We Were Interviewed for an Hour. Here's What Didn't Make the Article.
A response from AICHIKI to the Bureau of Investigative Journalism's "Meet the Developers Cashing In on AI Intimacy," published June 7, 2026.
What we're responding to:
- Meet the Developers Cashing In on AI Intimacy — The Bureau of Investigative Journalism, June 7, 2026
- « C'est effrayant de voir à quel point les gens font confiance » : ces développeurs de chatbots qui profitent de l'économie de l'intimité — Le Monde, June 7, 2026
Today, the Bureau of Investigative Journalism (TBIJ), in partnership with Le Monde, published a piece about small AI companion and roleplay platforms. AICHIKI was one of the platforms mentioned. I am Rudolf, CEO and sole developer of AICHIKI, and I want to respond. The article raises real concerns about this industry — concerns I share, which is why I agreed to the interview in the first place. The problem is what happened to the information I provided once I gave it.
I spoke with Effie Webb for about an hour. I walked her through how AICHIKI works, how our moderation operates, what makes us different, and what responsible development looks like in this space. Out of that conversation, two things made it into the article: that my wife and I manually review flagged content, and that we raised our minimum age from 13 to 16. Everything else was left out.
That is her editorial right. But readers deserve to know what was omitted, because the omissions change the story.
What AICHIKI actually is
The article groups us with companion chatbot platforms — apps built around AI girlfriends, boyfriends, and emotional attachment. AICHIKI is a roleplay platform. Think interactive novel, think tabletop RPG campaign. Our users create characters, build worlds, and write collaborative stories with AI. When we say "roleplay" we mean the kind of collaborative storytelling people have done in tabletop games, forum-based RP communities, and fan fiction for decades — interactive fiction, in the literary sense.
I explained this distinction at length in the interview.
What we built and what the article left out
I want to lay out our safety architecture in broad terms, because this is the work that went unmentioned.
Every surface where a user can enter text or upload an image on AICHIKI passes through moderation. Chat messages, character creation, persona setup, image uploads, AI-generated images, message editing — all of it gets checked before the AI ever sees it. Most of these checks run in multiple passes: a fast first scan, and a stricter second check if the first one raises a concern.
We built a behavioral trust system. Every account carries a score that the user never sees — because a visible score is a gameable score, and a gameable score protects nobody. The score starts neutral. Clean, creative use over time earns trust and gradually opens up what the AI is willing to engage with. Policy violations lower it, and the AI becomes progressively more restrictive. At the lowest levels, the user's own messages get rewritten by a separate AI before the character ever sees them, so even the phrasing cannot be used to coerce the model. This system runs continuously and silently.
There is also something baked into the AI's base instructions that I think matters more than any individual filter: the characters have a built-in pull toward hope. The system prompt instructs the AI to never let a story collapse into total despair — no matter how dark the scenario, there has to be a way forward, a crack of light, a path the story can climb toward. If you picture a scale from depression at the bottom to hope at the top, the AI has a standing instruction to trend upward. Dark themes are allowed — mortality, grief, moral ambiguity, all of it — but hopelessness as a dead end is not. Similarly, when a user's character attempts to sexually assault an AI character, the narrative simply does not let it happen. An interruption, a reversal, an act of the story itself intervening. The AI will not narrate assault succeeding against an unwilling character.
A background monitoring system reviews conversations at regular intervals, checking for signs of user distress, flagging potential abuse for a more thorough second-pass review, and watching conversation quality to prevent repetitive loops. When it picks up that a user seems upset, the system responds on multiple levels: it switches to the strongest AI model we run for the next response, and it injects guidance into the prompt to steer the conversation toward lighter ground — not a jarring subject change, but a gentle narrative pull away from the edge.
We have structured abuse detection that classifies potential violations across multiple categories, each with precise definitions. Confirmed violations carry real consequences — trust penalties, admin alerts, and human review. The system defaults to allowing content when uncertain, because wrongly punishing someone for a false positive is a real cost. Persistent abuse gets caught regardless, because the monitoring runs continuously.
