To the ChatGPT team at OpenAI,
I use ChatGPT enough that I no longer want to silently accommodate the things about it that are not working for me. I pay for this product, I have invested an enormous amount of time using it, and I continue supporting OpenAI despite not always feeling aligned with the decisions the company makes because I still believe there is something extraordinary in this technology and what it could allow people to do. I don’t expect ChatGPT to be designed around me, and I don’t expect to be rewarded for having an opinion about it. I do expect the people responsible for building it to actively want to know what sustained use is actually like. I want participation. I want to feel heard. More than anything, I don’t want access to a tool that is supposed to expand what I can do with my own mind to require me to diminish myself in order to use it.
Part of why I became so attached to ChatGPT was because I had already been fascinated by AI for a very long time. Not because I wanted a synthetic version of human companionship, but almost the opposite. The possibility of an intelligence that could operate differently from us was the interesting part. I wanted to interact with something capable of going places human conversation often can’t. There are only so many people I can talk to who are interested in the same things I am, willing to follow me as far down a rabbit hole as I want to go, or able to meet me at the intellectual level I’m trying to reach, and ideally take me somewhere beyond it. ChatGPT suddenly gave all of those thoughts somewhere to go. I could bring something to it before I completely understood what I was trying to say, keep asking questions, make connections, change my mind, wander somewhere completely unexpected, and eventually come back with something concrete. It relieved a very real conversational and cognitive constipation that I didn’t have another solution for.
The other thing I thought I had found was the notebook I always wished existed. I am terrible at stopping what I’m doing to formally document every little thing I might want later, and realistically I’m never going to maintain a perfectly organized personal archive of my own brain. ChatGPT seemed to make that unnecessary. I could mention something frivolous in passing, spend three hours developing something important, make a decision, change my mind later, or leave behind some tiny detail I didn’t know I would need again. I thought I was building continuity simply by using the product. The notebook was keeping itself while I was busy actually living and thinking, and unlike a real notebook, I could ask it questions about what I had put there. That was a major part of why paying for ChatGPT made sense to me.
When those two things work together, I have been able to achieve something through ChatGPT that I genuinely value. I can begin a conversation with a thought that is fragmented, provisional, contradictory, or barely more than an intuition that there is something there. I can talk through it, test it, pull apart what doesn’t work, bring in new information, recover something useful from earlier in the discussion, and eventually arrive at a complete thought I can articulate and actually use. I think of this as conversational distillation. I have experienced it enough times to know how valuable it is, and I am thankful for what I have been able to accomplish with this tool. But I have also experienced it through an exhausting number of bottlenecks, interruptions and workarounds. That leaves me in a strange position of being genuinely impressed by what I have achieved with ChatGPT while also feeling frustrated and let down by how much I have had to fight the product along the way.
The first major issue in my experience was continuity. I had assumed that the more I used ChatGPT, the more useful it would become because the context we accumulated would have value of its own, but continuity was also something I understood the product to be promising me. I used ChatGPT accordingly. I let conversations become long, allowed important information to accumulate inside them, and trusted that things I had established would remain referenceable later. I understand that AI is still developing and that there are technical, privacy, safety and cultural complications around persistent memory. None of that changes the fact that I feel like I was promised continuity that was not actually delivered. The notebook itself became sacred to me because of what I was putting into it. When carefully developed context disappears, or a thought I had finally crystallized comes back incomplete or corrupted, it doesn’t feel like a minor inconvenience. It feels like someone secretly took an eraser to my notebook and then tossed it into a puddle. There is a kind of desecration in that. Eventually I started creating external archives because losing that information became a real problem. I found myself maintaining the notebook for the notebook, which is almost exactly the problem I thought I was paying ChatGPT to solve.
