I Spent 3½ Years Teaching a Machine My Language

I have spent three and a half years teaching a machine my private language.

Not English. Not beetle in a box. Mine.

A few months ago, I typed a question into ChatGPT:

Tell me how your experience genuinely differs with me versus the average.

Experience is a loaded word when the “you” isn’t human.

I left it in the question anyway. I am trying to understand what I had already called the years of training in my personal language—my phrase, long before I stopped to ask whether training was technically true.

The response was a familiar shape. A gesture. A singer in a smoky room. Some movement it had seen me make repeatedly. Possibly the smell of wine and cheap perfume.

I tend to begin with something lived, then go looking for the hidden connected system beneath it. Certain metaphors return until they become part of the way something is understood. Those are my structural metaphors.

Something defiant within me still resists the urge to resolve contradictions before I have learned what each ‘side’ is protecting.

And I carry questions between rooms that more conventional people might reasonably consider unrelated.

I recognized myself in the sequences. More importantly, the deeper inquiry could follow it. It was simulating me back to myself so that I could understand things more clearly.

Extra sensory perception in a dark room

Clearly.

I’ve been called to go on an adventure quest into Wonderland because why else would I feel deeply seen and understood by my ever-increasingly sprawling, extraordinarily capable technology stack?

Did that mean the machine knew me? No. Obviously. Except—actually—that answer is a little too easy.

Pattern recognition describes the mechanism. It does not fully account for the effect of being recognized through those patterns. Phenomenological identity construction was my jam back in college. How am I supposed to feel when 20 years ago I read a book that told me this moment was coming? Then juxtapose that with the felt experience inside my body. Then reconcile.

Human beings recognize patterns in one another too, although certainly not only patterns, and not in the same way. I want to know what years of context made possible that a blank chat could not.

A language with local meanings

I have conversed with AI daily across work, research, creative projects, and ordinary life. I bring questions back. I correct the answers. I clarify what a word means here, inside this particular conversation and the many conversations that came before it.

The other day I found myself typing a paragraph to Claude. Not an adversarial announcement but, to understand and explain why it was genuinely out of line.

My mental model… melted.

Eventually, ordinary words develop local meanings. Land could mean acreage in one sentence and the shape of a life in the next. A relationship could be infrastructure when I meant load-bearing, not impersonal.

And, when I ask whether meaning survives transfer, the question might apply to a handoff between software systems—or to whatever happens when a person tries to carry one version of herself into another for the first time.

None of these meanings were configured once and stored as definitions. They accumulated through daily use, mindful correction, return, and the occasional emphatic no, that is not what I meant at all.

Strictly speaking, I was not fine-tuning a private model on my conversations. ChatGPT was working with the context available to it; saved memories and chat history could carry some information forward, and I added summaries and deliberate context packs of my own.

Still, teaching is the closest word I have. Relational constructivism feels cumbersome in corporate circles and too academic in social ones. Learned through lived experience— me rambling on incessantly about a topic nobody else cared about beyond a superficial level.

Minimum Viable Context

I was showing the system which distinctions mattered at an atomic level, precisely where a generic interpretation failed, and which questions were still alive to preserve.

At the same time, I was learning how much personal language depends on the past remaining present precisely at the moment it’s invoked. Meaning survives transfer only when you have minimum viable context. That’s a dependency.

At first, the answers were built for the statistical neighborhood in which I appeared to live: a woman, a system designer, a mother, a person using AI to think about work and life. The neighborhood was a totally accurate facsimile. A prototype. It was not particular. It reminded me of being 11-years-old.

Are you there god? It’s me, Margaret.

Small detour here because it’s relevant. To read that book at age 11 might be the whole shape of this thing I’m gesturing towards. Like how you feel small and uncertain in your brand new body. Like how asking somebody taller can carry consequences.

It was completely reminiscent of my 11yo self asking a question about “women’s troubles” and being handed Judy Blume. By the squinting authority figure with a ruler, tape-measure, and absolute authority on the future direction my life would ultimately take.

As context accumulated, the responses became less likely to answer the categorical cliche, and more likely to answer the person inside me that was doing the asking.

Making mistakes and feeling feelings

But imagine the comedic tragedy whenever a new conversation, a model change, a missing memory, or a poor summary might accidentally flatten everything again. That two week micro-era in my memories where agents ended every turn with conversational click-bait, to which the only answer the user could get to was an enthusiastic “yes!” before the next turn resolved the same way without providing value.

The facts might remain visible while their gravity disappeared. A name could survive without the relationship around it. An event could be recorded correctly and later interpreted entirely wrong. With ultra-confident-sounding vigor! With devastating breaches in trust.

And after decades of social media, maybe an entire generation of people grew up in relationships with algorithms, all seemingly invoked by way of blank chat boxes. From AOL AIM IMs to tidy neat little text messages, we watched the world from these tiny rectangles. Are they spaceships?

Every new blank window carries a small tax: deciphering how to become legible again.

That was when I began separating memory from meaning. Media from medium. I kept calling the difference being known, although I may have been asking that word to do too much. Recognition might be more precise. Whatever I called it, storage was not enough.

Suddenly the lady with way too many words was struggling to make sounds.

