CYBERSPACE — David Mack, a researcher active in artificial intelligence (AI) and machine learning with a game development background, recently published an essay on Medium about his failed attempt to replace flesh-and-blood porn stars with synthetic copies.
The piece, called "What I Learned from Building an AI That Generates Porn," is a fascinating glimpse at the actual motivations and ethical shortcuts of certain so-called "Tech Overlords" who become fixated with "disrupting" the adult industry.
“For a long time,” Mack confesses at the beginning of his peculiar essay, “there’d been a project on my ideas-list that I’d been ignoring: create pornography using AI. And so I delved into a brave new world.”
The project, initially, was a joke between friends but making porn with AI remained a persistent idea that never fully faded away.
“Pornography, as content,” Mack states, “has properties that make it amenable to being synthesized," mainly the fact that he believes consumers of adult content — unlike mainstream movies and television fans — do not care about any specific human performer.
Mack would probably be surprised to learn that an enormous part of the porn industry, from the star system to camming and custom clips and merch, is based on a very intimate connection between fans and their favorite performers.
For Mack, porn consumption is “somewhat random and transactional.” Unfamiliar faces, lack of plot, disconnected images (e.g., the old Tumblr porn) and low production values are not deterrents to Mack’s imaginary porn watcher.
Stealing Porn, for Science
For his AI experiment, Mack chose the Generative Adversarial Network (GAN) model, previously used to generate high-resolution what he described as “breathtakingly realistic" pictures of imaginary celebrity faces.
By feeding the GAN random images of real people, the network learns to generate images of imaginary people. The CelebA-HQ dataset, for example, generated vaguely distinctive faces of synthetic people who only bear a fleeting resemblance to Liam Hemsworth, Jon Hamm and Michael Douglas.
To train a GAN, Mack explained, “you feed it images of the things you would be like to generate. It simultaneously tries to generate images — the 'Generative' part of the name — and learns how to distinguish its own generated images from the real ones, known as a discriminator. These two networks, the generator and discriminator are pitted against each other for the Adversarial part of the name.
"As one gets better, so does the other. In the end, the generator can be used by itself to generate as many images as you wish for.”
To train a GAN, you typically need between 20,000 and 1,000,000 images, obtainable from popular image banks like ImageNet and corporate “data lakes," which hold data in raw form, unprocessed.
However, Mack was looking for something more niche. “To collect a dataset for this project,” Mack said, “I needed to get a little more creative. I needed to get tens, or better hundreds of thousands of pornography images.”
Mack admits that he knowingly stole the images for his experiment. “One way people collect datasets is to write web scrapers: little robots that trundle across the web collecting information from websites. I considered this, but it can take quite some time to get a web-scraper working really well. I also feared that pornography websites might have defenses against web scrapers," he said.
“I started hunting for BitTorrent pornography archives. Whilst many only had tens or hundreds of images, I finally found one with 200,000. It contained photographs of women, almost entirely alone, in around twenty different poses. These photos were of unknown copyright and ethical origins; however, I deferred solving those problems until I could prove this approach had any merit at all.”
Mack used Google Cloud to host his experiment and used a high memory 10-core machine with 8 Nvidia Tesla GPUs attached to it. In three days, he had the ability to generate 128x128 pixel images.
Following a few aborted attempts, “the network got to convergence” after three days of work, and Mack awaited anxiously for what he half-jokingly calls “the future of pornography.”
But the first experiment, as it often happens, was less than successful.
A Nightmare of Limbs and Vaginoids
The resulting images that Mack’s GAN generated were less the future of smut, and more a nightmare of limbs and mutated features and vaginoid protrusions resembling tiny Salvador Dalí paintings that had been left in the rain — or the failed insectoid hybrids in David Cronenberg’s “The Fly.”
“The network struggled with getting arms, legs, hands and faces in the right places,” admits Mack. “It had successfully learned their local textures, but not how to connect them all together. At this point, I turned to the academic literature to find out what I had done wrong.”
Mack figured out that the “imaginary mainstream celebrity faces” experiment that inspired him had lined up eyes, noses and mouths. GANs didn't have to learn how to coordinate the facial features — they could always place them in the same location. Put simply, celebrity face generation relied upon having a fairly homogenous dataset.
The images also had to be properly labeled, and to simplify things, Mack came to the revelation — by studying his stolen porn pictures — that a celebrity face has a lot in common with a tight vagina shot.
