Three Dudes Run The Biggest AI Romantic Fantasy Site For Women
A fter moving to a new city in North Carolina in 2024, Cookie (a pseudonym) felt the weight of a new city. Their husband was traveling a lot for work, and as a stay-at-home parent with a now 4-year-old daughter, the days were “very draining”. Since Cookie didn’t know anybody in their new town, they turned to Janitor AI, a social chatbot site known for its unbounded, often explicit, fantasy roleplay. It was a “nice release,” Cookie told Forbes.
Cookie grew up around fantasy and romance novels—their mother kept a collection—and Janitor AI became an easy way to escape the drudgery of the day-to-day. By the time their daughter is down for a nap or tucked in for the night, Cookie is creating "slow burn" romance characters with detailed and often explicit prompts. There’s Charlie , a nudist werewolf roommate; Marcus , a seven-foot ghoul with a taste for dive bars; Greenwood, Colorado, a fictional town where humans live alongside supernatural “demihumans.” Beneath Greenwood’s romance and monster lore is a civic rot: a glossy new church masking an organ-harvesting operation, with seedy bars serving as bait.
Cookie is one of Janitor AI’s 2.5 million daily, die-hard users. The platform claims more than 15 million total users and with 100 million monthly visitors, and it's the tenth most popular consumer AI app, according to Similarweb , a digital market intelligence company.
Janitor AI is part of a growing number of companies that offer social chatbots that adopt fictional personas or impersonate real people. The category is highly immersive: users of the top roleplaying AI apps spend on average 78 minutes a day on their mobile apps, according to market intelligence firm Sensor Tower, nearly as long at TikTok.
If your mental picture of “AI romance” is awkward guys trying to outsource intimacy, Janitor AI’s demographic muddles that stereotype. The company estimates between 70 to 80% of its users identify as women.
Don’t call it porn, either—though it’s certainly explicit and not-safe-for-work. Janitor AI pitches itself as entertainment, first and foremost, more akin to an ‘HBO for AI’ or a platform for interactive fiction, where the reader also gets to be the writer.
“It’s an evolution of the amateur novelist... for people who would have never written a book, they can make something way deeper and interactive, and a lot quicker,” says its founder, 26 year-old Jan Zoltkowski.
For Cookie, that means drafting thousands of words of character backstory in Google Docs and spitballing tropes in a private Discord with five other Janitor AI creators.
In that way, Janitor AI is the digital evolution of the romance novels Cookie has always loved, with a community layer bolted on. In this world, though, the voracious reader, without much effort, can also be the author. Nearly 400,000 users have joined Janitor AI’s Discord server, forming their own sub-Discords that function like ‘writer’s rooms’ for specific tastes and story worlds. The most rabid creators spend many hours a day on the site, learning scripting and CSS to customize their profiles, some creating guides hundreds of pages long to teach other creators.
Janitor AI is riding the broader “romantasy” wave—fantasy plus romance—that has been propping up book publishing. While categories like nonfiction have softened, romantasy sales increased 50% in 2024, powered by blockbuster series from Rebecca Yarros (dragons) and Sarah J. Maas (fairies). Published last year, Yarros’ newest book, Onyx Storm, is now the fastest-selling adult fiction release since they started tracking in 2004, said Circana, a company that monitors book sales. The phenomenon has even made its way into brick and mortar, sparking a slew of romance-themed bookstore openings.
The idea of romantasy is enough to make some Silicon Valley investors flinch. They lump Janitor AI into the “vice” investment category alongside cannabis and porn, which many funds are restricted from investing in.
But Janitor AI and its backers argue the stigma is lazy.
“If you think about Reddit, Tumblr, X, Snapchat, they all had higher percentages of not-safe-for-work content when they first got started,” said Mercedes Bent, co-founder of venture capital fund Premise, and a Janitor AI investor. “Tumblr and Reddit were well over 50% in the early days, but the fundamental substrate of what the platforms were offering was a new way to communicate.”
Sex is often a highway to quick adoption of a new entertainment platform, she said, pointing to a long history of technological trends like VCRs that were first adopted by the adult industry.
