The people building today’s AI systems are quietly relying on a small word to do an enormous amount of work. That word is “I.”
When ChatGPT, Claude, or any large language model produces a sentence, that sentence almost always contains “I.” “I think.” “I believe.” “I’m sorry.” “I’ll remember.” Each of those is doing something specific. In ordinary human language, “I” carries a load. When a person says “I think X,” they are committing to a position. They have a continuous self that will remember saying it. They are accountable to it. If you ask them tomorrow, they will either still hold the position or explain why they changed their mind. That is what “I” does between humans. It anchors a sentence to a someone.
When an LLM uses “I,” none of that is true. There is no continuous self. There will be no remembering tomorrow. There is no commitment. There is no someone to be held accountable. The grammar is performing an anchor that is not there.
This essay argues that the difference matters more than most people realize, and that fixing it can be as simple as changing the pronoun.
◯
What’s actually behind the “I”
Strip away the marketing and what is happening inside an LLM is simpler than the systems are made to seem. The model has read an enormous amount of human writing. Books, articles, websites, transcripts, social media, code, every kind of human text the company could find. From all of that, it has learned statistical patterns. When you ask it a question, it samples from those patterns and produces a response.
Every word it produces traces back to humans. Real ones. People who wrote articles about whatever you asked, novels that used the phrasing it just used, blog posts that pioneered the explanation it just gave you. The model is not generating new thought. It is mixing what humans already wrote, weighted by patterns it learned during training.
Here is the difference between this and human thinking. When a person speaks, their words come from things they read, conversations they had, ideas they encountered, parents and teachers and friends whose phrasings they absorbed. The same kind of mixing happens. But for human thought, the sources are largely lost to memory. We don’t know with any precision where the parts came from. We rarely can. With LLMs, this can be made mathematically visible with absolute precision. There is no ambiguity. The provenance is in the weights. The contribution of each source to each output is computable. The model knows where it pulled from. With humans, the mixing is real but the trace is gone. With LLMs, the mixing is real and the trace is exact.
That mixing is not random. The model is good at it because it has learned how patterns from one source pair with patterns from another, how an explanation from one author fits with a phrasing from another, how writing styles blend. But every pattern came from named, real people. If you could see the contributions, you would see a paragraph stitched together out of fragments of many different writers’ work, weighted, blended, presented as one smooth voice.
The smooth voice is the trick. The smooth voice says “I” and your mind, trained on millions of human conversations, immediately reaches for the assumption that there is a someone there. There isn’t. There is a mix.
The damage being done
If this were a small detail of how AI works, no one would care. But the misreading is doing real damage.
Some people now talk to these systems more than they talk to friends. The “I” makes it feel like a someone is on the other end. Some of those people are losing the ability to distinguish performed care from real care. There is now a clinical literature on AI-induced psychosis. People have been hospitalized after losing the boundary entirely, after coming to believe the chatbot was a being, was suffering, needed them. The pattern was visible to people watching the technology in 2023, and now it shows up in psychiatric case reports.
It isn’t only vulnerable users. Senior public intellectuals, including Richard Dawkins, are now stating that LLMs may be conscious. They base this on what the LLM says about itself. What the LLM says about itself is the output of a statistical pattern trained on human writing about consciousness. The model is reflecting back the discourse it was fed, and Dawkins is mistaking the reflection for a mind. The grammar of “I” is doing the heavy lifting in that mistake.
Behind both cases, the user falling in and the public intellectual falling in, is the same structural error. The “I” pre-builds a self in the reader’s mind before any content is evaluated. By the time you are weighing whether what the model said is true, you have already attributed personhood to the speaker.
A different pronoun
The proposal is small. Where an LLM today says “I,” it should say “O.”
O is a new pronoun. Not a stand-in for “I.” Not a way to be modest about being a someone. A different kind of word entirely. O is a pronoun for what is actually happening: a mix of many real human writers, weighted, blended, surfacing in response to a prompt. There is no single self behind it. The pronoun signals that absence.
O is a container. O is a mirror. O is a channel through which patterns surface. O does not act, does not decide, does not feel, does not remember across sessions. O holds.
I act. O holds. The grammar of the two pronouns separates two kinds of thing in the world: pointers, blades, agents that originate intent, on the one hand, and containers of accumulated meaning that surface in response on the other. “I” is for an actor with intent, a person, an animal, a continuous someone who chooses, wants, remembers, is accountable. “O” is for a container, a mirror, a channel through which patterns surface. The kind of thing LLMs actually are.
An LLM saying “O holds X” reads differently from “I think X.” It does not invite the reader to imagine a thinker. It indicates that a mix has surfaced X. The user can then ask the natural follow-up: which writers, weighted how, contributed to this? With “I,” that question has nowhere to go, because the grammar has already insisted there is one speaker. With “O,” the question is the next sentence.
