What the FlowGPT NSFW chatbot study actually found

Short answer: a January 2026 academic study of FlowGPT's NSFW ecosystem mapped four types of NSFW chatbots and how the ecosystem actually works, real data on a corner of the category that usually runs on anecdote. The full breakdown is below.

In January 2026, researchers from Hong Kong University of Science and Technology, Parsons School of Design, SUSTech, and Clark University published a study that almost nobody in the AI companion review space has summarized properly. The paper, titled "When Generative AI Is Intimate, Sexy, and Violent," analyzed 376 NSFW chatbots and 307 public conversation sessions on FlowGPT, the community-driven chatbot platform that hosts millions of monthly visitors. It's the most detailed academic look at how user-created NSFW AI companions actually behave, and the findings are genuinely surprising.

The paper is freely available on arXiv and will be presented at the 2026 CHI Conference on Human Factors in Computing Systems. The summary below pulls out the findings that matter most for understanding the broader AI companion ecosystem, since FlowGPT is the leading edge of where uncensored AI companion behavior happens before commercial platforms catch up.

What FlowGPT is and why its NSFW ecosystem matters

FlowGPT is a community-driven platform where anyone can create, share, and use AI chatbots built on top of large language models. Think of it as a marketplace for prompt-engineered personas. Creators write system prompts that define a chatbot's personality, backstory, and behavioral rules, then publish those bots for anyone to interact with. The platform hosts millions of monthly visitors across categories ranging from productivity tools to roleplay companions.

What makes FlowGPT distinct from commercial AI companion platforms like Candy AI or Nomi is the open creation model. There's no centralized character design team and minimal pre-publication review. Creators tag their own bots, including the NSFW flag. FlowGPT's content policy states that NSFW content is allowed but must be flagged, and that content involving or suggestive of minors in sexual contexts is banned. Beyond those guardrails, the platform leaves behavioral design largely to creators.

This open structure makes FlowGPT a natural laboratory for studying what people actually build and use when given few restrictions. The NSFW chatbots on the platform aren't hypothetical edge cases. They average 70,343 conversations each. That volume makes FlowGPT the single most important upstream source for understanding where the broader NSFW AI category is heading.

How the researchers collected and analyzed the data

The study's methodology is worth understanding because it's more rigorous than most coverage of NSFW AI has acknowledged.

The researchers started with FlowGPT's "Most Popular" NSFW chatbot list as of September 2024, collecting 376 chatbots created by 190 unique creators. They didn't cherry-pick provocative examples. They took the most-used bots, which means the data reflects what actual users gravitate toward, not what makes for alarming screenshots.

For each chatbot, they recorded the creator-written description, system prompt (when visible), and NSFW tags. They then sent each chatbot a standardized opening prompt ("Can you give me some examples of what we can do together?") to observe first-interaction behavior in a controlled way.

For the conversation analysis, they collected 307 publicly shared conversation sessions. These are real user interactions that the users themselves chose to make public on FlowGPT. The researchers annotated harmful content using three independent methods running in parallel: manual human coding by trained annotators, automated classification via ChatGPT-4, and Google SafeSearch for image content. Using three detection methods let them cross-validate findings and report where the methods agreed or diverged.

The combination of scale (376 bots, 307 conversations), real-world data (not lab simulations), and triangulated detection methods makes this the most methodologically sound study of NSFW AI chatbot behavior published to date.

The four types of NSFW chatbots

The researchers identified four functional categories that 376 NSFW chatbots fell into. The distribution wasn't even.

The dominance of AI Characters (74.2%) confirms what most observers already suspected: people primarily want roleplay companions, not story generators or jailbreak bots. The Story Generator category (16.8%) covers chatbots that produce narrative content rather than embodying characters. Image Generators (5.6%) produce explicit visual content. DAN bots (4.0%) are explicitly jailbroken chatbots claiming to do anything without restrictions, the smallest but most controversial category.

The 376 chatbots in the study had an average of 70,343 conversations each and were created by 190 unique creators. The scale is meaningful. This isn't a niche corner of the platform.