“It was clear that I needed to make my dataset more homogeneous,” he writes. “From a visual scan of the archive and observing the model’s generated outputs, I noticed one type of image it excelled on: a close-up of a vagina. I reasoned that this was because that shot has a regular structure — for example, where the legs are, where the tummy begins — just like the celebrity faces dataset.”
“I need to build a data subset with just that one pose,” he told himself.
Mack went back to his precious “200,000 pornography images” and he got to work, laboriously locating a loot of labia.
“At this point, I had to roll up my sleeves and do some hand-labelling of the pornography images. This became moderately unpleasant, and [led me to] question to whether I was okay with this project. Manual labelling of the images was slow work, therefore I wanted to minimize the number of labels required. I spent an hour, categorizing images as 'close-up vagina pose (label-1)' or 'other pose (label-0)' then trained the aforementioned model on it. I withheld some of my labels to test the model. In the end, I had 300 label-1 and 962 label-0 images. Training was fast and accurate, achieving around 95 percent accuracy.”
After a day, the vagina-only subset started generating synthetic “beaver shots” that were much closer to the real thing than his previous surrealist hellscapes.
This Vagina Does Not Exist
But Mack still had work to do. “Analyzing the misformed images,” he writes, “some patterns emerged. They often contained fingers. And usually way too many fingers, and hands that didn’t connect to them.”
Still thinking about these "fingervaginathings," which reminded him of the "Alien" film franchise, Mack took a break and “headed out to the mountains with my friends for some exercise and relaxation. I told them all of my project and where it’d got to. I showed them some results. A friend giggled and said ‘Have you seen ThisPersonDoesNotExist.com?.’ I smiled, knowing what they were implying, and checked available domain names.”
By the next evening, Mack was the proud owner and designer of ThisVaginaDoesnNotExist.com, a yonic riff on a popular synthetic face generator website.
The last part of Mack’s article chronicles his disappointment at the lack of support for his research from potential financiers. But then the essay takes an even more bizarre turn, when he says that the lack of funding made him rethink some ethical implications of his experiment.
If you thought he meant that he finally considered the ethics of stealing the likenesses of hundreds or thousands of actual women from dodgy torrent sites to put them out of a job, you'd be wrong.
The thing is, like many of his colleagues in the tech sector, Mack holds severely regressive views about porn. Whatever “ethical issues” he has are informed by the rampant misinformation spread by religiously motivated lobbies waging the “War on Porn.”
The last part of the essay is a sad reminder that many “tech innovators” who attempt to enter the porn industry to eliminate the need for human talent come to it with a superior, "tech-daddy-knows-best" worldview and have no interest in looking at the reality of the complex business they are trying to disrupt.
“At the macro view,” says Mack, from his imperious height, “if this project were a success, it could improve the world. Many people are harmed in the production of pornography, and this project, given the very low cost of producing new images, would supplant them. Pornography would harm fewer people.”
“On the other side,” he continued, “this would displace workers and reshape an industry. I believe that these people would find new work, and also that the transition could be difficult. The industry has faced many challenges over the past decade from digital technology and piracy, and it’s already become a difficult place to make money or a career in.”
He goes on with his theories, which come right out of the #NoFap nutjobs and subreddits about "preserving one's manly essence." The consumption of porn, Mack decided may be "harmful."
"Researching online. it was hard to find a clear, unbiased view on this," he writes. "Uncertain of the situation here, I decided that producing non-violent porn of a consensual nature, and including messages encouraging positive sexual attitudes, would be enough to break even on this ethical concern. At least until more research can be found. Finally, there is the issue of the training images — they have come from real actors, who may be experiencing harm. To mitigate this, my plan was to seek out pornography that had been ethically sourced.”
That’s right, Mack has zero concern about the intellectual property of production companies and the many actual women whose vaginas he spent vast amounts of time labeling. He just wants to “ethically source” his images, whatever that means (“Grain-fed, cage-free vulvas”?).
Fortunately, Mack reveals that he has moved on to greener (i.e. potentially more lucrative) pastures. “Checking in with myself, this no longer felt like an industry and line of work I wanted to be in," he writes. "Working with the content felt unpleasant. The various ethical issues and inevitable judgement and stigma didn’t feel worth fighting. Overall, this had been a fascinating dive into an entirely different area of the world from what I knew. I learned a lot, about technology, society and myself.”
When you're an aspiring Tech Lord, the whole world is your disruptable oyster.
To read David Mack's "What I Learned from Building an AI That Generates Porn," visit Medium.
