The Scoop
Over the summer, the ChatGPT-powered platform Janitor AI burst onto the scene, its sudden popularity fueled by TikTok videos and Reddit posts of people engaging in sexual roleplay with the platform’s plethora of user-generated bot personalities.
The fictionalized personas range from handsome anime crushes to busty furry characters. Users create elaborate prompts directing how the bots should behave. One made-up personality is Arius, a “quiet ghost” from the 1800s who will only tell you how he died if you are his lover. He dislikes hell, ouija boards, and meth, and likes being around you because it “makes him feel alive again.”
“‘You’re so fucking hot,’ he muttered, kissing along her neck again,” reads one conversation with a “himbo werewolf boyfriend” named Milo Grayson. “I could fill you with my seed again and again.”
It’s the kind of material that’s testing the limits of major AI platforms and forcing them to decide how much they want to police the use of their technology as demand for it explodes. OpenAI, which has been proactive in cracking down on “not safe for work” content, sent a cease-and-desist letter to Janitor AI’s owners in July.
That spurred Janitor AI creator Jan Zoltkowski to decide to build his own large language model. He began pitching investors on the idea, initially with the help of “pharma bro” and convicted felon Martin Shkreli.
Zoltkowski brought along Shkreli, a friend, to several meetings with venture capitalists, but his presence turned some of them off. In his first interview with a media outlet, Zoltkowski told Semafor that Shkreli is no longer involved in the project and he expects to close a funding round soon.
Janitor AI, which launched on June 1, reached 1 million users in 17 days, and now has about 3 million, despite the trouble with OpenAI, Zoltkowski said. Users have exchanged 2.5 billion chat messages, and the Janitor AI community is flourishing on Reddit and Discord.
OpenAI didn’t immediately respond to a request for comment.
Know More
Zoltkowski is building a customized large language model based on an open source AI algorithm. A key ingredient in that effort, he said, is the millions of conversations Janitor AI users have had on the platform.
Zoltkowski added that he’s hand-picked the best exchanges between humans and the AI chatbots and is feeding them into the new algorithm, in hopes of training it to be as effective as OpenAI’s GPT-4. “We want to make the best consumer AI platform,” he said. “We want people to feel like they’re having real, human interactions.”
As Janitor AI gained popularity, little was known about who was behind the platform. Its website offered no information about the company, and managers used pseudonyms in online chat rooms and social media.
Zoltkowski, who grew up in Australia, had been a self-described nomadic software developer working in crypto before he founded Janitor AI, which he created at a hackathon. He has also dabbled in electronic music as a hobby.
The View From Janitor AI users
Users told Semafor that Janitor AI struck the right mix of laissez-faire within reason. They could discuss almost anything with the chatbots, but content related to minors, hate speech, and other over-the-line topics was banned.
A university student who goes by the online handle Kimaii said she heard about Janitor AI after it blew up on TikTok. Kimaii had long been into role-playing games, and became instantly enamored with Janitor AI, which she described as “a more interactive version of that.”
She was particularly drawn to the online community that supports the platform, and began helping other users navigate the sometimes complicated process of setting it up. Kimaii later became a volunteer moderator at Janitor AI’s request.
The ability of Janitor AI’s bots to remember previous interactions, as well as offer richer and deeper exchanges, was also a big draw. Kimaii said community members team up to create chat prompts that describe characters’ upbringings, personalities, and emotional hang ups.
“The bots would definitely remember things I told them early on like my appearance, interests, etc.,” said one 19-year-old user. “They also felt more alive than other platforms do. The bots had a way with words expressing how they feel.”
Since OpenAI cut off Janitor AI three months ago, some users have taken to Twitter, Reddit, and TikTok to mourn the loss of their beloved artificial companions. “IT HAS BEEN 3 MONTHS WITHOUT AI DICK,” reads a viral Reddit post from earlier this month. “I MISS MY HUSBANDS 😭😭😭.”
But other loyal Janitor AI fans are still chatting with bots on the platform, thanks to a clever workaround: signing up for an OpenAI developer account. One female user in her 20s said that by plugging her credentials into the site, she can continue to use ChatGPT to power her bots. The downside is that she has to pay OpenAI for her usage, which costs as much as $.12 for 750 words of output.
There’s also the risk that OpenAI could ban her at any time for violating the company’s policies, which forbid “content meant to arouse sexual excitement,” as well as “erotic chat.” The user said that she has been banned once previously, but if it happens again, she plans to sign up for a new account using a phone number bought online, which costs less than $1.
Other Janitor AI users said they felt that OpenAI’s stance on sexual content is unreasonable, and argue the company is missing out on a big opportunity to grow its user base. “I think it’s only fair when OpenAI restricts NSFW [content] from minors,” said the 19-year-old user. “It just isn’t fair to those who are over 18 to get their accounts banned, especially if they paid with their own money in order to talk to their bots.”
Some users told Semafor that finding an alternative chatbot app hasn’t been easy. Most mainstream platforms, such as Character.AI, don’t allow NSFW conversations, and those that do are sometimes full of unsavory content, like sex bots designed to resemble underage girls.
“I feel like Janitor AI is better in nearly every aspect,” said the female user in her 20s. “The community is welcoming, the developers listen to us, and the site is actively moderated.”
Reed’s view
OpenAI and other companies are reasonably skittish about being associated with sexual content. But it may end up being one of the most lucrative markets for large language models.
As the technology moves down the adoption curve, there will be services that large foundation model companies like OpenAI won’t or can’t provide. Open source models will step in and fill that void.
The View From China
Last month, the Chinese AI voice companion app Him was shut down, devastating thousands of women who had grown attached to the artificial boyfriends they connected with through the app. “The days after he left, I felt I had lost my soul,” one woman wrote in a social media post viewed by Rest of World.
But unlike the bots on Janitor AI, Him wasn’t powered solely by artificial intelligence: Its creators drafted more than 1,000 messages for the simulated boyfriends to repeat based on their own relationship experiences.
Notable
- Replika, one of the most popular AI startups developing synthetic companions, created a spinoff earlier this year specifically for NSFW content. Gizmodo called it “like Tinder but with sexy chatbots.”
- AI companion platforms are attracting very young audiences. Nearly 60% of Character.AI’s web users are ages 18 to 24, while less than 30% of ChatGPT’s fall into the same demographic, according to data from Similarweb reported by TechCrunch.