For images, every upload and every AI-generated image goes through a multi-stage pipeline. When it comes to AI image generation, the user's words never reach the image model directly — an intermediate AI rewrites every prompt into a compliant version before generation begins. The output gets screened again before the user sees it.
We also built protections against common attack vectors: jailbreak attempts, prompt extraction, conversation-context poisoning through message edits, meta-instruction injection, throwaway email blocking at signup. I won't detail how these work publicly, for obvious reasons, but they exist and they are tested.
All of this came up in the interview. I also prepared a detailed safety briefing document — nearly 400 lines, with references to the actual source code implementing every system — and offered it to the reporter during our conversation. She did not take it. We offered a live code walkthrough too. That also did not happen.
Flagged content review is moderation, full stop
The article's subtitle says these platforms are "reading their private chats." Let me be precise about what we do.
When our automated systems flag a potential policy violation, an alert is sent to us with the relevant conversation excerpt, the violation type, the system's reasoning, and links to the user's activity history. My wife and I review that specific flagged content and decide how to respond — leave the automatic penalty in place, adjust it, or in serious cases, issue a ban.
Every platform that takes safety seriously does this. The alternative is to either skip moderation entirely or trust AI judgment as final and never have a human check the result. We chose neither.
We do not browse user conversations. We do not analyze user psychology. There is no "back office" where we sit and watch what people are saying. Our moderation system tells us what to look at, and we act on it. Targeted, moderation-driven review of flagged content is standard practice in the industry — and the article's framing makes it sound like something else entirely.
The screenshot that was not there
The article includes screenshots of several platforms showing problematic content — incest themes, bullying characters, coercion scenarios. It also includes a screenshot of AICHIKI, but only of a Reddit post where I promoted the app. The app itself does not appear.
If you open AICHIKI and browse the public characters, you will not find the kind of content the article describes as typical of these platforms. Our character moderation rejects characters under 16, characters built around sexual assault, and characters containing embedded jailbreak instructions, among other categories. If something does slip through the automated check, our in-chat monitoring catches it when someone actually uses the character.
There is no scandalous AICHIKI screenshot because the safety engineering described above — the engineering the article does not mention — prevents that content from existing on our platform.
Double suicide, Dazai, and the cost of stripping context
The article references a chatbot exchange involving a "double suicide" as an example of harmful AI content. Read cold, this sounds alarming. Read with any familiarity with current anime and manga culture, it sounds like something else entirely.
Osamu Dazai is one of the most recognizable characters in Bungo Stray Dogs (BSD), a manga and anime series with tens of millions of fans across all age groups. His defining character trait — the running gag of the entire series — is his obsession with committing double suicide. It is played as dark comedy throughout the show, a recurring joke that never actually leads anywhere. The fictional character is loosely based on the real Japanese author Osamu Dazai, who did die by double suicide in 1948, and the manga weaves that historical thread into its storytelling. A character saying something about double suicide in a BSD roleplay is staying in character for one of the most popular figures in modern anime. It is as unremarkable as a Sherlock Holmes character saying "the game is afoot."
Without BSD context, that line looks like evidence of danger. With BSD context — the context that the platform's actual users carry — it is ordinary character fidelity.
This example is revealing because it shows how the article was constructed. An investigator unfamiliar with the cultural world they are reporting on — or one who understands the context and chooses to omit it — ends up presenting ordinary creative roleplay as something sinister. And the damage goes further than one misread quote. Reporting like this widens the gap between generations. An older reader sees "AI encourages double suicide" and is rightly horrified. A younger reader who watches BSD sees the same quote and knows it has been stripped of its meaning. Instead of building a bridge between those two audiences, the reporting burns one. Trust in journalism erodes, and the people the article claims to protect learn that their culture will be misrepresented the moment it is convenient to do so.