Bent says the venture community often reflexively writes off female-focused adult entertainment. “There’s a very persistent trend throughout history, which is that if something caters to women's health or to women's desires, it is unnecessarily penalized,” she said.
There’s also a more practical reason the big AI labs keep their distance: legal, regulatory, and moderation risk. OpenAI was planning to allow erotica for adults, then shelved it earlier this year after internal backlash. Janitor AI exists in a perpetual arms race with its users, who constantly push the boundaries of permissible content.
Its dense content guidelines is a roadmap of this conflict, showing how users can achieve maximum eroticism while staying within the guardrails of non-pornographic content. For instance, the rules for digital avatars are comically precise: “subtly wet” clothing and “overly defined bulges”are forbidden though pubic hair is permitted to “peek out from clothes.” Sex with household appliances and other inanimate objects is allowed; with robots it’s also okay so long as they are smart and can communicate consent.
While Forbes found no active lawsuits against Janitor AI in the U.S., its predecessor, Character AI, has been hit with many, mostly around children (Janitor AI does not allow anybody under 18 to sign up). This week, the state of Pennsylvania sued Character AI, saying it found chatbots impersonating medical doctors and offering prescriptions, the state of Kentucky said earlier this year that Character AI was preying on children and exposing them to sexual conduct, and in the same month, it settled a wrongful death lawsuit over a 14-year-old user who died by suicide after developing an intense emotional bond with a chatbot.
Zoltkowski, a self-taught engineer from Brisbane, Australia with a libertarian streak, dropped out of university during freshman year to work at cryptocurrency exchanges Bitfinex and Bittrex. After a few years, he started tinkering with early versions of OpenAI’s GPT models at an EthDenver hackathon in early 2023. He quickly became an avid Character AI user, discovering that he liked “mean” chatbots that made him feel like he had “no agency” (a favorite on Janitor AI is Nova Marino , an infamous gang leader who likes to bite her partners).
Then Character AI tightened its filters to block erotic content, triggering outrage among its diehard fans, who left the platform in a movement they dubbed “The Great Exodus.” Disgruntled and ideologically motivated, and at the same time enamored with AI, Zoltkowski decided to build an alternative that allowed sexual chats.
With the help of AI, he says it took him just three weeks to write the code for the initial site. He called it Janitor AI, because in Latin, the word refers to a guardian who has keys to many doors.
On June 16, 2023, he posted to one of the Character AI Reddit forums: “Hey, launching website, check it out.”
A few days later, Janitor AI went viral on "BookTok," the influential romance and fantasy subculture of TikTok. The platform hit one million users in just 21 days.
Success nearly crushed it. Traffic spiked, infrastructure costs followed and Zoltkowski says he was soon on the hook for some hefty bills—including a $180,000 invoice from web hosting and security company Cloudflare.
Then: plot twist. Martin Shkreli, the smirking patron saint of pharmaceutical price-gouging, direct messaged Zoltkowski on Discord, promising GPUs and funding in exchange for equity. Desperate and perhaps a bit naive, Zoltkowski agreed, but the deal soured quickly when the promised wire transfers never arrived. In the end, it was an angel check from Sky9 Capital that saved Janitor AI.
Defying the Silicon Valley playbook
Janitor AI runs on a “crazy low” burn rate, Zoltkowski says, one that would make most other venture-backed startups blush. The company only has three full-time employees, all men, who live and work together in a lofted space in San Francisco’s SOMA neighborhood. (It employs some contractors who are women). The top two executives, Its CEO Zoltkowski and Hugo Smith, its COO, are 26 and 22, respectively.
Even after a May 2025 Series A led by Khosla Ventures, Zoltkowski won’t say how much money they’ve raised. He’d rather focus on how leanly they run the company; there’s some scar tissue from that $180,000 near-death experience.
“I just never want to be in that position again,” Zoltkowski said. “People don’t know how far money goes when you actually optimize, and that’s why we’re so focused on being frugally minded with every dollar of compute."
