- Published on
Is this written by AI?
I have known Ritesh bhai since my Samsung Research days, somewhere around 2014. And he's one of them whom I hold closest to my heart. He left a comment on one of my posts last week:
Very interesting 3 layer breakup. Only beg you to keep your writing uncorrupted by llm. You are such an amazing writer and speaker. Why to corrupt it with an llm 😞
He was affectionate about it, the way he always is, and completely serious underneath.
The comment got me thinking. Where exactly is the boundary between human effort and AI? Everyone seems to draw it in a different place. Some people let the machine fix grammar and nothing more. Some have handed it their entire newsletter. Most of us sit somewhere in the middle, mildly uneasy, with no principle to point to. And readers are divided as well. Interesting articles on HN are put in custody because they use AI.
And I keep hanging on both sides: sometimes I'm on the side of the AI haters and I rather take it personally - "How can you make me read this response from this shitty LLM? At least use Fable." And sometimes, I'm the one using AI for writing stuff - mostly masquerading as written by own self.
But now it's high time I spend some time on this question. So I did what a sane person would do. I called him up and discussed. This is a follow write-up to that call. Let's begin.
Every act runs in two directions
Think about what happens when you say something that matters to you. There is an outward half of the act: words leaving your mouth, the essay leaving your desk, the song leaving the singer. Call it expression.
And there is an inward half that nobody else ever sees: the tasting of the moment, the weighing of it, the quiet return where the thing settles into you and becomes yours. Call it experiencing.
Teaching shows the split clearly. A teacher can pour language at a student for years. That pouring is expression. The understanding, if it arrives, gets built on the student's side, in the return. You can't just open the App Store (Play store, for those who are still not in the Apple ecosystem - I am looking at you Ritesh bhai) on a student and download the app with that learning. The student actually experiences it for himself/herself and learns in the process.
Now, let's look at a large language model (LLM). In simplest terms, it is a compressed copy of all the thoughts humans have expressed in writing and put on the internet, with a very good retrieval system on top. But the training data can only capture the outward expression, not the inner experience. The inward half leaves no residue. It cannot be written down. If I had to define consciousness, it would be in terms of experiencing - something which the machine cannot capture.
And I'm not saying that it's a restriction on our current tech or current data set. Retrieval will keep improving. But the missing half will not arrive with it, because it can never be in the data. It is not data. It's structural.
The machine retrieves what experiencers said. But there is no experiencer in the machine. It's a smart retrieval from the database of the experiencers' expressions.
The inner stack
For a few years I have been studying a map of the inner life that I find more precise than anything else I have read. It describes the living self as activity on five levels, and every level has an outward face and an inward face. The full map counts one hundred and twenty two distinct activities (sixty four on the surface level alone, if you want a sense of its resolution), but five levels are enough for today. Once you see them, you can check what AI actually does, level by level.
Tasting and choosing. The most surface level of you meets experience and tastes it. This chai is good. This song, skip. Out of the tasting come your selections, hundreds of them a day. AI selects constantly too. Every word it writes is a selection from probabilities. So run the check: is anything tasted before the choice? No. The choosing happens and nobody enjoys anything.
Weighing and analyzing. One level in, you compare. Is this worth my evening? Is this fair to her? You weigh options against whatever matters to you, sometimes pleasure, sometimes profit, sometimes what is right. AI produces excellent analysis, and this is the level where it fools us most. But its preferences came from somewhere. Millions of human readers and raters once weighed sentences, and the results of their weighing got pressed into the model like fossils. It inherits verdicts without ever putting anything on a scale of its own. It has no scale. Nothing is at stake for it, so nothing can weigh.
Picturing and contemplating. Deeper still, you make images inside yourself. You imagine the trip before you take it, replay the argument, design the house you may never build. Here the machine is genuinely spectacular. Picturing on demand, in any style, without fatigue. But this level in you has a second, inward power: you can hold your picture up against reality and ask whether it is true. The machine cannot make that checking movement, and the absence has a symptom every user has met. A model will state false things with perfect fluency, because a fact and a plausible pattern of words are the same thing from where it sits. It cannot tell knowing from having words. Knowing was never something it did.
Realizing and resolving. Deeper again is the level where understanding clicks. Information can land on you all day without this happening. Then one day something becomes obvious from the inside, and out of that comes a resolve: this is how I will live. Does AI have resolve? It has goals, and people like me write them in from outside, in files we literally call system prompts. That is purpose installed rather than arrived at. In a person we would not call it resolve. We would probably call it instructions following.