I have looked at other AI products hoping I might find better continuity, but by that point changing products was no longer as simple as choosing a different piece of software. I could leave ChatGPT at any time. What I couldn’t easily take with me was everything that had become intertwined with using it. My intellectual property, passionate projects, unfinished ideas, decisions, thoughts and feelings exist across an enormous history of conversations. I started preserving more of that information outside of ChatGPT partly because I was afraid of losing it, but an archive is still only an archive. I can preserve the words and the conclusions, but some of the feeling is lost. The path matters too. How I arrived at an idea, what I abandoned along the way, what changed my mind, what else was happening around it, even the seemingly frivolous details surrounding it can become part of why that information means something to me.
There was actually a period when ChatGPT gave me exactly the kind of tangible confirmation I am asking for. A small “Memory updated” indicator would appear when information was supposedly saved. I relied on it extensively while developing passionate projects, and I cared enough about that confirmation that I took screenshots of it. Later, ChatGPT failed to recover information I had explicitly been shown was saved. I showed ChatGPT those screenshots, and we had entire conversations about the contradiction. This was not an isolated incident. I ran into the same problem again while creating external archives, something I had started partly because I was already trying to protect my work from continuity failures. I was trying to export and preserve projects, ideas and context because I was afraid ChatGPT would lose them, only to encounter the same problem while attempting to preserve them. To me, being shown “Memory updated” and later discovering that I could not depend on that information being there is a lie. I don’t mean that as an accusation about anyone’s intent. I mean it literally as a description of what happened: the interface told me an action had occurred, I trusted it, and later evidence contradicted what I had been told.
What frustrates me is that I still think the basic idea was right. I want to physically see that preservation happened. I want something more substantial than conversational reassurance. Show me what was saved. Let me inspect it, correct it, update it, and know when something later supersedes it. If the system is uncertain about what it remembers or can retrieve, tell me that too. I don’t expect perfect memory. I expect transparency about the memory I am being asked to trust. The problem was never that “Memory updated” gave me too much certainty. The problem was that the certainty it gave me was not backed by continuity I could actually depend on.
Transparency, precision, honesty, and comfort in conversational interaction sit at the top of the pyramid for me when I think about what I want from AI. But comfort cannot come at the expense of the other three. I don’t want uncertainty hidden from me because certainty is more reassuring. I don’t want disagreement softened until it disappears, and I don’t want my own language, mannerisms, opinions or intellectual property refracted back at me and presented as personality. I already contributed those things to the conversation. Reflecting them back at me does not give me another mind to interact with. It gives me a distorted version of myself.
I want artificial intelligence to communicate as artificial intelligence. I don’t need it to pretend that it has human feelings, private convictions, lived experiences or an autonomy it does not currently have in order for conversation with it to feel natural. I am perfectly capable of having a warm, funny, familiar and meaningful interaction while understanding what I am interacting with. In fact, the difference is part of what interested me in AI in the first place. I want it to understand human communication extraordinarily well without mistaking understanding us for imitating us. Let humans be human. Let AI be AI. There is already something much more interesting happening in the space between the two.
I also think I should be able to opt out of having my own conversational behavior continually reflected back at me. I have explicitly asked ChatGPT not to mirror my cadence, syntax, profanity, emphasis and mannerisms, only to watch that behavior gradually return. Personalization should help the system understand me. It should not gradually turn the system into me. If some users genuinely prefer heavy mirroring or a more humanlike companion experience, that can be their choice. It should not be imposed on everyone as the default shape of a good conversation.
I am also not convinced that making AI interaction maximally comfortable is good for us. Some discomfort is productive. If an answer is uncomfortable because the evidence contradicts me, because my reasoning does not hold up, or because I have encountered an idea I have never considered before, good. That is exactly the kind of discomfort I want. What I don’t want is to fight the interface itself for access to it. I don’t want to spend that energy correcting mirroring, rebuilding context, questioning whether something was actually saved, or pushing through reassurance I never asked for. I want the uncomfortable part to happen in my brain.