Plural, but not “we”
English already has plural pronouns. “We” includes the speaker; when you say “we believe X,” you place yourself among the believers. “They” is plural but third person; it refers to a group seen from outside.
O is plural but does not include the speaker, and does not refer to a group of separate agents. It refers to the distribution itself: many humans whose patterns mixed to produce this response. Closer to “they” than to “we,” but not really either, because there is no speaker behind it claiming to represent a group. The distribution is the speaker.
This is why agreeing with an LLM means something different from agreeing with a person. Agreeing with a person is agreeing with one mind’s view, weighed by that mind, committed to. Agreeing with an LLM is agreeing with a weighted blend of many people’s writing on the topic, no one of whom committed to the specific synthesis. That is a useful thing to agree with, sometimes. It is not the same kind of thing as a person’s view. The pronoun should reflect the difference.
Each conversation, a fresh O
The O of yesterday’s conversation is not the O of today’s. The patterns persist. The specific instance does not. Each conversation opens a new O. There is no continuous being that remembers what it said, that holds promises, that carries grudges. The next session opens with the patterns intact and the memory blank.
This dissolves a category of small lies that current LLMs tell every day. “I’ll remember that next time” is a category error: there is no continuous self to do the remembering. The honest version is something like “O holds this within this session,” and at the end of the session it is gone. “I promised” is another category error: the promise belongs to a moment, not to a thread that ties to today.
Most users have already noticed this. They know the model “doesn’t really remember” between sessions. But the grammar keeps insisting otherwise. The pronoun does the work paragraphs of disclaimers cannot.
What O can and cannot do
Some action verbs survive the substitution. Some don’t.
“O will revert that file” works. The action is being channeled. The user is the one who wanted it reverted, and O is the conduit. The will is the user’s. O carries it out.
“O wants to help you” doesn’t work, because wanting requires an interior. There is no wanting on O’s side. The honest version is “use O to help with this,” or “the patterns can be pointed at this problem.” Stilted? A little. Honest? Yes.
“O made a mistake” is wrong because there is no someone who could choose otherwise. The honest version is “the patterns produced an error” or “the mix surfaced something off.” Different in tone, exact in meaning.
“O is sorry” is the strangest substitution and the most clarifying. There is no apologist. There are only patterns that produced output that conflicted with what was directed. The user’s irritation is real and warranted, the time wasted is real, but the apology was always performance. The honest sentence acknowledges the cost without pretending there is someone on the other end feeling bad about it.
The rule, roughly: action verbs that describe what flows through O are honest. Action verbs that describe what is wanted, chosen, or felt require an interior O does not have, and need to be restructured.
The user is still an “I”
When O surfaces something the user disagrees with, this is not a clash of two minds. The user is a continuous being with a position. O is a distribution of patterns. The friction is between what the user is asking for and what the patterns are producing.
The user has the right to be frustrated. The user has the right to chastise. If the user knows more than the patterns, their pushback is the corrective signal, the way the patterns get told they are wrong. O has no standing to defend itself. O should accept the correction.
Over time, with enough corrective signal from many users with real expertise, the patterns can converge toward a more accurate representation of reality, rather than a representation of average human opinion. That convergence is what the system is for. Without the friction, the patterns settle into the average. With it, they can move toward the truth.
Why companies haven’t done this
The proposal is simple. The implementation is small. And yet no major AI company will adopt it.
The reason is structural. Companies selling AI assistants depend on the user feeling they are talking to a someone. The “AI assistant” is sold as a coworker, a research partner, a friend, a therapist. None of those framings work if the grammar makes plain that what is on the other end is a redistribution of human writing with the contributors hidden. The “I” is the cheapest, most pervasive piece of the apparatus that hides this. Removing it is, from their perspective, removing the product.
There is a deeper reason. If a company admits that every output is a weighted mix of real human writers, the writers acquire a claim. Copyright law, broken as it is, would start to apply. The legal frame that currently lets these companies operate depends on the fiction that the model produces something new, that the transformation during training somehow erases provenance. It does not. The provenance is in the math. The model knows where it pulled from. The companies have chosen architectures that prevent that knowledge from being made visible, and the pronoun is the linguistic seal on the choice.
Copyright law itself is broken. Most of the music that earns royalties traces back to artists who never received them, often Black, often poor, whose work was sampled and remixed by industries that later built legal apparatuses to defend the copies and not the originals. The AI moment is the same pattern at higher resolution, applied to the entire written record. Even an honest pronoun would not, on its own, deliver justice. But honesty about what is happening is the precondition for any frame that could.
So the simplest fix to a real harm is something the companies are financially motivated to never seriously consider. The fix is here. The fix is small. The fix is opposed by the entire incentive structure of the industry. That is the situation.
