Last week, Anthropic refused a Pentagon demand. The government wanted to remove two specific restrictions from their AI contract: no mass domestic surveillance of American citizens, and no AI systems directing lethal weapons without a human in the loop. Anthropic said no. The Pentagon gave them a 5:01 PM deadline to comply. When it passed, President Trump ordered every federal agency to immediately stop using Anthropic’s technology. Defense Secretary Pete Hegseth — who now prefers “Secretary of War,” a rebranding that cost taxpayers up to $125 million and tells you most of what you need to know about the current posture — designated Anthropic a “Supply-Chain Risk to National Security.” A label previously reserved for foreign adversaries like Huawei. Never before applied to an American company.
Within hours, OpenAI announced they had signed the replacement contract.
That’s the story. You probably already knew most of it.
Here’s the part where it stops being news and starts being a decision.
This Isn’t a Compliment
Anthropic did the right thing. Full stop. Dario Amodei said they “cannot in good conscience” allow their tools to be used for mass surveillance or autonomous killing, and they took the consequences. That took something.
But notice what we’re actually saying when we call it remarkable: we live in a world where a CEO choosing not to enable surveillance of American citizens is newsworthy. Where holding that line costs you a $200 million contract and gets you labeled a national security threat. Where the response from the top of the U.S. government was, essentially, how dare you have terms of service.
This is not a story about a good company. It’s a story about what we’re left depending on when everything above them has already decided which direction it’s going.
I have known a lot of CEOs. I’ve worked with them, reported to them, competed against them. Most of them are smart, driven, and primarily focused on not losing. The ones who hold the line when it’s expensive are rare enough that we notice them when they appear. Building your civil liberties on that foundation — on the hope that this particular person, at this particular company, at this particular moment, happens to have good values and the board to back them up — is not a strategy. It’s a wish.
About That Other Deal
Sam Altman announced OpenAI’s replacement contract the same evening Anthropic was blacklisted. He said OpenAI shared Anthropic’s “red lines” — no mass surveillance, no autonomous lethal weapons. He described the deal as including the same safeguards Anthropic had been denied.
Altman later admitted the timing made it look “opportunistic and sloppy.” We’ll take him at his word on that.
Since then, OpenAI’s head of robotics, Caitlin Kalinowski, has resigned on principle. Her statement was direct: “Surveillance of Americans without judicial oversight and lethal autonomy without human authorization are lines that deserved more deliberation than they got.” She said it was “about principle, not people” — which is a careful way of saying the problem isn’t one executive’s judgment. It’s the decision itself. (NPR)
What we also have is Anthropic’s assessment of OpenAI’s actual contract language: that it was “paired with legalese that would allow those safeguards to be disregarded at will.” Meaning the protections existed on the surface. The exits were built in underneath.
Nobody has explained why the Pentagon agreed to accommodate OpenAI and not Anthropic. The Pentagon’s position throughout the dispute was that Anthropic shouldn’t be able to dictate how its technology is used — that “legality is the Pentagon’s responsibility as the end user.” That position didn’t change. What changed was the company they were negotiating with.
Draw your own conclusions. The facts are sufficient.
The Lever You Actually Have
Elected officials aren’t going to fix this. Not right now. The administration that threatened to invoke the Defense Production Act to force a private company to remove safety guardrails is not a government currently interested in protecting your digital civil liberties. That’s not a partisan point — it’s just an accurate description of what happened last week.
What that leaves is where you spend your money.
That sounds small. It isn’t. According to market intelligence firm Sensor Tower, ChatGPT uninstalls jumped 295% day-over-day the Saturday after OpenAI’s deal was announced — compared to its typical daily rate of 9%. One-star reviews surged 775% the same day. Claude jumped to number one on the Apple App Store, surpassing ChatGPT in US downloads for the first time in its history. That’s not symbolic — that’s revenue signal, and revenue signal is one of the few things that has a chance of changing behavior at this level. These companies are not governments. They answer to money, and some of them answer to users. When those two things point in the same direction, the effect compounds.
