Actors, Artists, Models, Minors
AI Is A Hassle Lately
Two stories we are working on for a future post have been leaking out this weekend. Apparently, the folks at Twitter (now called X) spent just four months training a user AI tool. In X’s drive to be the “everything app” some early sneak peaks claim the tool will enable AI powered searching of posts, creation of well written posts, etc.
OpenAI is poised to release tomorrow a feature enabling each subscriber to create apps on top of their LLM tech with OpenAI taking 30% of the subscription fees for those apps. As many have posted on social media already, “That’s the sound of 1000 startups dying.” We shall see. But, on to today’s collection of AI being a hassle.
You Cannot Sound Like Me
Lisa AI is a mobile app that enables users to, among other things, create 90s Yearbook style images and “AI powered art.” In its push to market its product, it hired a company called Convert Software. Convert created a 22 second ad for Lisa AI. The ad used the voice of Scarlett Johansson taken from her role in the movie Black Widow and then cuts to a scene with a voice over where her voice is heard saying something positive about Lisa AI and encouraging with “I think you shouldn’t miss it.” The problem is, that second piece of audio sounding like the actress is actually an AI voice model tuned to sound like her. Her legal team apparently reached out to both entities who haven’t yet commented. But, what seems like an invasion of talent or privacy or some kind of legal concept we have not yet codified, might not be that at all.
In the 1970s through the 1990s, late night shows regularly filled their invited guest lists with impressionists.
No legal action was ever taken when people would learn to make their voice sound like a famous actor, politician, etc. In fact, many performers made entire careers in Las Vegas doing impressions of other people’s voices, mannerisms and facial expressions.
In the example of Ms. Johannson above, what if a voice over actress wanted work in that field and she just happened to have a natural voice that sounded just like Ms. Johannson? Is she prohibited from voice over work or narration or being the voice of an animated character? I doubt that person could be stopped from working in whatever field merely because her voice sounds like that of someone famous. In comes AI. A company tunes a voice model that sounds just like someone famous. I can see some hair splitting here.
If a given endorsement commercial actually claimed to involve a famous person without their consent, that seems a fairly clear cut case that would be prohibited. But, what about commercials, voice-overs, etc. that simply include the AI tuned voice model that sounds exactly like a famous voice? No claimed endorsement, but such a distinctive voice that the casual listener assumes that actor or actress X is the one endorsing that product. This is a much harder case to make it seems. For courts to weigh into such uses of AI voice clones would involve questions like “did it really sound exactly like [famous person]?” or “what if we tune the model to be just a slightly lower voice, etc?” It’s a struggle to see how the law could effectively regulate AI tools sounding just like a recognizable voice.
Artists Making Poison
As we have written about in previous posts, there are currently multiple class action lawsuits pending against OpenAI, Microsoft and others by authors and visual artists. Their claim has been that the use of their artwork to create various LLMs (Large Language Model) infringes their copyrights. In a recent case brought by an artist against the makers of AI art tool Stable Diffusion, the court has given the artist an opportunity to amend her complaint. The court has advised that the artist has to “plausibly plead” that the defendants’ product enables the creation of AI art by “expressly referencing” the artists’ work by name. The court has also asked that the complaint allege “how much of [the artist’s] protected content remains in Stable Diffusion or is used by the AI end-products.” This presents dim prospects for this artist and her complaint surviving a motion to dismiss. Essays by many legal experts have echoed what we have written here when these cases were first filed. Merely showing that an AI tool can produce work “in the style of” a particular artist does not change the nature of how art has always worked. Meaning, artists have always trained on the styles of previous artists and many have even openly sought to mimic those previous artists. An AI tool that does what no court has ever prevented humans from doing, that is, duplicating the style of a previous artists, does not appear to be violating any existing copyright regulations.
Seeing these lawsuits, developers have started to take a different approach. An open source tool called Glaze is now available for free Windows or MacOS download. A short video about the tool does is here. Glaze purports to enable artists to display their art but if an AI tool scrapes it for use in training an LLM, the content the AI tool sees is blurred pixels.
