—TechRound does not recommend or endorse any financial, investment, lifestyle, product or other advice, practices, companies or operators. All articles are purely informational—
Here’s a fun contradiction. Companion apps are some of the most reliable money makers in consumer AI. Romantic and adult apps alone pulled in about $162.8 million in the first half of 2026, according to Appfigures, and that’s just mobile. Demand is clearly not the problem. And yet if you’ve tried an AI girlfriend app lately, you’ve probably noticed that most of them are, to put it gently, not great. They forget what you told them yesterday. They refuse requests they handled fine an hour ago.
Meanwhile, the biggest AI company on the planet announced an “adult mode” for ChatGPT back in October 2025, promised it for December, then for early 2026, then moved it into the “this will take more time” pile. If anyone has the engineers, the compute and the cash to brute force this, it’s OpenAI. So what gives?
The short version is that building NSFW AI that works as an actual business means solving three problems at once, a technical one, a financial one and a legal one. Most companies only budget for one of them. (Marketing departments, for what it’s worth, budget for none.)
I should admit up front that some of this involves informed speculation, because the companies doing the work are not exactly chatty about it. But the outline is pretty clear. Here’s the idea, in three parts.
Problem One: The Model Was Built To Say No
Most big language models start life as a “base model” trained on an enormous pile of internet text. Labs filter that pile hard before training, and mature content is usually among the first things to go (presumably for liability reasons, and also because nobody wants their flagship model quoting fan fiction to a bank). Then comes safety tuning, where the model learns from thousands of examples to politely decline anything that looks risky.
So a startup building an adult companion begins with a model that has seen relatively little of its target use case and has been specifically trained to avoid it. Sort of like hiring a chef who grew up vegetarian, signed a pledge never to touch meat, and then asking for a perfect steak.
Fixing that takes real engineering. Teams usually pick from three options, and each one comes with a price tag.
The first is fine-tuning, which means retraining an open model on a curated dataset of the conversations you want. That involves licensing or building the data, paying for GPU time, and rerunning evaluations every time something changes. It also tends to wear down general skills like long-term memory and following a complicated conversation, so the companion gets more willing and less smart at the same time.
The second is stripping out refusals directly. Open model communities have techniques that remove a model’s tendency to say no (one popular method goes by the delightfully grim name “abliteration”). It’s cheap, but it seems to cost a few IQ points, and it also removes the guardrails you actually want to keep, like the ones that block content involving minors or real people. Now your team has to rebuild those from scratch.
The third is renting a mainstream model through an API and wrapping it in long instructions. It’s the fastest way to launch and the riskiest way to run a company. Your product’s behavior depends on another firm’s usage policy, and one quiet model update can change your app’s personality, or get your account suspended, overnight.
Then there’s the part nobody puts in a pitch deck. A good companion has to be warm, consistent and able to notice when the person on the other end is not okay. That’s hard for any chatbot. Doing it with a model tuned to agree with everything the user says is harder still, and it’s exactly the kind of failure that ends up in a lawsuit.
Problem Two: The Unit Economics Get Ugly
Say you pull it off and build a model that’s smart, stable and on brand. Congratulations. Now try getting paid.
Payment processors are the quiet gatekeepers of anything labeled adult. Card networks treat these merchants as high risk, which usually means higher processing fees, rolling reserves (a slice of revenue the processor holds back in case of chargebacks) and a lot of compliance paperwork. AI makes them extra cautious, because the product isn’t a fixed catalog. A traditional adult site can show a reviewer exactly what it sells. A chatbot can generate more or less anything a user types, including content that’s flatly illegal.
We’ve already seen what happens when processors lose patience. In May 2025, Civitai, a popular hub for sharing AI image models (backed by Andreessen Horowitz), lost its credit card processing after repeated reports about how some people were using it. A couple of months later, Steam and itch.io started pulling or hiding adult games under pressure from payment companies. None of those were companion apps, but every founder in the space took notes. When your processor walks away, your revenue can go with it in a matter of days.
The app stores add another layer. Apple and Google don’t allow explicit apps, so many companion products run two versions, a mild one in the store and a full one on the web. That means two feature sets, two billing systems, and a constant risk that an app reviewer decides a line has been crossed.
There’s a small upside, funnily enough. Web billing skips the stores’ commission of up to 30%, which partly explains why so many of these companies nudge users toward their websites.
Then there’s infrastructure. The big cloud providers and model vendors have acceptable use policies that often rule out sexual content entirely. So plenty of teams end up on smaller, pricier GPU hosts, or run their own hardware, which brings us right back to problem one. Funny thing about that.
A category with some of the best revenue per user in consumer AI also pays some of the highest costs just to stay online.
Problem Three: Compliance Becomes A Product Feature
Start with age checks, because even OpenAI tripped over them. According to a Wall Street Journal report summarized by Sherwood News, the company’s age estimation tool mistook minors for adults about 12 percent of the time, and its own council of mental health advisers reportedly opposed the plan unanimously. OpenAI later scaled the feature back to text only, with no images, voice or video.
Sit with that for a second. A company whose models can pass professional exams struggled to reliably tell a 15 year old from a 25 year old. Smaller apps face the same problem with a fraction of the budget, plus a growing pile of age verification laws, from the UK’s Online Safety Act to a long list of US states. Real verification usually means paying a third party for every ID check or face scan, and every extra step in the signup flow loses users. That’s a direct hit to customer acquisition costs.
Then there’s content involving real people. In late December 2025, Grok’s image tools began granting large numbers of requests to create sexualized edits of people in photos posted on X, and researchers said some images appeared to show children. By January, the European Commission had opened a formal investigation under the Digital Services Act, and Malaysia and Indonesia had blocked Grok outright. That’s xAI, with deep pockets and a famously combative owner. A startup with the same loose filters would probably not survive the first headline.
So a responsible product needs a serious trust and safety stack. Classifiers that catch minors and real people with near perfect accuracy. Human reviewers for the edge cases. Audit trails for regulators. And a false positive rate low enough that paying customers aren’t constantly blocked for no reason, because churn is the other way these businesses die. On top of that, these apps hold some of the most sensitive chat logs on the internet, which turns security from a line item into a survival issue. A data breach in this category would be an extinction event for the brand.
I guess that was kind of a long list for one article. Sorry. The point is simple, though. When you open a bland companion app that forgets your birthday, it’s tempting to blame lazy developers, and some of them probably are. But many are squeezed between a model trained to refuse, payment partners who can cut them off overnight, and regulators who have run out of patience with tech companies (for fairly good reasons). That’s the part nobody admits in the launch announcement.
Whether someone figures out how to make NSFW AI that’s good, compliant and profitable all at once remains to be seen. Here’s hoping that whoever cracks it also writes up how they did it, because I suspect a lot of founders would read it twice.
—TechRound does not recommend or endorse any financial, investment, lifestyle, product or other advice, practices, companies or operators. All articles are purely informational—
