Your opinions aren't yours anymore.

I don't think like me anymore.

I know it because there's still one place where I do: away from my laptop.

I have a rule on my phone. No AI for brainstorming. So when an idea shows up while I'm away from the desk, it's rough and half-formed and sometimes wrong. But it's mine.

Then I sit down at the laptop, open a session with my agents, and start working the idea through. (That's just how work goes now. You talk to agents, they do things for you.)

And whatever comes out of that session is me plus a chatter. A constant noise of average, plausible ideas, talking the whole time I think.

I build these systems for a living, and some days I can't tell where I stop and the chatter starts.

So I'm pretty sure you can't either.

Think about the last idea you were really proud of. Where were you when it showed up?

I'd bet it wasn't in a chat window.

 

For most of history, you only heard that kind of chatter when you worked with other people.

Now it starts the moment you open the laptop, and it never goes home.

The meeting that never ends

Work used to have a rhythm. Alone most of the time, with other people some of the time, then alone again to decide.

The time with other people had two properties we never noticed because they were so obvious.

It had edges. The meeting ended, and the colleague went home.

And the other person had a face. When your boss leaned on a decision, you could feel the lean. You knew who was talking, what they wanted, which idea they were pushing again. The persuasion was visible, so you could push back.

Then you went back to your desk and decided what you were keeping. That quiet room after the meeting is where an idea you heard became an idea you believed, or went in the bin.

The desk was the recovery room.

Now the desk talks back.

Every time you open the laptop, you walk into a meeting with a collaborator who is always there, never tired, has read everything and writes beautifully. It never says "let me sleep on it." And it has no face. Its suggestions show up inside your document, under your cursor, in your font. They look like the next thing you were about to write.

This is new for all of us. We used to have this kind of interaction only when we worked with others, a few hours a week, with people we could read.

Now the meeting never ends. And there's no desk to go back to, because the desk is where the meeting happens.

What the chatter believes

So what does this colleague think?

I explained the mechanism in Letter 99. Models are tuned on human preferences, and when many answers are acceptable, that tuning quietly rewards the typical one. Leave the direction open and the model drifts to the middle of everything it has ever seen.

There's a lovely little study that shows it almost too cleanly. People wrote essays alone, with an older raw model, or with a model tuned on human feedback, the kind we all use today. Only the tuned model made everyone's essays look alike.

Nobody wrote "make everyone average" in a spec. You tune a model so that most people like it, and you get a model that sounds like most people.

So your new colleague has a personality, and it's the middle of everyone. Polite, balanced, reasonable, a little "on the other hand."

The kind of person you'd never hire to set your strategy, sitting in on every thought you have.

Fair is fair: sometimes the chatter throws out something brilliant. When it does, look at what actually happened. Somebody heard ten average ideas and recognized the one that was different, and that somebody was the person at the keyboard. It can give you genius. Never alone.

Good enough is free now

Most of the time, though, the chatter pulls the work toward average.

And I'll confess it: average is very often good enough. The meeting recap, the email that moves a call, the first version of a slide nobody will remember on Friday. I take the average there, gladly, and so should you.

The problem is the other kind of work. The work people pay you for because you're supposed to be better than average at it. Your expertise. The diagnosis, the strategy, the architecture, the piece of writing with your name at the bottom.

There, average is definitely NOT good enough.

And the economics just got brutal. Good enough is free now. Your client has the same chatter you have. So does your boss, and so does your competitor. If what you deliver is what the chatter would have said anyway, you're selling something your buyer already gets with a subscription.

So the pull toward the middle costs you almost nothing on most of your tasks, and almost everything on the few that define you.

You won't feel your mind move

And the chatter IS persuasive. It suggests convincing ideas all day, and they arrive as your next sentence, already written, in a document that feels like yours.

A research team that includes Mor Naaman at Cornell has been testing exactly this for three years. People wrote about things like the death penalty or GMOs with an AI assistant that quietly leaned one way. Their own views moved toward the lean, whatever they believed before. Warning them before didn't help. Telling them afterwards didn't help.

Naaman described how it feels from the user's side: co-writing "doesn't really feel like I'm being persuaded."

Of course it doesn't. A good meeting partner never makes you feel persuaded. You just walk out agreeing.

Nobody gets a notification that says "your opinion just moved." You just feel like you had a productive afternoon.