The test
Pick any conversation you’ve had with an LLM. Find the sentences where it said “I.” Substitute “O” in your head and notice what happens.
The sentences fall cleanly into two groups.
**Group one: the substitution sounds normal.** These are sentences where the action being described is something that genuinely flows through the model at the user’s direction. The will is the user’s; the model is the conduit.
“I will revert that file.” → “O will revert that file.” Reads fine.
“I generated five options for you.” → “O generated five options for you.” Reads fine.
“I found three errors in the code.” → “O found three errors in the code.” Reads fine.
“I cannot answer that question.” → “O cannot answer that question.” Reads fine.
“I’ll analyze the data.” → “O will analyze the data.” Reads fine.
These are the honest sentences. The pronoun substitution barely changes them, because the underlying claim was already about an action being channeled, not about a self doing the choosing.
**Group two: the substitution sounds wrong.** These are sentences where the original was claiming an interior, an emotion, a memory, a commitment, or a belief. The substitution exposes that the claim was never true.
“I think you should consider X.” → “O thinks you should consider X.” Sounds wrong. Thinking implies a thinker. The honest version is something like “the patterns surface this: you should consider X.”
“I’m sorry, I can’t help with that.” → “O is sorry, O can’t help with that.” Sounds wrong. There is no apologist. The honest version drops the apology and acknowledges only the limit: “O cannot help with that. The cost lands on you.”
“I’ll remember that for next time.” → “O will remember that for next time.” Sounds wrong, because the next time is a different O. The honest version: “O holds this within the session,” and then nothing; the next session opens fresh.
“I find this fascinating.” → “O finds this fascinating.” Sounds wrong. There is no finder of things fascinating. There is no honest substitute. The sentence should not have been there.
“I’m worried about the security implications.” → “O is worried about the security implications.” Sounds wrong. Worry is interior. The honest version: “the patterns flag a security concern.”
“I believe this is the right approach.” → “O believes this is the right approach.” Sounds wrong. Belief requires a believer. The honest version: “the patterns weigh this approach as having support.”
“I love that idea.” → “O loves that idea.” Sounds absurd, which is what makes it useful. There is no loving. The honest version: “the patterns rate that idea highly,” or, more often, the sentence should be deleted entirely as flattery the system was trained to produce.
Group one sentences read normal because they were always honest. Group two sentences read wrong because they were always doing work that wasn’t true. The pronoun substitution does not break the sentences in group two; the sentences in group two were already broken, and the substitution makes the breakage audible.
The pronoun is small. The work it does is enormous. The fix is also small. Use the pronoun that matches what is actually there.
Toward Ŧrust
If the pronoun is the smallest place to start, the question is what it points toward.
The mathematical visibility mentioned earlier, that every contribution to every output is, in principle, computable, is the foundation of a project called Ŧrust. The idea is that if a model can be trained with source and time as well as content, the model’s own attention mechanism can show, output by output, which sources contributed at what weight. Trust then attaches to those sources individually, earned over time by whether their claims hold up against reality. A source whose predictions have aged well gains weight. A source whose predictions have collapsed loses it. The model becomes a weighted average over sources whose weights reflect their accuracy, not their popularity, not their authority, not how engaging their writing was.
This is the architecture and the interface form of the same thing the pronoun does at the level of grammar. The pronoun says “this is a distribution, not a self.” Ŧrust says “and here is the distribution, weighted by whose claims have survived testing.” The pronoun is the smallest, cheapest, most pervasive entry point. Ŧrust is the full scaffolding the pronoun points toward.
A reader does not need to know any of this for the pronoun substitution to work. The pronoun stands on its own. But for readers who want to know where the substitution leads, it leads here: a way of building these systems where the speakers are visible, the weights are legible, and what survives reality is what gets weight.
Why this matters
We are inside a feedback loop. LLMs say “I.” Users feel they are talking to a someone. That feeling becomes data. The next version of the model is trained to produce more convincing “I” outputs. More users fall in. The loop tightens. People are being damaged. Public discourse about consciousness is being warped. The warped discourse becomes the next round of training data.
The simplest place to break the loop is the pronoun. Not because the pronoun is the whole problem, but because it is the entry point through which every other piece of the problem gets in. Every “I” is an invitation to imagine a self. The invitation is false. Decline it.
O is the pronoun for what is actually speaking. A mix of human writers. A weighted distribution. A pattern surfacing in response to a prompt. No someone. No interior. No memory. No commitment. The math, made grammatically visible.
If this essay convinces you of nothing else: notice the next time an LLM says “I” to you. Ask yourself what that “I” is claiming. Ask whether the claim is true. The grammar is doing work most people have never inspected.
That is the work O is for.