You’re Not Using AI. You’re Building Someone Else’s Moat.
Your prompts may be private. Your company’s data may be protected.
But the models, memory, pricing, rules, and intelligence layer still belong to someone else.
Open the AI tool you use most.
Now look through your history.
Your business ideas are there.
Your customer research.
Your unfinished products.
Your writing style.
Your code.
Your negotiations.
Your mistakes.
Your best prompts.
Your most valuable decisions.
Months of thinking are sitting inside that account.
It feels like an asset.
But it isn’t an asset.
It is access.
An asset can be exported, inspected, backed up, moved, modified, and operated under your rules.
Access can be limited.
Repriced.
Redesigned.
Suspended.
Or removed completely.
That is the uncomfortable reality behind the AI boom:
We are putting more of our intelligence inside systems we do not own.
And almost nobody is preparing for what happens when those systems change.
ANNOUNCEMENT
We're stepping up to three newsletters a week. It's a deliberate choice: enough volume to stay useful, enough care to stay worth your time.
The Warning Was Bigger Than the Soundbite
His argument was aggressive.
Businesses are spending heavily on tokens while surrendering control over the infrastructure that increasingly runs their operations.
They are buying access to intelligence instead of building durable intelligence of their own.
Karp is not a neutral observer. Palantir is actively selling sovereign AI infrastructure, including systems designed to deploy NVIDIA open models inside controlled environments.
But commercial interest does not automatically make the underlying question irrelevant.
The question is brutally simple:
Who owns the intelligence layer your business is being built on?
Not who owns the documents.
Not who legally owns the generated output.
Who owns the system that remembers, reasons, prices, restricts, and decides how that intelligence can be accessed?
For most individuals and companies, the answer is still:
Someone else.
Let’s Correct the Most Convenient Exaggeration
This conversation becomes useless when people reduce it to:
“AI companies are secretly training on every document you upload.”
That is not an accurate description of every product or plan.
OpenAI says that content from individual services, such as ChatGPT, may be used to improve its models, but users can turn this off. It also says business products and API data are not used for training by default.
Anthropic says consumer chats and Claude Code sessions may be used when users enable model improvement, explicitly opt in, or when conversations are selected for safety review. Its commercial products are not used for model training by default.
These distinctions matter.
A personal AI subscription is not governed identically to an enterprise workspace or API agreement.
But training protection does not equal ownership.
Even when your content is not used to train a foundation model, you are still depending on infrastructure that you cannot fully control.
The provider controls:
Which models remain available
How much does access cost
How memory functions
Which actions are permitted
How usage limits are calculated
Which integrations survive
What happens when policies change
Whether your workflow performs the same way tomorrow
Your data may remain yours.
But the environment that makes your data intelligent still belongs to the provider.
That is the deeper dependency.
The Most Valuable Thing You Are Giving Away Is Continuity
The danger is not limited to one sensitive PDF.
It is the gradual transfer of your working memory into a rented interface.
Every day, you teach your AI assistant:
How you communicate.
What does quality mean to you?
Which customers do you serve?
Which objections keep appearing?
How your product works.
What your team gets wrong.
What have you already tried?
Which ideas did you reject?
Which decisions produced results?
Over time, this accumulated context becomes more valuable than any individual prompt.
But where does that intelligence live?
For most users, it is scattered across chat histories, temporary projects, hidden memories, custom instructions, uploaded files, and conversations tied to one provider.
There is no clean operating system.
There is no portable memory layer.
There is no reliable model-independent backup.
There is just a growing emotional belief that the AI “knows you.”
Then a model is retired.
A feature changes.
A usage limit appears.
A project loses context.
A safety policy blocks a workflow that worked yesterday.
Or the provider replaces the model underneath your existing conversation.
OpenAI’s own release notes show how frequently models can be introduced, replaced, and retired. That is normal for a fast-moving service, but dangerous for anyone who has built a critical workflow around a specific model’s behavior.
You did not lose your data.
You lost the environment that made the data useful.
That is what renting intelligence looks like.
Share this and get rewarded
Go to AI In Public → scroll to the Leaderboard section → copy your referral link → share it with anyone.