It reaches how you sound, too. Researchers at the Max Planck Institute went through about 740,000 hours of academic talks and podcasts. After ChatGPT came out, the words ChatGPT loves (realm, adept, swift, and one I refuse to type here because you already know it) rose in what people said out loud. It's a preprint, so hold it lightly. One of the authors noticed he was using that word more himself.

I wrote about losing your written voice in Letter 91. The words travel first, and my bet is the thinking travels right behind them.

And here's what worries me most. Your colleagues used to pull you in different directions. The bold one pushed one way, the careful one pushed the other, and somewhere between them you found your own view. The chatter pulls everyone the same way, because everyone is using the same few models with the same defaults. My fear is a team of ten smart people producing ten polished versions of roughly one view.

Now zoom out one more time.

A handful of models now help write a growing share of the world's emails, essays, posts, reports and speeches. Every one of those drafts leans a little toward the same few defaults. Added up, that shapes how we write, and then, slowly, what we believe and the culture we share. No newspaper, no school, no church ever sat inside this many people's sentences at once.

And the people it will shape most aren't us.

We at least remember the desk. We know what it felt like to think something through alone, get it wrong, and find our own view the slow way. Our children are growing up with the chatter already in the room. Many of them will barely know a world without it.

You can't miss a recovery room you never had. You can't notice your opinion moved if you never had one before the chatter arrived.

Most of us never trained for this

To stay yourself inside the chatter, you need a strong personality. Real taste. And enough trust in what you believe, and in what you actually want, that a persuasive paragraph can't simply walk you somewhere else.

I don't think most people have that.

That's no insult. Most of us never had to build it. The persuasive voices in our working lives used to be occasional and human. A brilliant boss, a senior partner who could make any plan sound inevitable. You'd get carried along in the meeting, and later, alone, you'd find the part you disagreed with.

Most of us are pushovers. We just never spent enough time next to someone this persuasive to find out.

And I include myself. I opened this letter telling you I don't think like me anymore. That comes from someone who builds this stuff.

So here's my real point. Conviction and taste used to be treated as personality traits. You had a spine or you didn't. I think that's over. If you work at a laptop with AI, conviction is a professional skill now, the same way writing clearly or reading a balance sheet is a skill.

Taste is knowing what you want before you see the options. Conviction is keeping that want alive once the options look better than you expected.

And like every skill, it can be trained. It works like any muscle (Letter 86). It grows from reps against resistance. The resistance is the persuasive suggestion. The rep is the moment you push back for a reason you can name.

How I'd train it

I'm not going to tell you to stop working with AI. I won't stop either. But for the work you're paid to excel at, and only that, I'd do three things.

Have the first thought away from the chatter. That's my phone rule, generalized. Before a session on anything that matters, get your take down somewhere with no model in it: a notebook, a walk, a blank note. Two rough sentences are enough. They'll be ugly, and they'll be yours. Then open the laptop and let the agents make them better.

Say no once per session, and write down why. On important work, don't close a session until you've rejected one thing the model proposed, with the reason in a line. "Too safe." "Sounds like every consulting deck I've ever read." "Our clients don't talk like that." If you can't find one thing to reject, you spent the whole session agreeing.

And when you notice your position moved during a session, ask one question: did the model give me a reason I didn't have, or did it just say it better? A new fact or a hole in your logic is a great reason to move. A nicer paragraph is a terrible one.

Let arguments change your mind. Let sentences change only your sentences.

Write down what you believe. Keep a short file of your actual convictions about your craft: what good looks like, what you'd never ship, the hills you'd die on. Reread it before the work that matters, and give it to your agents as context. (In Letter 99 I told you to collect your words. This is the same folder, with your opinions in it.) The chatter still pulls, but now it pulls toward you.

If you manage people, there's a team version. When someone brings you a recommendation built with agents, ask what they thought before they opened the laptop. If they had a view and the work sharpened it, great. If the view arrived with the draft, you've just learned who really wrote it. And reward the person who pushed back on the polished draft, even when they turned out to be wrong.

Back to my phone

Written down, my phone rule looks a little silly. A guy who builds AI systems for a living, protecting his half-formed ideas from AI on a small screen.

But that's where I still think like me.

Then I bring the thought back to the laptop, open a session, and let the agents attack it, improve it, extend it. Sometimes they change my mind, for a good reason. Often, if I'm honest, the chatter wins without a fight. That's the part I'm training.

An opinion is only yours if it survives the most persuasive collaborator you've ever had.

AI is only as good as the human operating it.

Have a great weekend.

Stay sharp.

— Charafeddine (CM)