When You Flatten the Spectrum, You Kill the Incentive
A broader reflection on the state of AI roleplay, investigative journalism, and why getting this wrong hurts the people we're all trying to protect.
Background reading:
- Our point-by-point response: We Were Interviewed for an Hour. Here's What Didn't Make the Article.
- The article that prompted it: Meet the Developers Cashing In on AI Intimacy — The Bureau of Investigative Journalism, June 7, 2026
- The French edition: « 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 we published a response to the Bureau of Investigative Journalism's article about AI companion and roleplay platforms. That piece was specific — it detailed what AICHIKI actually built and what the article left out. This one is about the bigger picture, because the problems with that article go well beyond us.
The article's broader thesis — that small AI platforms are popping up with weak safety measures and exploiting lonely users — is partially true. There are developers in this space who genuinely have no moderation, no age gates, and no intention of building them. Some said as much on camera. Report on that. Please.
The problem is that the investigation found a spectrum and published a monolith. Platforms with zero safety measures were placed next to platforms with multi-layered moderation infrastructure, and the article treated them as the same story. Evidence of difference was gathered and discarded because it complicated a clean narrative.
You can build a deeply misleading story entirely out of true quotes if you control which quotes appear and how they are framed. Every individual quote can be accurate while the overall picture they paint is false. And that is what happens when you write an article backwards — start with your conclusion, select evidence that illustrates it, and quietly shelve anything that points the other way.
There is a famous precedent for this kind of methodology in science. In the 1950s, the American physiologist Ancel Keys set out to prove that dietary fat caused heart disease. He gathered data from 22 countries. When the results came in, some countries did not support his hypothesis — their populations ate plenty of fat and had low rates of heart disease. So Keys removed those countries from the dataset and published his Seven Countries Study using only the data that confirmed what he already believed. That study shaped global nutrition policy for decades. Governments pushed low-fat diets based on it. The influence was enormous. And the whole thing was built on cherry-picked evidence — real data, selectively presented to support a conclusion that existed before the research began.
The TBIJ article follows the same methodology. The reporter gathered information from multiple platforms. Some of that information showed developers investing serious engineering effort into safety. That information was set aside. What remained told a clean, alarming story — technically built on real quotes, minus the data points that would have complicated things. Keys looked at countries that didn't fit his thesis and deleted them from the study. This article looked at safety architectures that didn't fit its thesis and deleted them from the story.
Flattening the spectrum has consequences beyond the developers who get misrepresented.
A developer in this space right now has two options. Spend months building trust systems, content moderation, abuse detection, image safety pipelines — real engineering work that costs real time and real money. Or skip all of that and launch a bare platform with an API key and a payment page.
If responsible developers and irresponsible ones receive the same press treatment — same article, same framing, same implicit accusation — the message to every new developer entering this space is clear: safety work does not pay. You will be portrayed the same way regardless. The rational move, for anyone without a strong personal conscience, is to skip the investment entirely.
And that is the real damage reporting like this does. Treating every platform the same actively disincentivizes the behavior the article claims to want. An incentive structure that depends entirely on individual developers having a conscience is fragile. It should reward doing the right thing, and reporting that cannot tell the difference between effort and negligence pushes in the opposite direction.
The loneliness question hangs over all of this, and almost nobody in the public conversation is engaging with it honestly.
Young people are turning to AI companionship and roleplay platforms because something in their lives is going unmet — connection, creative expression, a sense of being heard. The demand is real, and it is deep. Ban every AI platform tomorrow and that need does not disappear. It goes unmet, or it finds darker outlets with even less oversight.
The TBIJ article treats the demand as a given and focuses entirely on the supply side — who is building these platforms and how irresponsible they are. That framing lets the conversation feel productive without ever touching the harder question: why are so many people, especially young people, reaching for AI in the first place? If you are serious about protecting people, you have to engage with what is driving them there. Horror stories about what they find when they arrive are no substitute for that conversation.
The regulatory landscape is not helping either. The laws and frameworks being applied to conversational AI were designed for social media — content feeds, recommendation algorithms, follower dynamics, viral sharing. A chatbot conversation is structurally different from a TikTok feed. The harms work differently, the mechanisms work differently, and the interventions need to work differently too. But regulators reach for the tools they already have, and those tools were built for a different problem.
Treating AI platforms as products with product safety requirements — baseline standards, graduated risk profiles, room for different approaches to meeting them — is closer to right than most of what is currently being proposed. Product safety frameworks let you set real minimums while acknowledging that a two-person team and a billion-dollar corporation will meet them differently. Blanket bans and one-size-fits-all rules push responsible developers out and do almost nothing to stop the irresponsible ones, who will simply move jurisdictions or ignore enforcement.
What worries me most, though, is the cultural gap.
The people making policy decisions about these platforms — regulators, legislators, editorial boards — largely do not use them and do not understand the culture around them. They are relying on reporting that, as our experience shows, strips cultural context away for clarity and impact.
The TBIJ article surfaces a chatbot exchange about a "double suicide" as evidence of harmful content. To anyone unfamiliar with anime, that sounds like an AI encouraging self-harm. To tens of millions of fans of Bungo Stray Dogs, it is immediately recognizable as Osamu Dazai's signature character trait — a dark-comedy running gag that defines one of the most popular characters in modern manga. In context, it is about as alarming as a Sherlock Holmes bot saying "elementary."
An older reader sees "AI encourages double suicide" and is horrified, understandably. A younger reader who watches BSD sees the same quote and knows the meaning has been stripped out. The older reader walks away misinformed. The younger reader walks away having learned that the people writing about their world do not understand it — and that lesson generalizes fast. If they got Dazai wrong, what else did they get wrong? Why would I trust anything else in this article? Why would I trust the institutions behind it?
That erosion of trust is happening quietly and it is dangerous. A generation watching their culture be consistently misrepresented by the institutions claiming to protect them does not become more cautious. They stop listening. And when they stop listening, the next real warning — about a platform that genuinely is dangerous — gets ignored along with everything else.
Cultural context is a prerequisite for being taken seriously by the people you say you are protecting, not an optional layer of polish. Without it, you are speaking only to an audience that already agrees with you. The people who most need to hear the message have already tuned out.
AI roleplay and companionship are here. Nobody is putting that cat back in the bag. The question is whether we build a thoughtful framework around it — one that distinguishes between effort and negligence, engages honestly with why the demand exists, and respects the cultural world it is trying to regulate — or whether we keep writing the same alarmist story on repeat while the actual problems go unaddressed.
We have been trying to do this right. We would like the conversation around us to try, too.
— Rudolf, AICHIKI June 2026