How Academic AI Research Morphed Into Computer-Generated Porn

General Intelligence

These dangerous technologies are still in their infancy, and they’ll continue to become more accurate and convincing

Dave Gershgorn

Dave Gershgorn

3 hours ago·4 min read

Image for post

Imagined by a GAN (generative adversarial network) StyleGAN2 (Dec 2019). Image: thispersondoesnotexist.com/Karras et al. and Nvidia

OneZero’s General Intelligence is a roundup of the most important artificial intelligence and facial recognition news of the week.

ess than 10 years ago, some of the most basic artificial intelligence algorithms, like image recognition, required the sort of computing power typically found in data centers. Today, those tools are available on your smartphone, and are far more powerful and precise.

Like nuclear power or rocket propulsion, artificial intelligence is considered a “dual-use” technology, which means that its capacity for harm is equal to its potential for good.

Earlier this week Vice reported the latest example of one of these harms: Coders were using images of sexual abuse to train algorithms to make porn. The article details how nonconsensual images were compiled by an anonymous PhD student into a dataset and combined with off-the-shelf algorithms to generate custom videos.

The creator of the A.I.-generated porn, who posted it on platforms like PornHub and OnlyFans, told Vice that he used StyleGAN2, an open-source algorithm built by Nvidia. If you’ve seen highly realistic fake faces online, like ThisPersonDoesNotExist.com, they’ve likely been generated by StyleGAN2.

Like nuclear power or rocket propulsion, artificial intelligence is considered a “dual-use” technology, which means that its capacity for harm is equal to its potential for good.

But this technology didn’t show up overnight. There’s a clear path from some of the earliest modern image-generating algorithms to this phenomenon of A.I.-generated porn. Here’s what it looks like.

The leap to GANs

Image generation algorithms leapt forward in capability in 2014, with the creation of generative adversarial networks, or GANs. The idea, which A.I. researcher Ian Goodfellow originally thought up during an argument at a bar, was to pit algorithms against each other to generate the best outcome. To generate an image, you would have a “generator” and a “discriminator.” The generator would make images, and the discriminator would try to determine if it was real or fake, based on real images it had been trained on. Only the most realistic images would be accepted by the discriminator, ensuring that the final result was the cream of the A.I.-generated crop.

Making the technology useful

Goodfellow’s initial research on GANs performed well on industry benchmarks, but many of the images he created still looked like hellish blobs that only represented ideas in abstract and inhuman ways. By 2016, other researchers had started experimenting with the technique and found ways to make lifelike images, albeit at small resolutions. One of the standout papers of the time showed how researchers could generate realistic images of bedrooms, as well as rudimentary attempts at generating faces. This research again showed that GANs were able to adapt based on the kind of data they were trained on. The idea worked as well for faces as it did for bedrooms, meaning the networks were actually able to identify patterns in a variety of different types of images.

Image for post

Hell blobs. Credit: Goodfellow et al, 2016

Deepfakes go public

There are now multiple open-source, freely available methods for creating synthetic faces built on GAN architecture. And as cloud services like Amazon’s AWS and Google Cloud have become easier to access, so has the ability to train these algorithms. The most well-known in the A.I. research world is StyleGAN, made by Nvidia. It was released in December 2018, and while able to produce incredibly high-quality images of fake faces, the images also contained strange blobs and digital artifacts. Less than a year later, the Nvidia team released StyleGAN2, which fixed the algorithm’s architecture to prevent those blobs and artifacts from forming, as well as improving the fidelity of the images.

These algorithms are able to be adapted to different domains. By training the algorithms on pornographic images rather than just faces, the system was able to adapt to generating something it might not have been ever intended for.

GANS have also been ported to specifically make deepfakes, through open-source projects like DeepFaceLab and Wav2Lip. The ease of using these services can’t be overstated: The Wav2Lip project’s website shows how a single line of code can be used to automatically make the subject of a video lip-sync to any audio file.

These technologies are still in their infancy, and they’ll continue to become more accurate and convincing. Some of the applications of these technologies are genuinely entertaining — check out the Avengers singing the “Sweet Child O’ Mine” scene from Step Brothers — but ultimately, these algorithms are also now much easier for anyone to use for malicious ends. And without any recourse, deepfake’s harms might outweigh their slight entertainment value.

Latest posts