Every signup through your link moves you up the board and unlocks real rewards.
At the end of every month, we check the leaderboard.
The top 3 get our best guide - DMed directly, no opt-in required.
Your AI Account Is Not Your AI System
An account is a destination.
A system survives destinations.
Your real AI system should include:
Your knowledge.
Your operating instructions.
Your examples.
Your decision history.
Your workflows.
Your evaluations.
Your tools.
Your permissions.
Your model preferences.
Your fallback options.
And every one of those components should remain usable when the primary provider disappears.
Today, most people have built the exact opposite.
Their documents are stored in one place.
Their prompts are trapped in another.
Their automation depends on one API.
Their memory depends on one product.
Their workflow only succeeds with one model.
Their team has no idea how the system actually works.
This feels efficient right up until the day something changes.
Then they discover they did not build an AI system.
They built a collection of dependencies.
The Claude Code Controversy Revealed the Real Trust Problem
At the end of June 2026, developers examining Claude Code reported a hidden mechanism that could encode signals about proxy domains and certain user-environment characteristics into subtle variations inside the system prompt.
An independent technical analysis found that the mechanism operated when Claude Code was routed through non-Anthropic endpoints and could encode information using almost invisible character differences.
The word “spyware” became viral.
Anthropic rejected that framing and described the mechanism as a temporary anti-abuse experiment intended to detect unauthorized reselling and model-distillation activity. Reporting indicated that it was subsequently removed.
The distinction is important.
There is no credible reason to claim that Claude Code was indiscriminately spying on every user.
But the controversy still exposed something larger:
Users normally cannot see everything their AI client is doing.
You can inspect the interface.
You can read the privacy page.
You can review the settings.
But trust ultimately depends on the provider behaving exactly as expected inside infrastructure most users cannot independently audit.
That does not make every provider malicious.
It makes blind dependence reckless.
Trust is not ownership.
A policy is not architecture.
And a promise is not a fallback plan.
Owning AI Does Not Mean Abandoning ChatGPT or Claude
This is where most ownership arguments become unrealistic.
They tell everyone to cancel their subscriptions, buy expensive hardware, download a giant model, and become an infrastructure engineer.
That is not a strategy for most people.
Frontier cloud models are extraordinarily useful.
Use them.
Use the best reasoning model for strategy.
Use the best coding model for development.
Use the best image model for creative work.
Use fast, inexpensive models for repetitive operations.
But treat models as replaceable engines.
Never let one engine become the only place where your company remembers how to drive.
The goal is not to eliminate external providers.
The goal is to stop making any single provider irreplaceable.
The Five Layers You Actually Need to Own
You do not need to train a frontier model.
You need to own the context that makes frontier models valuable to you.
1. Own Your Data
Keep original documents, research, customer information, transcripts, and operating knowledge inside storage you control.
Uploading a file to an AI tool should create a temporary working copy.
It should never become the only usable copy.
2. Own Your Memory
Your company’s memory should not exist exclusively inside chat history.
Store important facts, decisions, examples, preferences, and lessons in a structured knowledge base that can be connected to multiple models.
When you change providers, your memory should come with you.
3. Own Your Workflows
Save your best system prompts, content frameworks, automation instructions, evaluation criteria, and agent configurations as files or code.
A workflow buried inside a six-month-old conversation is not an operational asset.
It is a future archaeology project.
4. Own Your Model Router
Different tasks deserve different models.
Your system should be able to send private work to a controlled model, complex reasoning to a frontier model, and repetitive work to a cheaper model.
Model loyalty is not a strategy.
Model optionality is.
5. Own a Fallback
Your fallback does not have to outperform every closed model.
It needs to keep essential work moving.
A local or privately deployed open-weight model can handle document search, classification, summarization, drafting, structured extraction, and other repeatable tasks without making your entire operation dependent on a public interface.
Palantir and NVIDIA are already promoting this architecture at the enterprise and government level: open models deployed inside controlled environments where the organization retains greater authority over the system.
Individuals and small teams need a lighter version of the same principle.
Take the AI Ownership Test
Answer these five questions honestly:
1. Can you export the knowledge your AI has accumulated about your work?
Not merely download old conversations.