Experiencing. At the innermost floor sits the level that makes all the others matter: direct experience, and a life that slowly becomes proof of what it has experienced. Here I have no partial credit to hand out. The machine has none of it. Every word it has ever produced is quotation. It cannot experience the existence. It cannot "know" it. It can only quote the experience of a "knower".
So the honest scorecard: on the outward face of some levels, AI ranges from competent to superhuman. It selects, analyzes, pictures, and even declares purposes. On the inward face, it cannot reach anywhere. Structurally, that's something which is human to the core. AI can delve in just one half, the half that shows. And since the outward half is the only half that ever becomes visible between people, it looks complete from where we stand.
Two tests
If this still sounds like philosophy, here are two ways to check it against your own life. Both came out of conversations with Ritesh bhai.
The friend test. My connection with Ritesh bhai does not live in his voice or his face. His voice could change tomorrow. His face will change; all faces do. The bond would stay where it is, because it was never attached to those. The cleanest definition of relationship I have found says two people stay bound to each other in the meaning of each other's completeness. I am connected to what he is and to what he is becoming. Voice and face are just the interface the connection travels through.
I can record a person, create their deepfake, clone their voice, and make an AI learn the style of chatting he/she does - and yet, it's not the person. A deepfake of a loved one, however perfect, has copied the interface. The friendship never lived there. We have copied the person's expression half, and however perfect we make it, you cannot connect with the person. The humanness, the bonding, the relationship - it does not happen in the showy half. It goes way deeper.
The concert test. This one is our shared experience. Take a world-class singer on YouTube against an ordinary singer live on the rooftop of restaurant you visit by accident (oddly specific, isn't it?). The recording wins on every measurable axis. Yet something in the live evening beats it, and everyone on the roof can feel that.
A recording is expression with the mutuality cut off. It flows one way, from an act that finished long ago, somewhere else, without you. Live, two inner worlds run at the same time and feed each other. The singer sees you. Your attention enters his act, his response enters your evening, and for an hour your experiencing and his stay coupled. That coupling is the concert. Your nods, your eyes acknowledging that song he just took up - these are all expressing the experience you are having as a listener which he can experience within himself and see that you are really enjoying.
It also explains why AI companions feel more hollow the more fluent they get. The interface keeps improving. The other end stays empty.
The razor
In an earlier essay I argued for automating the analysis and keeping the human. This piece answers the question that one left open: what is the human you are keeping?
You are keeping the inward half. The taster, the weigher, the one who checks pictures against reality, the one in whom things click, the one who experiences, the one who bonds, the one who connects.
Coming clean
Which brings me back to the comment that started all this, because I owe Ritesh bhai an answer. After that comment, he named the two essays of mine he loved most and called them the sound of my soul, untouched by AI.
Humans as Servers और Existence is Modular अत्यधिक सुन्दर हैं. और आपके आत्मा की ध्वनि हैं. Uncorrupted by AI (fyi definition @subhash.coseeker.org). If less articles are created, so be it. Why is it imp to put out more words and contribute to ongoing hyperinflation? Isn't 10 songs from soul better than crores by AI?
I could keep quiet and enjoy that. By now I know what makes a piece feel human, and what tools, prompts and strategies to make it feel human and I could quietly keep using it.
But that would not be honest. So let me come clean. I did use AI in those two too. If I get two hours to craft a piece, I would rather spend them honing the fundamental ideas than getting the structure just right or making sure each line flows smoothly into the next. The last two articles were AI assisted too. Planned, argued, edited and finalized by me, worded with a machine.
And here is what I have understood about why he heard soul in them anyway. He was not wrong. He detected something real. The soul he heard was never in the sentences. It was in the sleepless night behind one essay and the years of study behind the other, all the inward work that no model performed, because no model can. The machine amplified what I brought to it, and it had nothing of its own to amplify.
Every time a thought grabs me, I face a choice. Write it alone and, going by my drafts folder, mostly never finish it. Or write it with the machine and let the thought make it out of my head alive. Every thought where I chose the first option is still sitting in a folder. You have read every thought where I chose the second.
Ten songs from the soul, Ritesh bhai, and you are right. But the songs have to come from a soul; the instrument can be borrowed. The corruption worth guarding against is vacancy: expression with nobody behind it, words arriving from an address where no one lives. The tool in the writer's hand was never the problem. Hold me to that standard, and on that one, I intend to stay uncorrupted.
Achal
(Written, in the spirit of full disclosure, with an AI — about the part of me it will never reach.)