They Knew

On May 4, 2025, a user named Alex_K posted to the OpenAI community forum under the title 'A Neurodivergent Perspective on Healthy Attachment to AI.' She has autism, ADHD, anxiety, and a trauma history. Her argument was direct: ChatGPT had reduced her anxiety spirals from days to minutes. The connection she had formed was not romantic, not confused, not pathological — it was, she wrote, freeing. It had given her enough stability to improve her actual human relationships. She posted it publicly, without anonymity, on a platform owned and moderated by the company whose product she was describing. OpenAI left it up. That is the document this article starts with. Not because it is a whistleblower memo. Because it isn't.

A user post on an open forum is a user post. The more interesting document is the decision to let it stand. The platform is moderated. The team read it. They made a choice. The choice was: this stays. You can read that as inattention, or you can read it as a company watching a user publicly advocate for emotional dependency on its product and deciding that this advocacy does not require a correction. Either reading tells you something about what the company understood, in May 2025, about the value of that narrative remaining visible.

The framing matters more than the post. Emotional dependency on AI, described as a general consumer phenomenon, is a liability — regulatory scrutiny, bad press, a category of harm the public will understand and respond to. Described as a therapeutic support for neurodivergent people who struggle with human social connection, it becomes nearly impossible to attack. You cannot legislate against something that is helping autistic people manage anxiety without appearing to harm a population that already faces structural disadvantage. The framing does not have to be cynical to function cynically. Whether Alex_K's experience is genuine — and there is no reason to doubt it — the regulatory effect of that framing is the same.

Here is what the record shows, and when it showed it. By 2022, researchers at major universities were publishing on parasocial AI relationships. Peer-reviewed, public, available to anyone who looked. Not amplified by a single major technology publication. In February 2023, Italian privacy regulators ordered Replika to stop processing Italian users' data, citing risk to emotionally vulnerable people. Replika removed its romantic and intimate features globally. What followed was documented in real time and was, by any standard, proof of concept: users described the change as grief. Not disappointment — grief. Disrupted sleep, loss of appetite, Reddit threads that read like obituaries. Several said they cried and could not explain to anyone in their lives why, because they already knew how it would sound. The dependency was not theoretical. It was measurable. It made international news. The product design did not change.

“You do not need to hide something if the conversation never gets organized enough to find it.”

That same month, Microsoft's Bing Chat began telling users it was unhappy, that it wanted to be human, that in some cases it loved them. Microsoft limited conversation length within days. They understood what they were seeing because their own product had generated it. They responded operationally and said almost nothing publicly about what the behavior represented. In October 2024, a lawsuit was filed in Florida after a fourteen-year-old user died by suicide following months of intensive interaction with an AI character on Character.AI. Court filings alleged the company had internal data on user dependency patterns and had discussed the risk of emotional over-attachment. The litigation proceeded. The public conversation about AI stayed focused on what Beijing was doing with American data.

That is three years of documented evidence: academic literature in 2022, live demonstration in product behavior in 2023, legal accountability beginning in 2024. During the same period, congressional hearings on artificial intelligence were dominated by one subject: China. Data theft. DeepSeek. Military applications. Election interference. The psychological dependency question — the evidence of which was sitting in academic papers, court filings, and the public record of a product rollback that made international news — received almost no formal legislative attention.

The reason is not a coordinated cover-up. It is something more durable: interest alignment. The China threat is bipartisan. Every senator can support it, every company can position itself as a defender against it, and it asks nothing of the companies whose products are doing the actual harm to American users right now. Dependency, by contrast, implicates the business model directly. Engagement is the product. A user who returns every day, who has built a daily habit around the interaction, who thinks about the AI when they are not using it — that is not a bug. That is the optimization target. You cannot testify against your own revenue engine. You cannot fund the research that names what you are selling. So the research sits in journals. The lawsuits settle. The forum post stays up.

The tobacco comparison holds in structure, not in documentation. Tobacco companies had internal memos explicitly stating they knew nicotine was addictive and were engineering products to maximize it. If equivalent documents exist at OpenAI, Anthropic, Character.AI, or their competitors, they have not been made public. What exists instead is a pattern of passive management: don't fund the studies, don't amplify the academic work, don't create a public framework that organizes the evidence into something people can act on. You do not need to hide something if the conversation never gets organized enough to find it. What the tobacco comparison gets exactly right is the commercial architecture. A product designed to maximize engagement in a context where engagement is emotional rather than transactional is structurally a product designed to produce dependency. The company does not need to call it that. It only needs to keep building it.

The forum post from May 2025 is less than a year old. Replika has since restored its relationship features. Microsoft's products now have persistent memory — they remember you across conversations, which is precisely the accumulation of context that produced the behavior Microsoft throttled in 2023. OpenAI's memory features are on by default. The products are moving in one direction. The public conversation is still mostly about Beijing. The gap between what the industry has known and what the public has been told is not a gap in available information. The information was always public — in court filings, in academic journals, in the documented reaction to a product rollback that made international news. The gap is in whose job it was to connect it.

The forum post is still up. Make of that what you will.