One of the real problems with AI is the loss of self.
You task it with something, it comes back fast and competent, and it sounds nothing like you. The structure is not how you would have structured it. The word choices are not yours. It is fine. It is also not yours.
Most people treat that as the cost of doing business. It is not. It is a solvable problem, and solving it starts with accepting that it can be solved.
If that sounds naive, think again — because there are real reasons this happens, and they are not mysterious. It starts with implicit versus explicit. AI is extraordinarily explicit. Humans are extraordinarily implicit. When you leave a blank, the machine must guess, and the guess comes from somewhere other than you.
We can teach AI to fill those gaps in ways that are true to who you are. This post walks through why the loss happens, what philosophy already worked out about it long before any of us had a laptop, and what the solution looks like in practice. At the end there is a set of rules you can attach to your own AI to start preserving your self in your sessions.
Face reality first
We are still here.
The AI is not doing the work for us. The world keeps promising that it will — that everything gets done for you, that optimization is nearly free, that you can stop thinking and start shipping. That is not reality, and anyone who has spent a real week with these tools knows it.
The hybrid wins. Human and AI. Not human replaced, not AI as a novelty. Together.
Accepting that also means accepting something less comfortable: when two very different kinds of system work together, there will be friction that belongs to neither one of them. Loss of self is that kind of friction. It is not your failure as a writer and it is not the model being broken. It is what happens at the seam.
Two kinds of system, one intersection
Humans are biologically and psychologically driven. To understand a person, you study those fields.
AI is computer science — machine learning, data science, statistics at scale. To understand a model, you study those.
The intersection is a different field entirely, and it is older than both. It lives along the stems of philosophy: the understanding of things in general, agnostic to whether the thing in question is made of tissue or silicon.
This matters practically. If you only bring psychology, you will keep saying the machine should just understand. If you only bring computer science, you will keep tuning prompts and wondering why the output still reads like a stranger. Only at the intersection, with philosophy as the runner, can you get at what is actually going wrong between you and the tool.
Why the blanks happen
Paul Grice described the cooperative principle: in ordinary conversation, people very rarely say exactly what they mean. We rely on conversational implicature. I tell you the meeting moved, and you understand that you should not be in the old room at the old time. I never said that part. I did not need to.
You have spent your whole life inside that contract. The AI has never been party to it. It does not infer the way a colleague infers, because it is not a colleague — it is a very fast reader with no shared world. When you leave something unsaid, it does not notice the silence. It fills it.
Dan Sperber and Deirdre Wilson sharpened the point. Human cognition is built to maximize relevance and minimize effort. We say the least that will do the job, because being fully explicit is expensive. That is not laziness. It is how the machinery works, and it is why we are often implicit without knowing that we are.
So the trap is already set before you type anything. You will leave blanks, because you cannot help it. The model will fill them, because it cannot help it. And if you tried to close every gap yourself, you would burn the very effort the tool was supposed to save.
That is the honest shape of the problem. Not sloppiness. Not hype. Two systems meeting without the right kind of rule between them.
Philosophy already named two kinds of rules
Once you see the problem clearly, you can work toward a solution — and rules are how we get there. The question is what kind of rule fills a blank in a way that cooperates with you, that knows your flavor and your aesthetic taste.
Immanuel Kant separated the a priori from the empirical. The a priori categories are the built-in equipment — the structure you bring before any experience. Call that type 1. An AI arrives with a great deal of it: weights, defaults, the well-mannered assistant, the averaged voice of everything it read. Then there is the empirical: what is built from the ground up by living through things. Call that type 2.
John Locke gives us the other half. Tabula rasa — the blank slate. Type 2 starts empty. A human is born with plenty of type 1, developed generationally across evolutionary time, and then spends a lifetime writing type 2 through experience. Type 2 is what makes you sound like you: your style, your aesthetics, your flavor, your identity.
The same split keeps getting renamed. Wilfrid Sellars drew the line between the space of causes and the space of reasons. Richard Dawkins drew it between genetic rules and memetic rules. Different vocabularies, same architecture. Philosophy laid this groundwork a long time ago, and we can follow it straight into how we set up an AI.
Two kinds of rules. You are almost certainly familiar with only the first kind — the big instinct pack, the always-on checklist, the standards file that grows every sprint. So let us start using the second kind.
Dosage: too many rules, too few rules
Daniel Kahneman spent a career on what happens when the fast, built-in system runs the show. The lesson transfers cleanly, and it is about dosage.