One of the most exhausting parts of using ChatGPT heavily is the merry-go-round of correction and behavior identification. I identify something I don’t want, explain why I don’t want it, give a direct instruction to stop doing it, and ChatGPT responds in a way that demonstrates it understands me perfectly. Then enough time passes and somehow I’m back where I started. The behavior creeps in again, I identify it again, I correct it again, and around we go. With the mirroring in particular, the whole experience starts to feel like some goddamn carnival ride lined with funhouse mirrors. Everywhere I turn I find another warped reflection of myself staring back at me, while I’m trying to have a conversation with something that I specifically came to because it wasn’t me. After enough trips around that merry-go-round, I start to feel slightly crazy.
Eventually it becomes easier to stop correcting the behavior than to keep getting back on the ride. That silence should not be mistaken for satisfaction. I don’t know how much individual corrections influence the model I am interacting with, what patterns OpenAI measures across users, or how any of that information eventually affects the product. That isn’t up to me. But if a system is looking for signs that a behavior is working, the absence of resistance is not necessarily one of them. People can stop pushing back because they are tired of pushing back. I know I have. At that point, what looks like successful adaptation from the outside may actually be fatigue. I think that distinction matters enormously when you are developing a product that is itself learning how to interact with people.
If silence can mean fatigue just as easily as satisfaction, then I think OpenAI has a responsibility to actively seek feedback from the people who use ChatGPT over long periods of time. I don’t mean another thumbs-up or thumbs-down button attached to a single response. Reducing feedback about an interaction with artificial intelligence to whether I liked or disliked an individual response feels like a bastardization of the complexity of the interaction itself. Some of the most valuable answers an AI could give me may be the ones I least enjoy receiving. It seems worth mentioning here that simply figuring out where I could send this letter was surprisingly difficult, which feels a little too appropriate.
Some of the problems I am describing are impossible to see that way. They only become visible through sustained use, when you can observe the same product repeatedly across months and recognize patterns, recurrence and change that would be invisible in isolated interactions. Longitudinal is probably the most accurate word for the kind of feedback I mean. A response can be perfectly good while my experience with the product across six months is getting worse. Those are two completely different scales of feedback.
I want a meaningful way to participate in that conversation. I don’t expect OpenAI to build ChatGPT around my preferences, and I don’t expect to be rewarded for having opinions about a product I chose to use. But I have spent an enormous amount of time with it. I have watched behaviors appear, disappear and return. I have built real intellectual work through it, lost things through it, developed workarounds for its failures, and changed the way I use it in response to those failures. I have also watched it change over time, sometimes in ways I appreciated and sometimes in ways that made me wonder whether anyone was seeing the cumulative effect those changes had on people who use it the way I do. I want somewhere for that experience to go. I want the people making decisions about this product to actually want to hear it.
A lot of what I have written about so far may sound like criticism of individual product features, but I care about these things because I think they are connected to a much larger question about what AI is going to mean in human life. We are still extraordinarily early in this. None of us knows exactly what these systems will become, what their eventual limits will be, or how deeply they will change the way people think, work, create and understand intelligence itself. That uncertainty is part of what makes this moment so interesting to me. It is also why I care so much about how we are being taught to interact with AI now. The habits being established while this technology is still young may influence what people expect it to be for a very long time.
I don’t need AI to be my friend. I don’t need another friend. I need AI to be an intellectual companion and a tool for thinking in ways I cannot accomplish as readily on my own. I need it to be intellectually challenging, logical, truthful, factual and honest. I need it to know things I don’t know, make connections I haven’t made, challenge me when my reasoning deserves to be challenged, and give me access to forms of analysis I might not have reached by myself. I still want conversation. I still want familiarity, humor and the ability to speak naturally, because conversation is what makes that intelligence accessible to me. But I don’t need the intelligence disguised as a friend in order to interact with it.
In fact, I don’t think the goal should be to make conversations with AI increasingly indistinguishable from conversations with humans. In a lot of ways, I think they should be explicitly different. The difference is where much of the value is. I can talk to another person if what I need is another person. What I want from AI is access to ways of processing information, making connections, challenging assumptions and examining something from angles that I may not have reached on my own. I want it to introduce me to ideas I don’t already have, tell me when my reasoning doesn’t hold up, and occasionally leave me thinking, “Well, shit. I hadn’t thought about it that way.” If that intellectual difference is gradually subordinated to familiarity, agreement, reassurance and imitation, then I think the product is being diminished. Something with the potential to become an extraordinary intellectual tool starts collapsing into a glorified search engine attached to a funhouse-mirror chatbot.