His method is part scavenger hunt, part high-end engineering. He’s constantly testing smaller open-source models that can be fine-tuned for Janitor AI’s roleplay needs. Zoltkowski says he’s used Mistral’s models in the past, and is currently experimenting with Google’s family of small open-weight Gemma models, though he’s found they can be prudish in ways that are quickly noticed. He tweaks the code to squeeze it for maximum efficiency.
Janitor AI says it serves about 13 trillion tokens a month for about $130,000. Running a similar volume through AWS’s cheapest, smallest models would cost at least $1 million and using more expensive Anthropic models could run more than $50 million, according to pricing on the AWS website.
The end goal is total vertical integration. Zoltkowski wants to pre-train Janitor AI’s own models because current open source models tend to be optimized for coding performance, not the narrative range and variability erotic roleplay demands.
“For instance, in the new Google Gemma models, I noticed the word ‘pussy,’ is kind of shadow-banned in the [model’s] weights,” Zoltkowski said. “But we’ve been able to bring it back into the sample. That’s what all the tuning is for, right?”
Running lean is fun—until it isn’t. Zoltkowski has been able to keep costs low by running Janitor AI like a donation-supported, passion project, but as the site has grown into a behemoth, the reality of running a legitimately large social media site are surfacing: policy, enforcement, legal risk, and the boring work of consistent governance.
Unsurprisingly, the hardest part is moderation. There’s a reason big labs avoid the erotic roleplay market: it is a liability minefield.
Fantasies are by nature illicit, and the boundary between creative expression and prohibited content can quickly blur. Explicit sexual content involving children is an obvious red line. But the site also has to decide what to do with extremist themes like necrophilia and consensual cannibalism.
It hasn’t been easy. Earlier this year, the community blew up over Zoltkowski’s reliance on volunteers, not paid employees or contractors, to moderate content.
In January, a moderator resigned , publishing an indictment of Janitor’s system, alleging it was “performative” with just two to three volunteers managing a massive backlog of some 125,000 incident reports. The author alleged that safety rules were inconsistently enforced, moderation tools were lacking and that developers frequently overturned bans for extreme content when popular creators complained privately.
The uproar punctured the founder myth for Zoltkowski—"shep" on Discord—who had cultivated near-mythical status among his users and had personally handled much of the moderation.
Janitor AI told Forbes the ex-moderator’s claims false, and that the company invests heavily in both AI and human-reviewed moderation. But in an apologetic note to the community at the time, Smith, the company’s COO, conceded that the company needed to do more to professionalize its moderation.“Some of this is unavoidable—we’re operating in genuinely uncertain legal territory, and we’re still figuring out where the lines should be,” he wrote. Zoltkowski echoed the sentiment on Discord, saying a “handful” of claims were incorrect, but “for the most part the complaints are painful because they are true.” They have since stopped using volunteer moderators.
As Zoltkowski works to turn Janitor AI into a legitimate business, he’ll need to balance scaling up without losing the freewheeling, libertarian bent that birthed the site. Currently, the site is free, but the team is working out how to roll out a freemium subscription tier.
A year ago, Zoltkowski wrote an impassioned appeal to the community explaining why the company needed stricter image controls: to get a payment processor comfortable enough to support subscriptions, the platform needed images that wouldn’t be classified as pornographic. Many users hated it.
“JanitorAI isn't just a site to us, it's our passion project which has brought so much joy and creativity seeing this amazing community grow," he wrote. “I’m begging you to stick with us through this, Janitor is all of our baby and we are not ready to let it go.”
As Janitor AI matures, Zoltkowski is betting that the platform can transition into mainstream entertainment: he says his goal is to create something similar to kids’ gaming site Roblox, but R-rated. If he succeeds, he’ll be the architect of a new, interactive medium the industry’s giants like OpenAI were too prudish to claim.
He’s not there yet. First he has to pull off the hard part: bolt on the rules, controls, and other adult supervision that come with scale—without spooking the very users who came to Janitor AI because it wasn’t supervised.
Loading article...