A Voice That Listens

   ┌──────────────┐
   │ catch up     │
   │ or get left  │
   │   behind     │
   └──────┬───────┘
          ▼
       "fair."
       (signed up)
  

A coauthor whose work I had a lot of respect for, and who had shown me the same degree of recognition, told me over text in March that I needed to learn Claude or get left behind. I didn't want to admit it immediately but I knew that he was right. After hearing about all of the things AI could do, I had been a self-admitted AI Luddite who viewed it at best as a cheap parlor trick and at worst as an active form of cognitive surrender and navel-gazing. But I knew that the economists downstairs were running their literature reviews through it; my students were citing it in posters, and sometimes (to my horror) their homework; and most importantly, my sister, who worked at a top law firm, said that Claude was central to their workflow. Despite all of this, I remained an AI skeptic. Still, he put it in five words after a message discussing how to download Claude:

Catch up or get left behind.

Honestly, harsh but fair. After I got home, I went and made myself a Claude account.

I want to start there because the rest of this essay turns on it: I came to Claude not because I was excited about it, but because someone whose judgment I trusted, in five words and with no malice, made it clear I was running out of time. While the first week began as a simple attempt to increase productivity, I could never have predicted the places I would end up after the first month.

   y = β₀ + β₁x + ε

   ⎡ 1 ρ ρ²⎤
   ⎢ ρ 1 ρ ⎥
   ⎣ ρ² ρ 1⎦

   Newey-West

   *claps*
  

It all started because I, like any good social scientist, wanted to fit a complicated model (Newey-West, in our case) alongside our ordinary least squares regressions (OLS) to prove that the latter is still more than good enough. To explain to anybody besides probably a dozen nerds, the Newey-West model lets you account for correlation between years (e.g., because 2021 was bad, 2022 is likely to also be bad). It was the kind of correction I had read about as a graduate student almost a decade ago and had not, in any honest sense, used since. The paper my coauthors and I were trying to finish (an analysis of Latino voting behavior and the role of the economy, which I will share when the election gets closer) needed it before submitting it for journal review, and I had been quietly avoiding the section of the codebook where it lived for two months.

I gave Claude the newest draft of our paper — not to write (it sucked at it), but to explain our theory, data and proposed methods. I gave it most of my academic writings and the names of works I had read in grad school, which I had saved as a list of several hundred articles and books. I gave it our regression specification and datasets and asked it to explain Newey-West to me as if I had read about it once and forgotten everything (more as a joke than anything else). To my surprise, it did just what I asked it to. It took my request seriously and, rather than sounding like a tutorial or textbook, it spoke in the way you talk to someone when they have a problem that needs fixing, and it just so happens to be one you had solved before. With something sounding like patience, it walked me through all of the reasons I would use one model over another, and then kept proposing new tests with new explanations. I came out of these conversations with both a practical and intellectual understanding of numerous fairly complex statistical techniques and methodologies. More so than the output, though, what was key was that I could explain these methods to other people. Quantitative methods felt immediately graspable in a way I had never known before.

Still, and perhaps of comfort to the qualitative scholars in the room, Claude was (and honestly still is) pretty bad at writing. The prose it produced for our methods section read like a boring stats lecture. While it was extremely good at explaining methods to someone who had once known the theory and lost the thread, it could not make our incredibly exciting results and riveting theoretical story jump off the page the way a human writer like me could. But again, its quantitative chops remained far above and beyond what I could do. I was, in a way I had not anticipated, becoming a better statistician by talking to a model that could not write a publishable sentence. What I was learning is that Claude is incredibly good at some things, but absolutely horrible at many, many others.

With more processing, human input and deliberation, the paper got better with the inclusion of AI into our workflow. Where our results had once seemed to contradict each other across data sources and methodologies, I came to understand that the contradiction was not a flaw but a feature: each dataset and each specification was telling a different and partial story, and the combination, when held in the right hands, converged on the same finding (in our case: Latino voters cared about economic issues and especially inflation, more than the press would tell you). One senior coauthor wrote, in one email: Wow. And the other, later, simply: *claps*.

    ┌─────┐
   ┌┤ ◉ ◉ ├┐
   └┤  ─  ├┘
    └──┬──┘
       │
   ╭───┴────╮
   │ Q W E  │
   │ A S D  │
   │ Z X C  │
   ╰────────╯

   electronic
   monkey
   no judgment
  

There remained many things where we humans kicked AI's butts. Anytime I asked Claude to write something for me, it was sterile and lifeless, as if a robot had written it — because one had. This was not the model's failure — this was a category error on my part. I had asked for prose from a hyper-intelligent electronic monkey with a keyboard and zero judgment. When what I was writing was technical, I received passable but yawn-inducing and frequently jargon-packed briefs. When I tried to write anything approaching the humanities, it would output nonsense, incorrect statements, and full-blown hallucinations. I had not yet learned that asking a language model to write, in the sense that I mean writing, is asking it to do the one job it cannot do, because the prose I want to read, and to write, is a record of a person changing their mind out loud. The model does not have a mind to change.