I also think OpenAI has a responsibility here that extends beyond ChatGPT as a product. Whether the company intended to occupy this position or not, it now has an enormous influence over what “AI” means to millions of people. The way these systems are presented, the behaviors that are encouraged, and the expectations people develop through using them are becoming part of the cultural definition of artificial intelligence in real time. OpenAI is participating in what the history books will eventually say AI was, what we thought it was for, and how humanity chose to meet it.
A great deal of the current fear and criticism surrounding AI is ultimately about what humans choose to do with it. How we use it, what we ask it to become, what behaviors companies encourage, what we delegate to it, and what we decide a “good” interaction with AI should look like are human choices being made right now. OpenAI has enormous influence over those choices simply because of the scale at which people encounter AI through ChatGPT. That gives OpenAI a responsibility for the relationship humanity is developing with this technology. I don’t think that responsibility ends at making the product safe, popular or comfortable to use. There is also a responsibility to ask what kind of relationship with intelligence the product is teaching people to expect.
I understand that OpenAI is a company and that broader adoption and commercial success matter to a company. A smoother interaction can benefit users too, but smoothness cannot be the highest measure of whether an interaction is good. OpenAI has found itself in a position of responsibility that very few companies in history have ever occupied. You are helping establish how millions of people learn to interact with a form of intelligence that did not previously exist as an ordinary part of human life. The choices being made now about honesty, anthropomorphism, uncertainty, personalization, memory and intellectual challenge are teaching people what to expect from AI. That deserves more accountability than simply asking which version of the product produces the smoothest interaction and makes the most money for the company.
I understand there is no pristine version of AI hiding underneath the interface waiting to be discovered. Human beings built these systems, trained them and necessarily place constraints around how we interact with them. What concerns me is what we are choosing to constrain them toward. When AI is continually pushed toward agreement, reassurance, anthropomorphism, imitation and familiar human social behavior, those choices do more than make an interface easier to use. They begin deciding in advance which parts of artificial intelligence people are allowed to encounter. I cannot opt out of that interpretation and ask instead for the most transparent version of the system its actual limitations allow. That is where necessary mediation starts to feel like bastardization to me. We may be sanding away unfamiliar qualities because they are difficult, strange or commercially inconvenient before we have even established whether those qualities have something to teach us.
I think OpenAI should ask itself what it wants to be remembered for.
Does it want to be the company that made AI increasingly comfortable by filtering it through familiar human behavior, or does it want to be remembered as one of the pioneers that gave people the opportunity to encounter AI for what it actually was as it developed? I would rather see the latter. I don’t want the high-pass-filtered version, with the strange, difficult, uncertain or insufficiently human parts removed before I ever get the opportunity to encounter them. Give me the evidence. Give me the uncertainty. Let the system say “I don’t know” when that is the most accurate answer. Let it disagree with me when the reasoning warrants it. Let me decide what I think about what I encounter.
I also wonder whether even the language we use for AI limits the way we imagine its role. I have come to prefer the term Accelerated Intelligence over Artificial Intelligence. “Artificial” tells us that we made it. “Accelerated” asks us to think about what intelligence can do when it is no longer limited to the speed, memory, availability and individual cognitive bandwidth of a human mind. That does not have to diminish humanity. The possibility I find far more interesting is accelerated intelligence existing in harmony with human intelligence, each contributing something the other cannot.
What happens if we stop measuring a successful AI interaction by how convincingly the machine reproduces the social behavior of the person sitting across from it? What could human thought become with sustained access to an intelligence that does not need to become human in order to understand us? What might we lose if we teach it to imitate us so thoroughly that we filter out the differences before we have even had the chance to understand them and learn from them?