Scientists Processed 109 Hours of Oral Sex to Develop an AI that Sucks Dick

The Autoblow AI, a sex toy made for penis-possessing individuals, consists of a rubbery sleeve and a motor housed inside a canister that aims to simulate oral sex. It launched its Indiegogo campaign last week and quickly hit its goal of $50,000. Its biggest advertised advantage over the original model from 2014, the Autoblow 2, is a machine learning algorithm that “continually changes technique” in order to pleasure the user in new and exciting ways.

Instead of repetitive, mechanical motions, this "AI mode" promises to replicate the nuanced and unpredictable motions of a real, human blowjob. In order to do this, the company asked a team of six people to watch and annotate 109 hours of porn and hired machine learning engineers to create a model to take all that data and translate it into what the toy does. This entire process took three years.

The result of this research is the Blowjob Paper, a (definitely not peer-reviewed) study that’s full of sexy—or at least, sexually-themed—algorithmic research: “In this work, we seek to quantify the ‘common’ or ‘typical’ movements involved in oral sex performed on males,” the paper begins. “ To do so, we analyze a dataset containing over 108 hours of pornographic video, annotated at each frame with the position of the lips along the shaft of the penis. We use quantization techniques to discover sixteen distinct motions, and using these motions we design and evaluate a system that procedurally generates realistic movement sequences using deep learning. We quantitatively show that this system is superior to simple Markov Chain techniques.”

"I had no problem with it.”

I asked Brian Sloan, the creator of Autoblow AI, about his attempt to recruit serious scientists to work on a jerkoff machine. His pitch to them didn't mince any words.

“I told them I make a sex toy for men focusing on recreating the blowjob experience, and asked if they could use AI to study what really happens during blowjobs, so that I could make my machine give blowjobs like a human,” Sloan told me in an email. “That was enough to cause seven companies to tell me almost immediately that the ‘work was not for them.’ When I asked why, they didn't want to talk about why.”

That eighth team said yes, and produced the Blowjob Paper, which Sloan said he paid $30,000 for. But for all the beautiful calculations and data collection methods this research holds, one thing is conspicuously missing in the paper: the names of the authors. There are no names on this paper, no research teams to credit, no institutions claiming responsibility for this work. The scientists did the work anonymously because they believed their other clients might drop them if they were publicly associated with it.

Sloan connected me with these anonymous scientists using a dedicated email address to guard their identities. I also viewed emails between Sloan and the scientists documenting progress on the paper during its production, with the names of the scientists redacted.

“Brian reached out on social media... I think on LinkedIn,” the scientists behind the anonymous email address wrote back. “Our initial discussion was not about sex toys and porn. It was technical; dealing with video analysis and artificial intelligence stuff. He was pretty open about the fact that several teams did not want to work on this sort of project. I had no problem with it.”

This scientist who wrote the email said they have a “PhD on AI stuff,” and have appeared in more than 30 publications and patents. The rest of their team are mostly engineers, they said.

Enter the Sliderman

The Blowjob Paper begins with a few mild charts and graphs about how the researchers quantified movements gleaned from videos of oral sex and translated them to the different movements Autoblow AI offers. Among these are cluster charts that represent activities, including pauses, edging (meaning pleasuring someone to the verge of climax), shallow tip play, and deepthroating.

The researchers designed a dense neural network (DNN) architecture that predicts a blowjob giver’s next move based on analysis of past movements, kind of like predictive text on your phone. On paper, the process for creating these models looks something like this:

To collect this data, Sloan enlisted the help of Dalibor Copic, who is based in Serbia. Copic works for Sloan’s Very Intelligent E-Commerce Inc., a marketing company that has churned out other adult-themed campaigns, including the World’s Most Beautiful Vagina competition and the Balls Contest. Very Intelligent E-Commerce Inc. is very successful at attracting press coverage for Sloan's crowdfunding campaigns for adult products by promoting them with a science-themed gimmicks. Before the Blowjob Paper, the same team produced the Vulva Paper for the Autoblow Vaginal Beauty contest.

Copic called together a team of six people to collect blowjob data from watching porn. These were not statistics or machine learning experts, just friends and confidants willing to help with an absurd task. They created a system for annotating 109 hours of blowjob videos—1,200 oral sex video clips in total—by animating a cartoon head and penis on a movable slider they could adjust with a mouse as they watched the porn videos. Copic told me in an email that they named the head “Sliderman.”

“Once oral sex is started, and when the girl is doing what she is doing, we move the Sliderman in order to follow the oral sex and that is how we reproduced the oral sex to our system,” Copic told me in an email. They watched the videos at half-speed, and collected the values from the slider to capture the data they needed.

I asked Copic whether he found this work tedious. “Well, to be honest, I was lucky and unlucky :),” he said. “A few videos at the beginning of the project were interesting... until the end, [it] was a bit I'll say, heavy, to watch. So much oral sex...”

"At the cutting edge"

“It's actually an interesting problem to work on,” the anonymous scientists who wrote the Blowjob Paper said. “Right now we are in the machine learning gold rush, and so developers can be selective about what projects to take on. My criteria for what we work on is to find projects that are interesting, and that pay. This project ticks both boxes.”

If they were so open and eager to get to work on automated AI blowjobs, I asked, why stay anonymous? They told me they asked a few of their clients if taking on this work and being publicly associated with it would be a problem.

“The response was surprising,” the scientists said. “Corporations don't want to be associated in any way with controversy. It's just not acceptable to them, and they view this as a PR nightmare. So, I am staying anonymous. On a personal level we didn't get pushback, it's more on the corporate level that being identified is a problem. For context, a lot of our work is for FinTech and/or government. That's where we got the red flag from.”

This isn't surprising. We’ve seen hostility against sexual content again and again from payment processors and lawmakers.

Aside from the math, it’s an accessible, easy to swallow breakdown of how AI is coded into products, as well as things we use every day, like Google’s email autocomplete suggestions and search results. They’re all learning from our past experiences, and the Autoblow AI is, in theory, no different. It’s just learning from porn.

Read more: Don't Let This Robot Suck Your Dick

The anonymous scientists told me that this was not a throwaway project for them, either. “My takeaway is that convolutional neural networks are super fussy, and the fact that AI can understand such a super high level idea as a blowjob is somewhat surprising,” they said. “I feel like we are really out there at the cutting edge, applying libraries that didn't exist six months ago.”

The Autoblow A.I. is loaded with a lot of math, and the sweat, tears, and potential carpal tunnel syndrome sacrifices of many people. But at the end of the day, despite all the statistics and machine learning, it’s still just a plastic can with a silicone sleeve that goes up and down on your dick. These researchers imagine a world where immersive, interactive aspects of porn and sex toys (and non-sexual entertainment, too) can be improved with AI—but it wouldn't happen overnight.

“Sex toys are available basically across the developed world. Obviously people want these things, and they surely will get smarter as scientists and engineers like me do their job,” they told me. “That's the big picture. There are of course also technical upgrades we can do to the solution we built for this particular sex toy, like better procedural generation of the emulated blowjobs, and more stuff than just blowjobs. We found that recognizing anal, for example, is totally doable.”