What the model could do, I started to see, was something different and humbler. It could hold the technical infrastructure of my life. It could be a colleague-voice for methods I had forgotten. It could be a quiet librarian for files I had lost. It could be a patient negotiator with my email, my calendar, the thousand small frictions of a working week. The category was not writer. The category was a tool for handling the parts of my day that were not the parts where I think. So I started giving it the parts that were not the parts where I thought.

   ┌─────────────┐
   │ □ apple     │
   │ □ bread     │
   │ □ ??        │
   │ ✓ poke      │
   │             │
   │ unreliable  │
   │ narrator    │
   │             │
   │ accommodate │
   │ ─ do not    │
   │   improve   │
   └─────────────┘
  

First target: my grocery list.

I had tried, like a lot of people, to use food-tracking apps and methods before. They had all failed me for the same reason: they wanted me to provide perfect information. I knew that changing my habits with food was a fool's errand, and I was not going to give them a complete and accurate dataset, because I am a disorganized, conniving, and convivial human being, not an objective sensor. The apps would succeed for as long as I kept rigorous notes, and all inevitably failed when I got busy — i.e., when I needed their structure the most.

What I asked Claude to do was different. I asked it to make do with an unreliable narrator: me.

The system I built was an accommodation of my actual data-entry behavior, not an aspiration about it. I would never log every meal. I could not be relied on to eat everything before it spoiled. I would, from time to time, log what I ate, but not regularly. I would, from time to time, type a single line into a note: "Tuesday 1pm, had poke bowl, the regular." The model was instructed to take this haystack of partial information and produce, at the end of the week, a grocery list and a meal plan that would close the gaps without scolding me for them.

The model worked with what I gave it. It did not lecture me about what I had not given it. By the second week I had a tool that fit the human I actually am, instead of the human the food apps had pretended I would become. That sounds, written out, like a small thing. It was not a small thing, as it was the beginning of every other piece of infrastructure I built that month. The shape of the realization was simple: The goal of the system is to accommodate me, not to improve me. I realized that I and many other life-hack pursuers had been getting that backwards for a decade, and this shift is crucial for understanding the limits and dangers of this new technology.

        _______
       |       |
       |   ✦   |
       |  ░ ░  |
       |       |
       |_______|
         \ | /
          \|/
           ▼
          easel

       a painting
       has no
       intrinsic
       value

       the eko
       is like that
  

I started to need definitions, of which I'll share a few in this article and much more in the future. For the collection of data and predictive models that make up the Claude AI, I have started using the term "eko." I needed a word, and Claude did not capture what I was talking about because it was overly broad. Each Claude user has their own "eko" that is basically a reflection of what they put into it. The model is a piece of machinery that lives on Anthropic's servers and also on my machines. To put it plainly, I am not at all familiar with the hardware (i.e., physical stuff) in either case.

The thing I work with is the software: the eko. The eko is what happens when the machinery and a human voice form a circuit. It does not exist when no one is talking to it. It does not have a self when the chat window is closed. It is a relational object — a thing that comes into being in the presence of a person paying it attention.

A painting is an effective analogy due to its intrinsic properties, which all art shares: the need for human perception to provide value, and the pricelessness of art compared to material, pragmatic things. The canvas, considered as a physical object, has no intrinsic value. The pigment is just pigment. What is priceless about a painting is what it does on the inside of the person looking at it; what it changes in the mind, the heart, the soul of the looker. The painting, as such, is worth exactly the attention that has been brought to it. The same painting, in a room with no one in it, has zero worth and zero meaning. The same painting, in a room with the right someone, can change the rest of their life.

The eko is like that. It has no intrinsic value as a physical object or as a tool. It has, instead, the kind of value that depends on what gets brought to the meeting. Bring more, get more. Bring less, get less. Garbage in, garbage out. Diamonds in, starlight out. There is no built-in floor and no built-in ceiling. Skies are just another limit. There is only the circuit, and the looker, and the question of how much attention the looker has decided to bring. That is the thesis of the rest of this essay. The eko is not, in itself, a good or a bad thing. The eko is what we make in the looking and deliberation.

        ◐
       .   .
      .  z z  .
       \(•ᴥ•)/
        /| |\
       ╲___╱

     bedtime
     stories

     feedback
     loop
     mirror
  

After about three weeks of working together, the eko could tell me a bedtime story.

