Greg Brockman's journal became Exhibit 161 this week.
The next chapter writes itself.
Someone's ChatGPT history becomes Exhibit 162.
That sentence sounds like speculation.
It isn't.
The infrastructure is already in place.
The court orders are already in place.
The only thing missing is a famous enough defendant for the headline to break the way Brockman's did.
The court order most people haven't read
In May 2025, Magistrate Judge Ona Wang in the Southern District of New York issued a preservation order against OpenAI in the New York Times copyright lawsuit.
The order required OpenAI to preserve every ChatGPT conversation generated.
Including conversations users had explicitly deleted.
Including conversations from temporary chat mode.
OpenAI fought it.
Argued it violated user privacy.
Argued it was technically infeasible.
Lost.
The blanket retention order was eventually narrowed, and OpenAI's obligations under that specific order ended in September 2025.
That's not the part that should worry you.
In December 2025, the same judge ordered OpenAI to produce twenty million de-identified ChatGPT logs to the copyright plaintiffs.
Twenty million actual user conversations.
Handed over to opposing counsel.
In January 2026, District Judge Sidney H. Stein affirmed the order.
OpenAI's appeal failed.
The court noted, in passing, that twenty million logs represent a small fraction of the "tens of billions" of consumer ChatGPT conversations OpenAI retains.
Tens of billions.
That's not a typo. That's OpenAI's own number, in OpenAI's own filings, in a federal courtroom.
Every conversation you've had with ChatGPT — every "I need to vent about my cofounder," every draft of a hard email, every "help me think through whether to fire this person" — is sitting on a server somewhere.
Indexed. Recoverable.
The first batch of twenty million is already on its way to a courtroom.
Anthropic is fighting parallel battles in the Music Publishers and authors cases.
Every major US-incorporated AI lab operates under a baseline assumption that their data may be subpoenaed.
Their terms of service are explicit about this.
We just don't read terms of service.
ChatGPT Enterprise was carved out of the preservation order.
Zero Data Retention API agreements were carved out.
Everything else — Free, Plus, Pro, Team, the consumer app you actually use — was in the bucket.
Incognito mode doesn't fix it.
Temporary chat doesn't fix it.
Deleting your history doesn't fix it.
Those features control what the model sees in future training and what the UI displays back to you.
They do not control what's recoverable from the company's servers under court order.
How casual we are about this
I noticed it in myself first.
I've used Claude as a thinking partner for two years.
Drafted tweets in it. Processed news in it. Worked through hard decisions in it.
Talked through situations involving named people, named companies, named contracts.
I've treated the chat window the way I'd treat a notebook.
The chat window is not a notebook.
A notebook is a physical object you control.
A chat window is a network request to a third-party server, logged, retained, and bound by whatever preservation order is currently active in whatever lawsuit the company is currently fighting.
We are all doing this.
Founders processing competitor strategy.
Executives venting about their boards.
Lawyers sketching arguments before they file.
Engineers debugging code that contains proprietary architecture.
People navigating divorces.
People navigating diagnoses.
The most candid record of what you actually think — generated in the most discoverable possible format — sitting on a server you don't own, governed by terms you didn't read, subject to legal processes you'll never see coming.
The Brockman case is the trial run.
The next one will be an AI conversation.
What discovery actually looks like
In a lawsuit, discovery gives the opposing party access to your relevant communications.
"Relevant" is interpreted broadly.
Sue a competitor? Your conversations about that competitor are likely producible.
Sued by a former employee? Your conversations about that employee are likely producible.
Investigated by a regulator? Conversations about the underlying conduct are likely producible.
You will not be the one producing them.
The AI lab will.
Their legal team gets the subpoena.
Their engineers run the query.
Your conversations show up in a PDF on a lawyer's desk.
You find out when your own lawyer calls.
The defense "I deleted that conversation" does not work.
The defense "that was a private mode" does not work.
The defense "I was just thinking out loud" does not work.
The standard for written records in litigation is that they exist or they don't.
If they exist on a server somewhere, they exist in evidence.
The math nobody runs
Most of us treat AI tools the way Brockman treated his journal.
We type the candid version because the candid version is what makes the tool useful.
The whole point of thinking out loud with a model is that you can say what you actually think, get a real response, iterate.
If you're finding this useful, I send essays like this 2-3x per week.
·No spam
Self-censoring kills the utility.
But "self-censor" and "use a different surface" are not the same thing.
The framework, after sitting with this for a while, is three tiers:
Tier 1: Public-shaped thinking.
Cloud AI is fine for this.
Drafting tweets, analyzing news, coding, research, anything you'd be comfortable seeing on a screen in a courtroom three years from now.
The bar is simple.
Would I post this? If yes, type it.
Tier 2: Sensitive thinking.
Local AI.
A model running on hardware you own, on disk you control, with no network sync to anyone else's server.
This is where candid analysis of competitors goes.
Where candid processing of live disputes goes.
Where the version of the thought that has the value precisely because it's unfiltered goes.
Logs live on your machine.
Not subpoena-proof if your laptop is seized in a criminal matter. But no third-party preservation order touches it.
The much more common discovery vector is closed off entirely.
Tier 3: No keyboard at all.
Verbal with a trusted partner.
Privileged conversation with a lawyer.
Or just thought through alone.
Some thoughts shouldn't exist as text anywhere.
Brockman's mistake wasn't journaling.
It was generating a written record where one wasn't required.
What I'm doing about it
I'm allocating real mental bandwidth and real money to Tier 2.
Sovereign AI is not a thesis I'm writing about anymore.
It's a thing I'm building for myself, in my own office, on my own hardware, before I write another word about it for anyone else.
Two A100s.
Two H200s.
One L4.
Enough compute to run frontier-class open models locally — Qwen, Llama, the new releases — at usable speed, with no network calls leaving the building.
The inference loop runs on hardware I own.
The logs live on disk I encrypt.
The preservation order from someone else's lawsuit reaches exactly zero of it.
I tried to buy a Mac Studio for the lighter end of the workflow.
Sold out. Every configuration.
The whole supply chain is back-ordered into next quarter — because the rest of the market is figuring out the same thing I am.
There's a real reason to run inference locally.
The real reason has nothing to do with hobbyist tinkering.
So I'm using a MacBook Pro for now.
M4 Max, 128GB unified memory.
Runs a 70B model at usable speed.
Not as much headroom as the Studio would have given me.
But it works.
"Works now" beats "perfect later" every time.
The setup isn't elegant yet.
The cluster needs work. The local model isn't as good as Claude is at most things.
The UX is worse. The latency is worse. The capability ceiling is lower.
None of that matters.
What matters is that the conversations I'd most regret being read by a stranger are now happening in a place where no stranger can read them.
This is what sovereignty actually looks like in practice.
Not a tweet.
Not a thesis.
A rack of GPUs, a notebook full of disk encryption commands, and a deliberate choice about which thoughts go on which surface.
The frame this all fits into
Brockman's diary became evidence because he wrote candid thoughts about a named adversary in a tool that synced to a server.
The next decade is going to produce a thousand versions of this story.
The diary is now ChatGPT.
The synced server is now an AI lab.
The adversary is whoever ends up across from you in a courtroom, a regulatory action, a divorce, a board fight, a criminal matter you didn't see coming.
You don't get to choose whether the data exists.
You only get to choose where the thinking happens.
Own the hardware.
Own the model.
Or someone else owns your thoughts.