is found to be grounds for a successful tort action then it'd establish a legal precedent
It's not about similarity. It's about property. Johansson developed her voice into a valuable asset. It's valuable because it's Scarlet Johansson's voice.
Tweeting Her explicitly tied it to Johansson, even if that wasn't the case up to that point.
>> It's valuable because it's Scarlet Johansson's voice.
> I think demonstrating that this is a substantial part of the attraction of OpenAI's tech will be difficult.
I think it's totally irrelevant if her voice "is a substantial part of the attraction of OpenAI's tech." What matters is they took something from her that was her valuable property (her likeness). It doesn't matter if what they took makes op 99% of the value or 0.00001%.
It does when it comes to this being a useful topic.
They didn't take her likeness; they recorded someone else. The only claim she has is that someone who sounds like her will add value to their product more than if the person didn't sound like her. At which point the question is: how much value?
(Even that isn't a claim in and of itself, of course, but it might be the basis for a "I'll make people not like you so pay me restitution from your marketing budget to avoid a court case" shakedown.)
> The timeline may not matter as much as OpenAI may think, though. In the 1990s, Tom Waits cited Midler's case when he won a $2.6 million lawsuit after Frito-Lay hired a Waits impersonator to perform a song that "echoed the rhyming word play" of a Waits song in a Doritos commercial. Waits won his suit even though Frito-Lay never attempted to hire the singer before casting the soundalike.
----
> The only claim she has is that someone who sounds like her will add value to their product more than if the person didn't sound like her. At which point the question is: how much value?
That may be relevant when damages are calculated, but I don't think that's relevant to the question of if OpenAI can impersonate her or not.
Yeah, but it's not Scarlett Johansson's voice and therefore not her property. It's one that sounds similar, but is different, and thus belongs to the true voice actress.
> it's not Scarlett Johansson's voice and therefore not her property
It's not her voice. But it may have been intended to sound like her voice. (I believe this less than twenty-four hours ago, but I'm hesitant to grant Altman the benefit of doubt.)
If it were her voice, would you agree that seems distasteful?
> one that sounds similar, but is different, and thus belongs to the true voice actress
They marketed it as her voice when Altman tweeted Her.
> They marketed it as her voice when Altman tweeted Her.
Even that is not open and shut. He tweeted one word. He certainly wanted an association between the product and the movie, but it is a much more specific assertion that that one word demonstrates an intent to associate the product's voice actress with the voice actress who portrayed the comparable product's voice actress in the movie.
Nobody stops anyone from suing, but the less-famous actor would have to make a plausible case that the big-budget movie intended to copy the voice of the less-famous actor.
> for example, if a human can read one book to learn something, then it’s ok to scan millions of books into a computer system because it’s just another kind of learning.
Since this comes up all the time, I ask: What exactly is the number of books a human can ingest before it becomes illegal?
This is a bit like someone saying they don't want cars traveling down the sidewalk because they're too big and heavy, and then having someone ask how big and heavy a person needs to get before it becomes illegal for them to travel down the sidewalk.
It misses the point, which is that cars aren't people. Arguments like "well a car uses friction to travel along the ground and fuel to create kinetic energy, just like humans do", aren't convincing to me. An algorithm is not a human, and we should stop pretending the same rules apply to each.
Is that a good example? People have been arguing in court about that exact thing for years, first due to Segway and then due to e-scooters and bikes. There's plenty of people who make arguments of the form "it's not a car or a bike so I'm allowed on the sidewalk", or make arguments about limited top speeds etc.
> first due to Segway and then due to e-scooters and bikes
Those aren't cars.
But you've identified that the closer something comes to a human in terms of speed and scale, the blurrier the lines become. In these terms I would argue that GPT-4 is far, far removed from a human.
Legally they're vehicles sometimes, and sometimes technically supposed to not drive on sidewalks. Perhaps that's Segway equivalent to fair use scientific researches on crawled web data.
Yes. It is pertinent not only to this particular instance (or instances, plural; AI copyright violations and scooters on sidewalks), but illustrates for example why treating corporations as "people" in freedom-of-speech law is misguided (and stupid, corrupt, and just fucking wrong). So it is a very good example.
It's easy to explain the difference between a person and a car in a way that's both specific and relevant to the rules.
If we're at an analogy to "cars aren't people", then it sounds like it doesn't matter how many books the AI reads, even one book would cause problems.
But if that's the case, why make the argument about how many books it reads?
Are you sure you're arguing the same thing as the ancestor post? Or do you merely agree with their conclusion but you're making an entirely different argument?