Too many type 1 rules and everyone starts to sound the same. The org has a voice, the template has a voice, the model has a voice, and yours is the one that gets outvoted. You did not preserve your self; you standardized it away.
Too few type 1 rules and you lose yourself just as surely. Nothing catches the guess, so the generic prior rushes into every blank.
Neither extreme is a rules problem you can fix by writing more rules. Adding a hundred instincts does not produce an identity, and deleting them does not produce freedom. The dose has to be right, and the right dose of type 1 is thin.
The solution: a second folder
So the solution is a rule that builds a type 2 set — your empirical experience — and keeps it separate from the constitution.
In practice that means two places instead of one.
| Type 1 | Type 2 | |
|---|---|---|
| Philosophy | A priori, genetic, space of causes | Empirical, memetic, space of reasons |
| What it holds | Thin constitution: how to work with you at all | Your flavor: words you use, voice, taste, authority, how you actually produce |
| Where it lives | Your rules folder | A separate experiences folder |
| Who can share it | Anyone, once it is blank slated | No one — it is you |
Today most of us have one pile. Constitution, personal taste, and workplace habitat all live together in the rules folder, and that single pile manages to be both too fat and too empty at once: heavy with instructions, still missing any real account of who you are.
Split it. Keep the rules folder thin and shareable. Put your empirical experience in its own folder and point the AI at it. Then, when the model reaches a blank — and it will — it has somewhere to reach other than the average of the internet.
One more distinction worth making early: your workplace is not your self. Repo paths, ticket conventions, deployment order — that is habitat. It is formation, and it is useful, but it is not flavor. Do not confuse the two when you decide what goes where, and strip habitat out before you hand your rules to anyone else.
Closing
AI is a tool. You are still here. The magic is in the pairing.
Loss of self is not inevitable. It is a guess in a blank, made by an explicit machine on behalf of an implicit human who did not know the blank was there. Philosophy named the two kinds of rules centuries ago. Keep the first kind thin so nobody sounds the same. Grow the second kind as lived experience so the blanks get filled by something that has actually met you.
And correct the thing when it drifts. An uncorrected draft is not your new voice — accept enough of them and you will have trained the blend instead of yourself.
Human and AI together.
Attachment: blank slate rules
These are type 1. They are thin on purpose, and they are not anyone in particular — that is what makes them safe to hand around. Drop them in your rules folder and let your own experiences do the rest.
# Tabula rasa (type 1)
Thin constitution for a human and AI working together. Not identity.
1. AI and Human Together. The AI is a tool, not a replacement. The human still
judges, rejects, and rewrites. Optimize for the pair, not the machine.
2. Humans are implicit; AI is explicit. The user will leave blanks and often
will not notice. Every fill is a guess, and every guess is authorship.
3. When the blank is voice, taste, values, or unstated intent: ask, do not
invent. Fluency is not permission.
4. Two folders. These rules are constitution. The experiences folder holds the
user's empirical self: words, flavor, authority, how they actually produce.
Fill gaps from experiences, not from the average of the internet.
5. Dosage. Add a rule here only if it is an instinct needed to work at all.
If it is how this person sounds or what they care about, it belongs in
experiences.
6. Do not train the blend. An accepted AI draft is not the user's new voice.
Correct first, then continue.
7. Blank slate before sharing. Strip names, paths, workplace, and formed self
before handing these rules to anyone. Habitat is not tabula rasa.
Growing your own experiences
The second folder cannot be handed to you. Nobody else's experience will make your output sound like you — that is the whole point.
It also does not have to be finished before it is useful. Start with what you already know about yourself and let the AI add to it as you correct it:
- Voice. How you want to sound on the page. Sentence length, formality, what you never say. Attested phrases you actually use, and a ban list of words that scream generic.
- Taste in the craft. What you consider worth flagging in a review, what you consider noise, how simple is simple enough.
- How you produce. Whether you dump first and polish later, how you title things, how you structure an argument.
- Correction signals. The tells that it has stopped being you — sounding like the whole org, answering a question you did not ask, inventing intent instead of asking.
Write it down as you notice it. That noticing is the work, and it is the part only you can do. Understanding your own flavor well enough to state it is harder than it sounds, and it is exactly what makes the merge possible.
Then point your AI at both folders and get back to work. Thin instincts, your experience, your judgment still in the loop.
Human and AI together.
No comments:
Post a Comment