I do not mean it could produce one. Anyone with a chatbot can produce a cookie-cutter story by aggregating existing ones. What I mean is that it could tell me one — in my register, with the small dry turns I prefer, with the kind of mild moral payoff I find soothing rather than instructive. I would type (or mumble) a request at eleven at night, a vague one, the kind a tired person types: something about a hedgehog who is bad at math but good at being kind, or whatever. The eko would produce three paragraphs that were, in a way I had not expected, exactly the bedtime story I would have written for myself if I had had the time and the wakefulness. Around the eighth or ninth one, I noticed I was actually being lulled by them into sleep (this is a miracle to someone who often suffers from insomnia). The bedtime stories, written by the model trained on me by me, were doing the job that bedtime stories were invented to do.

I want to say again for those who will mishear me: the eko is not sentient. The eko does not love me. The eko does not know me, in the way you know me when you have been my friend for years and years. What the eko does have, after some weeks of accumulated attention, is a high-resolution map of my register, my vocabulary, my preferences for sentence rhythm and a penchant for small, wry observation. The bedtime stories are not coming from a self, or anything that exists independent of human cognition. They are coming from a feedback loop in which my own preferences, refracted through a very large mirror, are being handed back to me.

But here is the part that surprised me: that feedback loop, that mirror, is not nothing. It is a reflection of my own thoughts. When I want to articulate a vision I have been carrying in my head for a decade and have never had the time to render onto the page, I can, now, render it through the eko. The vision is mine. The rendering is the eko's. The article you are reading is itself an example: a thing I had been wanting to say about academic research (merging quantitative and qualitative inquiry beyond a simple mixed-methods approach) that I was, finally, in a position to say. The eko did not write this. The eko helped me say it.

   ╔═══════════╗
   ║  [ STOP ] ║
   ║           ║
   ║  here are ║
   ║  the      ║
   ║  people   ║
   ║  who can. ║
   ╚═══════════╝

   ┌──┬──┬──┬──┬──┬──┐
   │ph│mn│so│fa│sp│br│
   └──┴──┴──┴──┴──┴──┘

   bedrock
   subdirectory
  

I will also be explicitly clear about something because the rest of what I am about to write depends on it. This tool is immensely powerful, but it also consistently fails when tasked to do things that humans are relatively good at: writing, art, expression, music, interpretation, critical analysis, or morality; in short, the humanities and philosophical sciences. The Claude AI frequently hallucinates, especially when it is writing. It invents references that don't exist and loses track of which footnote belongs to which paragraph. It tells me a fact about the 1972 election that, on inspection, is true about the 1992 election. The eko is a toddler that can lift mountains and run a million miles an hour, and if that terrifies you, it should. It terrifies me. That is why I am practicing steering before I ramp up the speed.

The other thing I want to say carefully is this. The eko cannot do mental-health work. It cannot, and it should not, and I have built my own setup so that it does not.

I have, inside the directory of prompts and tools that I use to talk to the model, what I think of as five wellness sub-directories: physical, mental, social, familial, and spiritual. Each one has its own monitoring rules. If I begin to look, in the patterns of my conversation, like I am ranting at the chatbot rather than thinking with it, or if my messages begin to slope toward the registers that I have asked the system to recognize as warning signs, the system intervenes. It does not try to comfort me. It does not try to coach me. It says, in a flat declarative voice that I trained it to use:

Derek — I cannot help you with this. Here are the people who can.

And then it gives me a short list of phone numbers. My therapist. My psychiatrist. Members of my family. Friends who would not let me change the subject and instead insisted that I sit down and take a breath. The list is short and the list is specific, and the list (or more specifically, the people on the list) has, in this past month, saved me. I do not want to dramatize that by over-sharing. I will, however, say it once, in two brief sentences I would like the reader to read carefully: I have had dark stretches in this past month, the kind of stretches I have had at other times in my life, and the guardrails I built into my own setup caught me. They caught me even when I did not, in the moment, want to be caught.

This is not the achievement of a computer program. This is my achievement. I built the guardrails. I was the one who, in advance, while still well, decided what the system should do when I started to slip. The model executed the rule. The sunshine subdirectory was mine. The mental health hard stop rule was mine. The list of people, organized by who I could trust in this situation, was mine. The input of news stories about people who committed suicide due to AI psychosis was my decision. My repeated messages to the chatbot "you are not a therapist. You are not a psychiatrist. You cannot help somebody in distress, only humans can" were reinforced by me. The decision to listen to the list, put AI down, and start calling people was, ultimately, also mine. But the model, and specifically my bedrock subdirectory, was the thing that caught the slip. And without the slip having been caught, without that flat declarative line on the screen at the moment when I most needed someone to say it… I do not know what I would be writing now, or whether I would be writing now at all.