This type of data poisoning as it has been termed is of increasingly interest to artists and those developing and improving LLMs. Although appreciating the desire of artists to not have their styles so easily duplicated, the notion of embedding anything termed “poison” inside of an image file harkens back to email attachment viruses that were often embedded in images with provocative names.
Models and AI, No, Not AI Models
A fashion model walked the runway for a well know fashion designer. Some time later, she noticed her image, or mostly her image, appearing in that designer’s Instagram feed. However, curiously, she noticed that the image of her at that runway show was not quite accurate.
The upload by the model quickly went viral on TikTok. The designer was not merely an up and comer trying to make their name. This designer has worked with the most well known names in fashion in culture like Beyonce and Jennifer Lopez. The model, Shereen Wu, is an independent model who said she was not paid for her work in the show. She accepted the job because she was counting on the exposure from her participation in the event to advance her career. However, the modified image essentially removed her identity from her participation in event. The designer removed the image once it was made known that it had been altered. The model and designer both theorized that an AI tool was used to replace the models face with the face as it appeared in the modified version. The controversy caused some in the industry to worry about a future where advertisers lock in on what “beauty” looks like and modify existing models’ images with AI to match that standard.
As a potentially ominous development for the future of modeling, companies are beginning to consider implementing completely AI models in place of human models. One example is the clothing company Levi’s. They announced earlier this year that they intended to use AI to augment the appearance of their clothing items for their online store. Their justification for the use of AI was to increase diversity and inclusion. AI, they reported, would enable them to display their products across a wider array of model sizes, ethnicities, etc. Which, of course, brought the obvious question to Levi’s “why not just hire human models across all those various features or characteristics you are using AI to reproduce?” The Levi’s response, if they were honest, would be one simple word - money.
With my ChatGPT Plus subscription I was able to create three AI models in three different ethnicities wearing essentially the same shirt and jeans in less than 2 minutes. Models used to doing commercial work for product labels, magazines or even online products have a right to worry about the future of AI and their livelihood.
Parents, High Schoolers and AI Image Manipulation
A New Jersey high school recently had to tackle something with no precedent in the world of student discipline. It started with teachers and fellow students noticing a small group of boys acting secretive and out of the ordinary with their phones. What was eventually discovered is that some person, perhaps a student, perhaps not, was uploading images of female students to an online AI powered tool which removed the clothes from the images. The resulting image was an undetectably altered image of a real female student modified to make it seem as if that student had posed nude for an image. Complicating the investigation and resolution of the issue was the students’ use of SnapChat. SnapChat automatically deletes shared content after set duration of minutes or hours. By the time authorities at the school and local law enforcement were made aware of the issue, most, if not all, of the shared images were unavailable for viewing to even confirm the events had occurred. Secondly, many of the female students were merely told that their images in this altered form were shared, but never saw them. This then presented the obviously distressing outcome for those students of not knowing what other students saw and how it might affect their reputation.
For adults, this kind of circumstance might be treated less seriously. Even states like California with revenge porn laws don’t have coverage for this situation. This is more like revenge fake porn if anything similar to what the revenge porn laws were designed to regulate. For impressionable and immature high school students, this kind of reputational damage could be sufficient to sideline their lives for months or years.
There is no solution to this problem, just trade offs. There is no way to stop open source tools that will enable the editing of images. And, editing of images means editing pixels. Every digital image is merely a matrix of pixels. There is no way to enable image editing, especially with AI, and somehow lock out the ability of these tools to do what these clothes-removing tools accomplish. And, surprisingly, this feature is not merely relegated to some obscure software. No, there are multiple, as in more than 10 at this point, such web applications that can be used for this purpose.
Conclusion
We stand at the collision of technological breakthroughs and societal norms. A collision that will continue to happen. This reality makes it imperative for us to navigate the digital terrain with caution and conscience. Regulation, education, and ethical AI development are important, but favoring one over the other in all circumstances seems to be an imprecise approach. As with any legal issue, if any of the pending class action lawsuits resolve with conflicting legal conclusions, the Supreme Court will undoubtedly weigh in. The future of AI and all it claims to be able to provide may well depend on how those matters are resolved.