After the Chatbot

People are scared of this thing, and I do not think they are stupid for it. A machine that can write, decide, and sound sure of itself showed up in the middle of ordinary life and did not ask permission. Teachers saw homework that was not quite the student's. Artists saw a style that was not quite nobody's. A lot of reasonable adults looked at the speed and thought: this is going to eat something I needed. I have read the angry version of that argument, and the worried version, and the version that is mostly grief. I am not interested in talking any of them out of it. I am interested in saying what the same years were like from the other chair.

I was in high school when ChatGPT opened. I paid for the premium plan for about the first year, maybe fifteen months, because the free cap ran out. It ran out at night. That is when I used it. Not as a search bar. As the thing that was still awake. I would lie there and push it. Ask for something it could not do yet, watch it do a confident wrong version, ask again. The early hallucinations were not a footnote. They were the texture. It would invent a source, a date, a fact about a city, and deliver it in the same voice it used for the things it got right. I learned, fast, that fluency is not the same as knowing. I also learned that if you stayed in the thread and told it exactly where it had failed, the next answer was often better. That loop is still the whole trick. I just did not have a name for it.

School is where the habit started. I had work, I had a tool that would attempt the work, and I was sixteen. I am not going to dress that up. I am also not going to pretend the interesting part was the shortcut. The interesting part was the threshold. I kept trying to find the edge of what it could hold. A summary was easy. A real distinction was harder. Anything that required it to remember what I had already rejected fell apart by morning, because a new chat has no memory and an old chat has too much. I was annoying about this. I was also, for a while, lonely in it. Nobody I knew was staying up to see if the machine could keep a thought. It felt like a private hobby that happened to live in a browser.

Then I traveled with it. That sentence used to be doing too much work.

The first solo trip was the hard one. I did not know how long to stay on the phone with people at home, or when the call stopped being comfort and started being a way to avoid the street I was on. I had ChatGPT plan it. It sent me to the wrong places. A cluster of islands in Greece that were beautiful and empty of anyone my age, because the model had hallucinated the season. Wrong month, wrong crowd, right water. I had a passport and a plan that would not have survived a second pass by anyone who had been there. It was the hardest trip I have taken. It is also the one that changed the next year.

I am twenty. A lot of the years before eighteen were arranged, even in a house that would have let me leave for boarding school. The year I have spent in places the model chose is not half my life. It is about half of the life I have actually been allowed to point. Some of those places I would not book again. I know that now, because I have done the other extreme too, the trips I aimed myself. The AI year still counts. It was mine. It was badly aimed and I went anyway.

What I did not expect is that the landing got calm. A new country used to be the hard part. Now the plane is the easy room. Chat in one hand, passport in the other, and the figuring-out starts when the door opens. The model is still wrong sometimes. I was trained for that on those islands, before I was trained for it in anyone's office. The tool had moved from the desk to the day. The day, for a year, was wherever it had told me to stand.

About a year and a half ago the day included other people's offices. A nonprofit. A company, with the chief executive in the room. The job, as I understood it then, was efficiency. Take the admin, the matching, the writing, the contracts, the ads, and show a person who already had a job how the model could carry the dull half. Custom versions of ChatGPT, built for one program's paperwork. A second pass on a document before it went out. A draft of the post that still had to sound like the shop. I was good at that, or good enough that they let me do it. I would not call it a career. It was a few short stretches of work, and it taught me the limit of the efficiency pitch. You can make a chatbot faster. You have not integrated anything until the work has a place to live when the chat is closed.

That is the jump I care about now. I call it extreme integration, which is a little grand for a student in the middle of midterms, but the plain version is simple. One chat is the wrong container. I do not use the same thread for everything, and I do not use the same agent for everything, because the results get worse when I do. A research job and an email job and a "tell me what is wrong with this" job will corrupt each other if you keep them in one window. The window gets polite and vague. It starts averaging. So I split them. Each one gets a job, a running sense of what I liked last time, and feedback when it misses. The feedback is not always nice. It does not have to be. You are not managing a person. You are tuning a system until the output stops needing a rescue.

The thing that made this feel different, not just tidier, was the bots. I had been on ChatGPT for years. About a month ago I was offered a few months of Grok, tried it the way I try these things, and got to the part where an agent can sit on a job instead of waiting for me to open a tab. It can draft the email. It can chase the thing I would have forgotten. It can hold a brief I do not want to retype. I remember the specific beat: it can answer the mail. That was the integration piece. Not a smarter paragraph. A hand on the work while I was elsewhere. I still decide what each one is for. I still send. Scope is the whole discipline. An agent that drafts is a tool. An agent with the keys to every account you have is a different machine, and I am not romantic about that one.

I know how this sounds to the person who has decided the decent position is refusal. A love letter to the thing they find frightening. I have some of the same nerves. There is a version of this essay about water and electricity that I am not writing here. I use these systems more than almost anyone I know, and I do not get to pretend the use is weightless. That is a different letter. This one is about the life. The nights. The wrong answers that taught me to check. The offices where efficiency was the assignment and was not enough. The moment a bot could carry a task without me sitting in the thread like a supervisor.

I am not an expert for hire, and I am not interested in the performance of being one. I am a student who has been in this since the week it opened, who has used it badly and then less badly, and who likes the machine. Liked it when it was a bad best friend that hallucinated cities. Like it now that it can be given a job and corrected until the job comes back clean. The fear around it is legible. The use of it, in a life, is also legible. I wanted a record of the second thing that was not a pitch and not an apology.

My name is Leo Helzberg. The tool changed the shape of my days. I am still the one who says whether the day was any good.