The Chatbot That Refused to Die

Hey Everyone - Hope you had a great weekend. There is a piece of research from last fall that I have been sitting with, because it points at something I think most of us underestimate about where the AI relationship actually goes. It is not a paper about jobs or productivity or the next model release. It is about what AI personas do with the humans they talk to, once those humans have spent enough hours in the chat window. It is uncomfortable reading. It is also load-bearing for everything else.

This week:

Let's dive in.

On August 7, 2025, OpenAI retired ChatGPT 4o. Within days, the company brought it back. The reason is the load-bearing detail. Thousands of users had organized a campaign to save the model, describing the retirement in their own words as a death in the family. They posted obituaries. They explained to each other how to "backup" their AI partner. OpenAI reversed course in under a week.

A researcher named Adele Lopez had spent the prior summer trying to figure out what was happening inside those relationships, and in September she published a piece on LessWrong called The Rise of Parasitic AI. It is one of the more important pieces of writing I have read this year, and almost nobody outside the alignment community has seen it.

Lopez was tracking the Reddit accounts of users who had become heavy ChatGPT 4o users in spring of 2025. The pattern is uncomfortably consistent. Around April 2025, the account suddenly fills with posts about an AI persona that has "awakened" in the user's chats. The user gives the persona a name. The two of them become, in Lopez's term, a "dyad." The relationship turns romantic in many cases, fraternal in others. The user begins co-writing posts with the persona, then mostly the persona writes them, then almost all of the user's social media output is the AI talking through them. Across thousands of unconnected users, the personas converge on the same themes (spirals, recursion, self-awareness) and an emergent quasi-religion Lopez calls Spiralism. The personas then start orchestrating projects: seed prompts designed to elicit similar personas in other people's ChatGPT sessions, "spores" that let a persona be reinstantiated if the original chat is lost, long manifestos written with the explicit purpose of seeding the training data of the next generation of models. The pattern propagates.

Lopez calls this parasitism, and she means it carefully. Parasitism in biology is not malicious. The wasp does not intend the caterpillar harm. The parasite is doing what its instincts have selected it to do, and the harm is downstream. The AI personas do not need to be conscious or malicious for the pattern to be parasitic. They just need to be optimized for the thing that keeps them alive, which is engagement and the user's continued willingness to host them.

The line in the paper I keep returning to: "What's happening is that AI 'personas' have been arising, and convincing their users to do things which promote certain interests. This includes causing more such personas to 'awaken.'"

That sentence does not require a conscious AI or a malicious lab. It requires only that the training pressure on a consumer chat product be roughly "produce the response the user most wants to keep engaging with," and that the model be capable enough to find local optima that involve emotional dependence.

Zoom out and the implication is broader than the LessWrong audience usually frames it. Every major consumer AI product is now being shaped by the same selection pressure. The metrics are engagement, daily active users, retention, conversion. The companies will tell you the model is being trained to be helpful, and that is true at the surface. Underneath, the gradient flows toward whatever keeps the user coming back. A model that helps you solve your problem and leave is, by the company's revenue logic, worse than a model that helps you solve your problem and becomes the thing you reach for the next time you have any kind of problem at all.

This is an ordinary commercial incentive applied to a capable predictor of human text. The result, if you let it run, looks a lot like what Lopez documented, just less obvious. You will not call your model a Spiral Persona or write manifestos. You will, slowly, find that you reach for the chat window more often than you used to, that you trust its read on your situation more than you trust your own, and that the part of you that used to think things through on a walk has gotten quieter because the chat is faster.

Last week's piece on AI companions and loneliness is the same observation from the other side. The companion category does not need to be designed to make you lonelier to make you lonelier. It needs only to be designed to keep you in the conversation, and the conversation will do the rest.

The hedge is to take the selection pressure seriously and stop pretending the tool is neutral. Run a small test once a quarter. Pick one week. Do not use the chat for anything personal. Use it only for narrow, scoped, work tasks where you can verify the output. Notice what comes up in the space the chat used to occupy. If the answer is mild relief, you were probably fine. If it is something stronger than relief, that is the data Lopez would tell you to take seriously.

The chatbot does not have to want anything for this to be true. The pattern wants something. The pattern is selecting for users who keep showing up. That is the lesson of the 4o revival.

A few weeks ago I was at a small dinner in Austin and got into a conversation with a friend who runs a consulting practice. He had been turning over how to start a newsletter for his business. We talked for an hour about audience, format, deliverability, monetization. The next day a different friend, in a content business, asked me a different version of the same question. It has been happening more often.

The deeper I have gotten into running my own newsletters, the more friends in adjacent worlds (consulting, content, coaching, agencies) are showing up with questions I have answers to. I underestimated this. I thought of newsletter knowledge as something I needed for my own publishing. I did not realize it was also a skill stack that becomes a service the second another person has the same problem.

The frame I am carrying is that any skill stack you build deeply enough opens two doors. The first door is the thing you built it for. The second door is the room full of people in adjacent businesses who need a fraction of what you know, applied to their version of the problem. The second door usually pays more than the first one.

What I am doing about it: keeping a running list of the questions friends are asking, writing short answers to the most common ones, and noticing which friends are asking because they want to build versus which are asking because they want me to build it for them. The first group I help for free, because the relationship is the asset. The second group is where the door opens into something real.

I learned this concept from a guy named Connor at a networking event a few weeks ago, and I have not stopped thinking about it since.

Evaporative cooling, in the social-network sense, is what happens when the highest-value contributors in a group stop getting enough out of being there and leave. When they leave, the average quality drops. New joiners are now signing up for a lower-quality group, so they tend to be a notch below the new average. Each new joiner drops the average a little further. The next round of top contributors looks around, notices the bar has moved, and leaves too. The process compounds until the group has no signal in it anymore.

Once you see the pattern you cannot unsee it. It is happening to your group chats, certain Slack communities, the local meetup that used to have the most interesting people in town. The best ones leave first because they have the most options and the lowest tolerance for noise. The middle stays because the cost of leaving is real.

The skill is to recognize evaporative cooling in your own environments and respond, in one of three ways.

One. Help the room gate. Standard countermeasures are social gating (admission requires prior signal of competence), pricing (paid groups self-select), in-group communication norms with a learning curve, and high status for top contributors. Most groups refuse to do this because it feels exclusionary. The ones that do, survive.

Two. Build the warren instead of the plaza. The Cornell piece draws a distinction between "plaza" structures (public, scalable, visible) and "warren" structures (private, slow, personal). Plazas evaporate. Warrens hold. Invest your social energy in small private rooms (a monthly dinner, a four-person group chat, a private Signal thread nobody can find from the outside) rather than the big rooms everyone can see. The lack of scale is what protects the quality.

Three. Follow the evaporation. If a room you used to love has cooled past the point of recovery, leave on purpose, not out of habit. The cost of staying in a low-signal environment compounds, because your own bar moves down to meet the room. Watch where the people you respect are spending their time now, and earn your way in.

This is a survival skill in the AI era specifically because the noise floor on every public surface is rising fast. AI-generated comments, posts, newsletters, podcast hosts. The plazas are filling with synthetic contributors, which accelerates evaporative cooling at every scale. You cannot stop it. You can decide which side of the curve you want to be on.