Then again, bicycles are neither people nor cars, and yet they make claim to both sidewalk and the road, even though they clearly are neither, and are a danger and a nuisance on both.
Plagiarism is not illegal, it's merely frowned upon, and only in specific fields. Everywhere else, it's called learning from masters and/or practicing your art.
"But what if a person was so thoroughly replaced with robot parts to be just like a computer" is just "if my grandma had wheels, she would be a truck, therefore it's not so easy to say that cars aren't allowed to drive inside the old folks home".
People and software are different things, and it makes total sense that there should be different rules for what they can and cannot do.
> Surely there’s some kind of difference between “voice impression for a two-line cameo in one episode of an animated sitcom” and “reproducing your voice as the primary interface for a machine that could be used by billions of people and is worth hundreds of billions of dollars.”
There are too many differences to understand what you're saying. Is the problem too much money is in the company doing it? Fox is also pretty wealthy.
I think the pertinent question is: does having it sound like Scarlett Johansenn mean they get to access billions of people? If not, then while she might get paid out a few million, it'll be from OpenAI's marketing budget and not because of actual value added.
How unique is a voice? I'm sure there's enough people out ther who sounds like Johansson. There's probably some argument for voice + personality + face + mannerisms, some gestalt that's more comparable to copying the likeness "person". But openAI is copying a fictional character played by Johansson, it's not her. Do actor/esses get to monopolize their depiction of fictional characters? Especially when it's not tied to physical represenation. What if OpenAI associate it with an avatar that looks nothing like her. I'm sure hollywood and/or actors union is figuring this out.
> “Do actor/esses get to monopolize their depiction of fictional characters? Especially when it's not tied to physical represenation.”
If Annapurna Pictures (the production company that owns the rights to “Her”) made a sequel where the voice AI is played by someone else than Johansson but sounded the same and was marketed as a direct continuation, I think there would be a lawsuit.
She didn’t write the script or develop the character, but I think there’s enough creative authorship in her voice portrayal that it would be risky for the production company.
But OpenAI isn't making a sequel to Her, which I feel like there would be prexisting legal text in contract about repraising role in event of franchise if johansson has leverage, or ability to cast close facsimile if studio has leverage. Right now Johansson has leverage in court of public opinion, not necessarily law. What if OpenAI used a cartoon cat avatar that sounded like "Her", what if they have one interaction that doesn't comport to "Her" personality from the movie, thereby indicating a different being. Is there some comprehensive method acting documentation outlining the full complexity of a fictional character. Seems like there aremany ways for openAI to make voice sound like her, but not embody "Her" but they'd rather strategically retreat out of optics. But IANAL, but I am interested in seeing how this will get resolved in court.
They aren't making a sequel, they are doing an unlicensed video game conversion.
Reading all these musings here about a possible "there is no bad publicity" approach, I'm starting to wonder if the plan for if Johansson signed up was achieving court drama publicity by getting sued by Annapurna Pictures. "Can a three-letter tweet be the base of a copyright case?"
It's entirely routine for actors and actresses to be replaced in follow up works. The industry is full of examples, but here's 1 off the top of my head:
In iron man 1, Rhodey is played by terrance howard. For iron man 2, he wanted too much money in the contract, so they replaced with with don cheadle.
Wouldn't it be a dumb world to live in if a single actor in the cast can halt the production of a new work via lawsuit because they own the character?
"How unique is X" is something we can start to get quantitative answers to with strong machine learning models, and for most things people care about, it seems like the answer is "not very unique at all".
It's not a fallacy. Behind the AI are 180M users inputting their own problems and giving their guidance. Those millions of books only teach language skills they are not memorized verbatim except rare instances of duplicated text in the training set. There is not enough space to store 10 trillion tokens in a model.
And if we wanted to replicate copyrighted text with a LLM, it would still be a bad idea, better to just find a copy online, faster and more precise, and usually free. We here are often posting paywalled articles in the comments, it's so easy to circumvent the paywalls we don't even blink twice at it.
Using LLMs to infringe is not even the intended purpose, and it only happens when the user makes a special effort to prompt the model with the first paragraph.
What I find offensive is restricting the circulation of ideas under the guise of copyright. In fact copyright should only protect expression not the underlying ideas and styles, those are free to learn, and AIs are just an extension of their human users.
Is there a name for this AI fallacy? The one where programmers make an inductive leap like, for example, if a human can read one book to learn something, then it’s ok to scan millions of books into a computer system because it’s just another kind of learning.