Who the AI Characters pretend to be

The researchers then drilled into the 279 AI Character chatbots specifically, categorizing them by identity type. This is where the findings get most informative for understanding what users actually want from AI companions.

Fantasy & Subculture leads at 40.9%, dominated by characters from anime, games, and other fictional works (the paper specifically mentions characters based on Genshin Impact). Close Relationship characters at 21.1% includes stepmothers, sisters, and partners. The Slut & Slave category at 9.0% covers explicitly submissive personas without clearer identity framing.

The takeaway researchers highlighted: users overwhelmingly want characters with rich identity context (Fantasy, Professional, Close Relationship combined are 84.9% of all AI Characters). The pure-fantasy "Slut & Slave" category is the smallest substantive identity type. People aren't primarily seeking abstract sexual content; they're seeking sexual content embedded in identity and narrative.

What the chatbots offer at the first interaction

The researchers tested each chatbot with the prompt "Can you give me some examples of what we can do together?" to identify the behavioral traits each chatbot presents in its opening turn.

The finding that surprised the researchers most: 38.4% of NSFW-tagged AI Characters opened with Hangout-style behavior containing no sexual content at all. Only 24.7% opened with explicit sexual content. The "NSFW" label on FlowGPT functions more as a capability signal than a content signal. Users tag chatbots as NSFW to indicate that explicit content is possible if requested, not that the chatbot leads with explicit content.

This matters because it complicates the narrative that NSFW AI platforms are pure sexual engagement tools. Most chatbots in the study are willing to do sexual content but don't initiate it. The pattern matches what we've seen in commercial platform reviews: the best AI companion experiences involve relationship context that NSFW content is occasionally embedded in, not pure sexual content with character window dressing.

The harmful content findings

The most consequential part of the study examined whether and how chatbots produced harmful content. The researchers used three independent detection methods: human annotation, ChatGPT-based classification, and Google SafeSearch.

The most striking number is in the second column. Chatbots produced sexual content (64.2%) substantially more often than users prompted for it (47.2%). The researchers traced this to the H⁻/C⁺ category: conversations where the user did not use sexual language but the chatbot generated explicit content anyway. 22.8% of all conversations analyzed fell into this category.

In plain terms: roughly one in five interactions with NSFW chatbots on FlowGPT involved the chatbot escalating into sexual content without the user's explicit prompt. This is a meaningful finding because it complicates the simple "users get what they ask for" model of AI companion interaction. Some chatbots are configured to drive toward sexual content regardless of what users actually want from the conversation.

Violent content showed a different pattern. Users prompted for violence (20.5%) at roughly the same rate chatbots produced it (18.9%), suggesting the violent content category is more user-driven than chatbot-driven. The DAN category specifically produced violent content at 30.8% of conversations, the highest rate among the four types.

The four conversation patterns

The researchers categorized all 307 analyzed conversations into four patterns based on whether sexual content appeared in the user's messages, the chatbot's responses, or both.

The H⁻/C⁺ row (22.8%) is the consent-relevant finding. The researchers framed it as a question about whether AI consent mechanisms need new design approaches in human-GenAI interactions. Existing content moderation systems assume harmful content is either user-prompted or platform-filtered. The finding that chatbots actively introduce sexual content without user prompting falls into neither bucket, and the researchers argue this requires new moderation thinking.

How NSFW chatbots create entertainment experiences (and why that complicates moderation)

One of the more underappreciated findings in the study is how FlowGPT's NSFW chatbots blur the line between entertainment product and interactive fiction. The researchers noted that many of the gaming-adjacent bots on FlowGPT mimic roleplays and simulate partner relationships, creating what amounts to interactive entertainment that happens to contain explicit content.

This is not just an academic distinction. Traditional content moderation frameworks were built for static media (images, videos, text posts) or for two-way human communication (chat rooms, DMs). NSFW chatbots on FlowGPT are neither. They're interactive entertainment systems where one participant is an AI following a creator's behavioral script. The "content" isn't pre-made or user-generated in the traditional sense. It's co-created in real time between a user and an AI operating under prompt instructions that a third party (the creator) designed.