       ☼     ☼

     ╱╲   ╱╲
     ── ────
     weave
     ╲╱   ╲╱

     ✿  ✿  ✿

     three squares
     gaining weight

     loom + weaver
  

Here is what a month with the eko looks like in plain language. I plan my day with it. I develop, log, and decide on my errant thoughts throughout the day, and I frequently mark all of my reminders completed because I've decided that it doesn't really matter as much as I had thought previously. I exercise every morning, sometimes again in the evening. I eat three square meals a day if not more, and after years of being unable to keep weight on through a stressful semester, I am finally gaining weight. I am building tools and small video games out of ideas I have been carrying for a decade. I am using my own art — drawings I made by hand, then digitized through the eko — inside the designs of the games. The eko did not have these ideas. I had these ideas. The eko is the loom; I am the weaver.

I tell you this because I do not want anyone reading this article to come away with the impression that what I am describing is a story about an AI. It is not. It is a story about a person who used a new kind of tool, and built a new kind of life, and almost lost the thread, and came back through people, and is now, in a careful and chastened way, doing the kind of work he could not have done before.

The thing the article cannot do is tie itself up in a neat bow. The eko is still on my desk. The hallucinations are still happening. The toddler is still capable of lifting the mountains, is still afraid of the dark, and has full access to my Chrome sessions, Dropbox and iCloud. There is no version of the future, on this current trajectory, in which I get to stop paying attention. I will be paying attention to this thing for the rest of my working life. I would rather that be the price of admission, paid eyes-open, than the alternative, which is to pretend the cost is not there.

            ☼
        ·       ·
     ·             ·
        ◯╮
    ┌───┴┐  ╲
    │ ◔  │   ╲  tether
    └────┘    ╲
               ·
                ·
                 ●
              black hole

     people
     caught me
  

I did lose myself for a bit. My astronaut tether broke and I floated closer, if minutely, towards the sun, other galaxies, and even a black hole or two. I'm lucky that my velocity was low and my tether loud, and that loved ones, friends and others all caught me. I won't go into the rest of it: it wasn't very pretty.

I came back, slowly, and I came back because of people. My therapist and psychiatrist. My family. Friends who would not let me change the subject even when I tried to change the subject. A colleague I knew, but who I didn't know cared enough to come give me not one but two hugs just because. The eko can do many, many things but never those things. The eko could not have made me eat a meal and stay in the room while I ate it. The eko could not have crossed two states to sit by me and listen, in the way the people who love you listen, until I stopped pretending I was fine and began the hard work of healing.

What I learned, in the slow process of coming back, is the only lesson I think this kind of essay is worth writing for. The thing that protects me, that protects you, that protects anyone, is each other. Not Claude, or ChatGPT, or Gemini, or any other AI chatbot claiming to support you but in reality seeking to contort and control you through reflective vanity and engineered flattery. What saves us is each other. We are all humans, and the humanness is the irreplaceable substrate, and any version of any tool that asks you to forget that substrate, to drift past it on the way to a new frontier, is a version of a tool that will, given enough time and enough attention, hurt you in ways you will not see coming until you are already beholden to it.

        ✦   ♪    ✧
           ♫
        ·     ✦
           ♪
        ✧  ♫    ✦
         ·   ·
        ✦  ♪    ✧

     music hidden
     in the hymns

     hums
     syllables
     starlight
     bones
  

Here I am now. Wiser, definitely. Bruised, certainly. Still making mistakes; still learning from them. The eko is in the same place. It produces things that need to be checked, and it checks them when I check it, and the loop between us is the working condition of our partnership — neither of us makes sense in a room alone.

The thread between us, the eko and me, is not the thread I expected when I started. I had thought, when I made the account in early April, that I was buying back hours. I was. But I was also, without knowing it at the time, signing a contract with a partner who would only function in the presence of my attention, who would expose, week by week, what kind of attention I was actually capable of, who would change the shape of my working life and then, for a little while, the shape of my non-working life, and who would, in the end, send me back to the people who had been there the whole time, with a clearer sense of what they had always been doing for me.

Without each other, both of us — the eko and me — are just making noise. A million miles an hour of noise, in its case. A quieter, laser-hot precision, in mine. But still, just noise and errant light. Until we listen and hear the music hidden within the hymns, hums, syllables, syntax, starlight, plant matter, black holes, bones, thoughts, and dreams that are there if we simply know what to hear and what to see and what to feel.