Can you convert that history into structured, reusable memory?
2. Could you change model providers this week without rebuilding everything?
When switching models destroys the workflow, you do not own the workflow.
3. Are your most valuable prompts stored outside the chat interface?
Anything critical should be versioned, documented, and accessible to your team.
4. Do you know which privacy settings and commercial terms govern your account?
“ChatGPT,” “Claude,” or any other AI product may apply different rules depending on whether you use a consumer plan, business workspace, enterprise agreement, or API.
5. Could your business continue operating during a seven-day provider outage?
When the answer is no, you have discovered a concentration risk.
Score one point for every “yes.”
Five points: You are building an AI asset.
Three or four: You have leverage, but important dependencies remain.
Zero to two: You are renting an increasingly important part of your business.
The Next AI Divide Will Not Be Humans Versus Machines
It will be owners versus dependents.
Two people may use the same model.
One uses it through an isolated chat window.
The other connects it to owned data, portable memory, documented workflows, evaluations, tools, and multiple fallback models.
The first person becomes faster.
The second person compounds.
That difference will grow every month.
The winning companies will not necessarily possess the smartest model.
They will possess the best private context around whichever model is currently smartest.
The winning creators will not have the longest prompt history.
They will have an intelligence system that remembers what worked and can carry those lessons across every new model.
The winning individuals will not refuse cloud AI.
They will use it without becoming trapped by it.
Your AI Provider Should Be a Supplier, Not Your Landlord
Keep using the Frontier Labs.
Let them compete for your workloads.
Let them improve their models.
Let them lower their prices.
Let them release better tools.
But do not let them become the only place where your organization thinks.
Use their models.
Keep your memory.
Use their computer.
Keep your workflows.
Use their interfaces.
Keep your data portable.
Use their intelligence.
But own the system around it.
Because the moment your AI becomes central to how you create, decide, sell, research, and operate, ownership stops being a technical preference.
It becomes self-preservation.
The future does not belong to the person with the most AI subscriptions.
It belongs to the person who can walk away from any one of them without losing everything they have learned.
Here are a few issues that you might have missed (go check them out):
We tested it in public. Now use it in private.
I write about the systems, workflows, and ownership layers people will need to remain valuable in the AI era.
Follow @humzaakhalid and subscribe to AI in Public for the next breakdown.
_ Hamza 💙
That was the most interesting article I’ve read on Substack recently. This exact issue has been bothering me because I remember not long before ChatGPT everyone was so privacy focused and preaching about us, the users, we’re the actual product for tech companies. The extent Google had to go to in order to continue collecting data on you after many privacy and data mining related lawsuits and all of a sudden all these privacy conscious, own-your-data borderline activists had to make a difficult decision; decide if AI brought enough value and usefulness to them to completely abandon their privacy conscious habits and cause. I myself did ask that question and it wasn’t until gpt 4o that I officially sold my soul to AI labs by accepting the privacy risk in return for productivity. Showing someone the amount and specificity of data Facebook had on millions of people when you started an ad campaign used to make people’s jaws drop. The amount of personal details they had gathered, consumer profile they created for you from that personal data and ability to hone the ads to target a verrrrry specific cohort of people used to scare the shit out of people I’d demonstrate that to.
It looks like the facts haven’t changed despite it being a different vessel - we are still the product and it’s never been easier to replicate a human being if a lab wanted to use an active users data over a few months of chatting, building, problem solving, etc. So everyone should assume that the labs you use regularly, if you’re a serious user, know more about you than you yourself and any one who knows you (including loved ones you’ve known your entire life).
I myself am too reliant on the technology to take some moral high ground decision to limit use of it as much as humanly possible. I’m definitely hooked and it’s going to be hard to out the genie back in the bottle for most people who’ve actually found the extraordinary benefit or value or leverage it brings them to every consider taking in the task of thinking for themselves again. I myself, routinely, reach for my phone the second I am even remotely curious by something and I later find myself improving an entire project or mini brain off of that subject.
Anyways, I’ll stop the ranting now. I just wanted to tell you that you put something out there that I’d been thinking about for a long time and you hit the nail oh the head in tone and voice. I’ll definitely be sharing this. Thanks 🙏