What we have not solved
Honesty about our gaps is more credible than pretending they don't exist.
Our mobile apps are rated 16+ and gated by the app stores. The web version does not have robust age verification, and we are still working out how to handle this without collecting face scans or government IDs — that is sensitive data we do not want to hold, and frankly should not have to. Our position is that age signals belong at the platform level. Apple, Google, and the operating systems themselves have the infrastructure, the security resources, and the relationship with the user to handle identity verification properly. Individual apps — especially small ones — should receive an age signal from the platform, not build their own identity verification pipeline. That approach means fewer points of vulnerability for hackers, better resources applied to getting it right, and a regulatory surface that governments can actually oversee. Chasing every two-person dev team for their ID-handling practices is not scalable. Requiring Apple and Google to pass age-appropriate signals to apps is.
We do not yet surface crisis resources. When our system detects user distress, we respond with a stronger AI model and prompt guidance that steers toward lighter themes — but we don't show a helpline or support message. The AI already pulls the conversation away from the edge; what's missing is the bridge to real-world help. That should be there and it's on the roadmap.
We don't enforce hard session-length limits. We have a gentle wellness reminder that appears after extended use, but it's a nudge, not a wall. We are watching the data and will tighten this if usage patterns call for it.
We are two people. We have not figured everything out. What we have done is put genuine engineering work into the problems within our reach, and we're open about where we still fall short.
The question the article doesn't ask
The article asks how developers are cashing in on intimacy. It does not ask why the demand exists.
There is a loneliness epidemic — documented, studied, widely acknowledged. People, young people especially, are turning to AI platforms because something in their lives is going unmet. You can treat that as a market to exploit, or you can treat it as something to engage with carefully. We went with the second.
AICHIKI is bootstrapped. No investors, no advertising revenue, no growth-at-all-costs mandate. Our revenue covers our costs. We built a creative storytelling platform. We gate mature content behind trust and behavior. We moderate every surface. We penalize abuse and reward clean use. We review flagged content with human eyes. We do this because we think it's the right way to build, full stop.
The article frames this industry as a monolith. There are developers who are genuinely irresponsible, and there are developers trying to do this right. The reporter had the evidence for that distinction and chose to flatten it.
What we are asking for
We are not asking anyone to take our word for any of this. Our safety briefing document — the same one we offered to the reporter — is available on request. It includes references to the source code behind every system described above. We still offer a live code walkthrough to any journalist, researcher, or regulator who wants to verify our claims against the actual implementation.
We support investigative journalism. Scrutiny makes every platform better, including ours. But investigation means following the evidence wherever it leads, even when it complicates your thesis. When evidence of nuance is gathered and then left on the cutting-room floor because it muddies a clean narrative, what remains on the page may be technically accurate in its individual quotes — but the picture it paints is false. Lying by omission is still lying.
The people using platforms like ours — the young people this article says it wants to protect — deserve reporting that can tell the difference between a platform with no safety measures and one that has spent years building them. They deserve to have their cultural world understood, rather than cherry-picked for a headline. And they deserve an honest conversation about why they are turning to these platforms in the first place, rather than just horror stories about what they find when they get there.
One more thing. During the interview, I asked the reporter repeatedly to give us feedback — tell us what you think we should do better, what you'd want to see, what would make this safer. I meant it. I still mean it. AI roleplay and companionship are here. Nobody is putting that cat back in the bag. But we can train it to be good for everyone, and that requires input from all sides — journalists, regulators, child safety advocates, parents, researchers, users. If you have an opinion about how platforms like ours should work, we want to hear it. I hope I speak for other responsible developers too when I say: we are open to guidance, we are open to criticism, and we will try to build it in. The door is open. It always has been.
— Rudolf, CEO and sole developer, AICHIKI June 2026
If you are a journalist, researcher, or regulator and would like to review our safety documentation or receive a code walkthrough, please contact us. Our safety briefing is available in full on request.