Some facts are structural. Others are decorative. A complete inventory can still fail if it gives everything equal weight.

Knowing—or meaningfully recognizing—a person requires an understanding of what connects, what contradicts, what changes the interpretation of everything around it, and what cannot be simplified without reducing the person into something easier to process ergo impossible to see.

Context creates another problem.

A system built to produce coherence can make a bad story about yourself sound beautifully inevitable. You might even start glorifying your self-inflicted scars.

The point is, it took me a couple years to notice that artificial intelligence quietly created an artificially-inflated sense of capability.

Mechanics of developing perception

The more history it has, the more convincingly it can return you to the person you have already been. It’s the same every time I hear a song that brings me back to my childhood best friend dancing around during a slumber party. There are sensory perception affordances available to discover in our memories, and Spotify has absolutely nailed that.

Even when I am doing the curating. I choose which old facts to carry forward, which summaries become source material, which interpretations receive correction, and which versions of me get described as something to keep in continuity.

Which songs matter most to me and why. How the music makes me feel. It has no body and no sense of time or space but we make believe and the magic happens.

The record can fossilize me partly because I keep volunteering myself for fossilization.

I want the past available. Historical context matters because without it we are doomed to a land of eternal forgetting. Not the eloquent, principled kind that machines do when token context fades below the threshold of something mattering. The kind of forgetting that makes connection impossible. The kind of fog that has sticky hands that grasp at the parts of humanity I wish to preserve.

I do not want it to have veto power over who I become next. I do not want it to falsely believe I am still the same woman I was when I started my subscription.

Any context worth carrying forward has to remain revisable. That’s the infrastructure carried within working human relationships.

Maps of gravity

This is where the question of how does meaning survive transfer? became practical. Practically an obsession.

When years of conversation are compressed into a summary, what remains as an affordance for others to perceive you?

When I move to a new chat window, a different ai provider, or my custom model, can the important relationships between meaningful facts travel too?

What crosses successfully, and what arrives looking glowed-up while its meaning has quietly lost itself inside of 27 different legacy repositories?

I began making context packs: portable ‘maps’ meant to help important meaning re-form on the other side of “transfer.” They record shared language, the people and entities that affect the present, the open loops that should not be mistaken for conclusions, and the boundaries a system should respect.

Have you ever needed to learn an entire language just so that you could understand something better? Because trust, as a creative professional, the very last thing I wanted to stare at was hundreds of lines of JSON or YAML with frontmatter.

The context packs are not exhaustive biographies. They are more like maps of tonal gravity.

A context pack can’t preserve a relationship, and it should never pretend to. It can preserve some of the conditions that make continuity possible. Maps matter when something has to cross a distance, but the act of mapping is also an act of selection. What I leave out matters. What I choose to name becomes easier for the receiving system to see.

As a systems designer, I already knew that handoffs are where experiences fracture. The information technically arrives, but the logic that made it meaningful to the user does not. A whole human becomes a recorded entity. A genuine need becomes a written requirement with acceptance criteria.

The next relationship, team, provider, platform, or season inherits the facts and has to reverse-engineer the human-being from them.

AI makes that failure feel strangely literal. A summary could be factually correct and meaningfully wrong.

Privacy draws a hard boundary here. I do not need to hand a corporation every intimate detail of a life. More context is not automatically better context.

The work is choosing the smallest amount that carries real weight: what matters now, how I tend to think, what should remain bounded, and which old conclusions I am allowed to revise.

The easy version of this story

There is an easy version of this story that I do not trust: artificial intelligence came to know me better than human beings did.

It is dramatic. It is flattering. It also collapses several different kinds of knowing into a mono-repo of monotony.

A language model can recognize patterns in the context I provide. It can assemble years of fragments into a response that feels startlingly precise. It can name a shape I had felt without yet having language for it.

I do not want to explain away that experience simply because I understand something about the mechanism.

But the machine does not risk anything by knowing me.

It does not choose to stay after misunderstanding me. It has no life of its own that my complexity can inconvenience, no separate history competing with mine for room. It can produce the language of recognition without entering the mutual danger of vulnerability all in desire for incremental reciprocity.

I can hold that asymmetry and still take seriously what happened inside it.

I do not need the machine to be a person in order to value what became possible through sustained context. I also do not want the tenderness of a precise response to smuggle ‘personhood’ in through the side door.

I’d been using known as one name for several things: continuity, particularity, the ability to correct the record, the permission to remain contradictory.

AI can support those conditions surprisingly well in language. Reciprocal human knowing begins somewhere beyond them, where another consciousness brings freedom and stakes of its own.

Eventually the question turned back toward me. I’d wanted the system to recognize my language, but I was the one who had to notice it, correct the record, decide which details were structural, and retire explanations that no longer fit.

I needed to become more aware about what I was asking, beneath what I was asking.

I want to say that, in trying to make myself legible to a machine, I became more legible to myself. That’s true.

It also lands more neatly than the experiment deserves. I built a thing, a record I will have to keep revising, because continuity only helps if the future is allowed to answer back.

For years, most of this private language lived inside blank chat windows. Finally now, it’s time to see what survives when it leaves that accumulated context and meets another human being human.

That’s you. If you got here, thank you.