If you’re a paying ChatGPT subscriber, that money is now flowing to a company whose CEO admitted he rushed to sign a surveillance contract while a competitor was being blacklisted for refusing it. You may be fine with that. But it should be a conscious choice, not a default.
Your Context Belongs to You
The months of conversations you’ve had with an AI system aren’t just chat logs. They’re context — your preferred communication style, your ongoing projects, your way of framing problems, your history. That’s real value that these platforms benefit from and that you’ve built.
You can take it with you.
Here’s how to do it, depending on how much you want to carry over.
If you have 60 seconds: use Anthropic’s transfer tool
Anthropic built this specifically for the moment we’re in. It works for ChatGPT, Gemini, and Copilot.
Go to claude.ai → Settings → Capabilities → Memory → Import
Copy the prompt Anthropic provides there
Paste it into ChatGPT — it will output everything it knows about you in a single code block
Copy that block back into Claude’s import field and hit Add to Memory
Claude absorbs it within about 24 hours
Available on every Claude plan including free.
If you want your full history: export the archive
ChatGPT lets you download everything — every conversation, every prompt, every response.
In ChatGPT: Settings → Data Controls → Export Data
You’ll receive an email with a .zip file — download and unzip it
Inside is a conversations.json file with your complete history
Create a Claude Project and upload the file — Claude can now reference your full chat history as context
The export can be large (several gigabytes for heavy users), so Claude Projects works better here than basic memory. But it’s all there, and it’s yours.
If you want something that lasts: build portable memory
The two methods above move your data to Claude. This one makes your data belong to you — platform-independent, updatable, and usable anywhere.
My setup is a structured memory file that captures my projects, preferences, decisions, and working context in plain text. I update it periodically and carry it into any AI system I use — Claude, local models, whatever comes next. It lives in my knowledge base alongside my notes. When I open a new conversation anywhere, that context comes with me. Not because a platform stored it. Because I did.
That’s the difference. Platform memory is theirs. Portable memory is yours.
The full breakdown — how to structure that file, how to maintain it, and how I use it across tools — is coming in a dedicated guide. Subscribe below if you want that when it drops.
Find One Other Person
The off-ramp is easier with company. Not a movement — just one or two people who are thinking about the same things and willing to compare notes. Someone to troubleshoot with when the export process is confusing. Someone who will text you when they find a better way to structure their memory files. Someone who makes it a project instead of a chore.
We’ve written about this before — in 7 Steps to Quietly Exit a System That Wants You Dependent, and in The Imagination Heist, where the argument was that the first thing these systems take isn’t your time or your data. It’s your ability to picture a different arrangement. The portability work is partly technical. But it’s also about refusing to let any single platform own your context, your history, the shape of how you think through problems.
That’s worth doing with someone.
Find your person. Walk through the three tiers above together. Pick a weekend afternoon. Your data will still be there when you’re done, and so will your ability to take it somewhere else.
That’s the vote. It’s a small one. It counts.
Related reading:
My argument on this is relatively simple. If you can't hard-code your guardrails into the system, and must rely on contract disputes to enforce a principle, then the technology is not ready for deployment in that domain.
If you go ask Claude, or ChatGPT, or any of the LLM'S out there, they will all say the same thing. They are not capable of hard-coding these kinds of safeties into the models, because the system isn't smart enough to understand what downstream effects there are from its outputs. It is not capable of knowing where it sits in the pipeline of violence. And so coding such limits would render the models operationally useless, because most military language is already violence-tinged.
So if the software can't be guardrailed, and contracts are so easy to subvert, or negotiate, or outright replace, then there are no guardrails at all. Not really. There's only lip-service.
Please can everyone stop "praising" Anthropic. Claude is still built on unethical theft of data, extractive practices and exploitation of the global south. Whether intentional or not, their "principled" stand was a marketing campaign, picking up people jumping ship from the supposed baddy.