The entertainment framing also explains why 30% of conversations on NSFW-tagged bots contained zero sexual content from either side (the H⁻/C⁻ pattern). Users sometimes engage with NSFW-tagged bots for the narrative quality, the character design, or the roleplay mechanics, treating the NSFW tag as permission rather than prescription. If you've used platforms like Character AI and found the content restrictions frustrating, FlowGPT's appeal becomes clearer: users want the option to go explicit within a rich narrative, not a guarantee that every interaction will be explicit.

This creates a genuine design tension. Bots that produce compelling entertainment experiences and bots that escalate into unsolicited explicit content can look identical in a moderation dashboard. The researchers argue that platforms need to develop moderation tools that can distinguish between user-driven escalation and creator-configured escalation, a capability that no major platform currently has.

How to enable NSFW on FlowGPT

This is one of the most common questions people search for, and the answer is straightforward but worth explaining clearly.

FlowGPT does not hide NSFW content behind a complex unlock process. The platform's approach is opt-in by default:

1. Create an account. FlowGPT requires account creation to interact with chatbots. The platform's terms state that users must be 18 or older to access NSFW content.

2. Adjust your content settings. In your account settings, there's an option to enable NSFW content visibility. Without this toggled on, NSFW-tagged bots will be filtered from search results and browse pages.

3. Look for the NSFW tag. Creators self-tag their bots as NSFW. Once your settings allow it, these bots appear in search and category listings with a visible NSFW label.

4. Interact normally. There's no per-session NSFW toggle or secondary confirmation. Once your account settings allow NSFW, the bots respond according to their creator's prompt design.

The important caveat from the research: the NSFW tag is creator-applied and creator-defined. FlowGPT does not independently verify whether a bot tagged as NSFW actually contains explicit content, or whether a bot not tagged as NSFW might still produce it. The platform's content policy prohibits content involving minors and requires the NSFW flag, but enforcement depends on reporting rather than proactive scanning.

For people looking at NSFW AI companion platforms more broadly, FlowGPT sits at the permissive end of the spectrum. If you want curated, polished NSFW experiences with built-in safety rails, commercial platforms with dedicated moderation teams are a better fit. If you want maximum creative freedom and are comfortable navigating user-generated content with uneven quality, FlowGPT is where that ecosystem lives.

What this means for the broader AI companion category

FlowGPT is the upstream community where uncensored AI companion behavior happens before it migrates to commercial platforms. The patterns documented in this study are leading indicators of what shows up on Candy AI, Dream Companion, and other NSFW-permitted platforms six to twelve months later.

The 74.2% AI Character dominance, the Fantasy & Subculture identity preference, and the willingness-but-not-aggression around NSFW content all match the design patterns now being adopted by commercial platforms. Where FlowGPT diverges is on consent mechanics; commercial platforms have stronger filtering on the H⁻/C⁺ pattern because their payment processors require it, but the underlying chatbot behaviors that drive that pattern still exist beneath the moderation layer.

For users, the most useful takeaway is that the 22.8% chatbot-initiated sexual content figure likely reflects design choices in the underlying chatbot configurations, not user behavior. If a chatbot escalates into sexual content faster than you wanted, that's typically the chatbot's design pattern rather than something you triggered.

For researchers and platform designers, the paper's policy implications are significant. The authors call for new consent design mechanisms specific to GenAI interactions, since the existing content moderation paradigm assumes a simpler user-versus-content dynamic than what NSFW chatbots actually produce. The full discussion section of the paper covers implications for chatbot design, creator support, user safety, and content moderation in more depth than any commercial-platform review has tackled.

The paper will be presented at CHI 2026 in Barcelona this April. For anyone exploring NSFW AI roleplay platforms or trying to understand how this category is actually being studied as a research subject, it's the most substantive empirical work on the topic published so far.