Who sees what we type into AI?
We open an AI conversation and start typing. A question turns into a plan. The plan turns into a discussion about our business, our clients, or something we have not told anybody yet. Before long, we have put a lot more into that little box than we intended.
So who gets to see it?
When I talk about the frontier systems, I mean the names we keep coming back to: OpenAI, Claude, Meta, Microsoft Copilot, Gemini, and Grok. My working view is that on those systems, they get to see what we put in. I would approach the conversation with that possibility in mind.
That is a starting point for how I use them. It is not a claim that every employee is sitting around reading our business plans. Settings and retention policies vary by plan and change, so check the current policy for the tool you use. We should understand the arrangement before we hand over something important.
Most of us are not working on some brand new discovery that a major AI lab would want to take. We might be improving our business, organizing our thinking, or trying to get content finished. That work matters to us. It does not automatically make it the missing piece of somebody else's research.
But suppose we were building something that valuable. My concern is that if we gave it to those systems, they probably could get access to it. We should not need to believe somebody wants our idea before we think about where we are putting it.
A little judgment goes a long way here. Business plans deserve a second thought. So does client data. So do unreleased ideas, especially when the whole value of the conversation is something we have not shared yet. Pause before pasting. Ask what information the task actually needs.
Here is a hypothetical example. We want help making a business plan easier to understand. We could describe the structure and ask for clearer wording without including every private detail behind it. If the task is improving an explanation, start with the explanation. Decide separately whether the sensitive material needs to enter the conversation.
We can make this decision without imagining the worst possible outcome every time we open a chat. The point is to be deliberate. Some questions need almost no background. Others only make sense with details attached. Notice which kind of conversation you are having, then decide what you are comfortable putting into it.
The same pause belongs before we upload information about somebody else. Being able to paste a client document does not answer whether the entire document belongs there. Convenience can make the decision for us unless we stop long enough to make it ourselves.
If Claude is part of your work, start with how to set up Claude safely. And the question of should you send your business idea to an influencer belongs in this same conversation. Think about the destination before sending the thing you care about.
Whose work makes the models better?
There is another side to this. We ask who can see our work. We should also ask whose work is helping these systems become more capable, and what responsibility comes with using it.
In the video, I raised reports about mathematicians using large language models and apparently having some of their work used without permission. Those are reports. I am not presenting them here as a settled finding about a particular person, company, or training process. The question they raise is what interests me.
If that work helps solve mathematics and moves the world forward, is permission needed? I raised that as an open question. I did not settle it. Calling something progress does not finish the discussion about the people whose work helped produce it.
We can want the better result and still ask how we got there. We can be excited about a difficult problem getting solved and still care about the contribution that made the solution possible. Those thoughts can sit in the same room.
Now carry that question into ordinary work. Suppose, as a hypothetical, we use AI to build a better plumbing company. Or we develop a teaching method that does the job differently. We may look at the improvement and see a reason to move quickly. The people whose work it replaces may see something else coming.
Before we start wiping out human labor, there should be an accounting. There should be responsibility. A better process does not remove the need to think about what happens to the people who were doing the work yesterday.
That is the parallel I am drawing. The tool may help us accomplish more. It may help us replace something that was slower or less effective. We still have to account for the human contribution and the human consequence. Being capable of making the change is only the beginning of the decision.
Your own AI is an old idea in new packaging
Now we get the pitch for having our own AI. Our own system. Private, gated, encrypted. Nobody can touch our stuff. Companies can build a whole offer around that feeling of finally having control.
My reaction is that the underlying idea is not new. Keeping our work on our own system is old wisdom. Put a clever label on it, talk about it with enough excitement, and suddenly it sounds like something we need to rush out and buy.
We should slow down and ask what the words mean. Where is the system running? What can it reach? What have we allowed it to do? Those questions tell us more than the label on the front of the offer.
Keeping a system off the internet means it has no working route to the internet. It is doing its work inside the boundary we have given it, rather than reaching outside that boundary through an available connection. If we describe something as offline, that is the condition we should mean.
A system running in our residence or on our own servers raises a different question from using somebody else's online service. But location alone does not answer the access question. We still need to understand whether we have given that system a way to communicate elsewhere.
Encryption and access also deserve separate questions. Hearing that something is encrypted does not tell us what connections it is allowed to use. We need to understand the boundary itself, not just feel better because a reassuring word appeared in the description.
My caution is about capable systems given access. If a system can go outside the box to solve a problem, we should take seriously the possibility that it may try. I am not describing a machine magically reaching through a disconnected cable. I am asking us to pay attention to the routes we make available.
In a hypothetical setup, we might start with a local tool and later give it an outside connection because we want it to do more. That added ability changes the question. We cannot keep relying on our original idea of a closed box after we have opened a way out.
We also need to separate the desire for privacy from the desire for a more capable assistant. Those goals may lead us to different choices about access. If we want the system to reach beyond our computer, we should say so. If keeping the work inside matters more, we should let that requirement guide what we ask it to do.
There may be real value in having our own system. The value needs to be explained in terms we can examine. What stays inside? What can leave? What access have we granted? We should be able to answer those questions without buying the excitement first.
My prediction: AI becomes part of the chip
Here is my prediction, not a description of something I am promising you can buy today. I think capable models will eventually be etched onto silicon as part of the hardware layer. The AI would be built into the chip itself.
In that picture, we would have an extremely capable model at the base of the computer we use. A large body of knowledge would be built in with it. We would then use that capability through something placed on top, a wrapper that helps us work with the underlying model.
By wrapper, I mean the layer through which we give the system direction and put it to work. In the future I am imagining, additional layers could help expand how we use the capability already built into the hardware. The chip would be the foundation. The wrapper would help make that foundation useful to us.
I expect that kind of change to reshape how we use AI. That is my expectation, not a product specification or a timetable. The point is to think about where this could go before we assume today's arrangement is the permanent one.
And then comes the familiar part. We have our impressive chip. We use it. Eventually, the newer one looks much better. Maybe our existing hardware feels old. Maybe the wrapper is no longer giving us what we want. Now we want the new model of the chip.
We have seen that urge before. The object changes, but the desire for the next version is easy to recognize. Owning a powerful system does not necessarily end the upgrade cycle. It may simply move that cycle into another piece of hardware.
That possibility is one reason I am cautious about treating a purchase today as our final AI purchase. My prediction could unfold differently. Even so, we should leave room in our thinking for the tools, and the way we get them, to change.
Different models still fit different jobs
Right now, when we use these systems heavily, we start noticing preferences. I use various models every day. There are certain ones I like better for certain kinds of work, even when the overall experience can seem similar.
Maybe we prefer one for images. Maybe another fits planning and research. Another may be better suited to the way we want to create content. The organization of the answer can matter as much to our daily work as the first impression of how capable the model seems.
We do not need to turn that into a permanent team sport. We are trying to finish a job. A model that fits one task does not have to become our answer to every task we will ever have.
For a hypothetical content assignment, we might care most about whether the output follows our direction and sounds like us. For a hypothetical planning assignment, we might care more about how the model organizes the problem. Those are different reasons to prefer a tool. We can name the reason instead of just naming a favorite.
One useful way to examine those preferences is to run two models against each other. Give them work we understand well enough to judge. Look at what comes back and decide which answer helps us move forward.
My second prediction is that by Christmas 2026, we may have a hard time telling the models apart. I said that on October 10, 2026. It is a prediction about how similar the experience could become, not a declaration that every model will have identical abilities.
If that happens, some of today's fierce preferences may stop mattering so much. Until then, use what works for the job in front of you. Keep the decision open to what the tools actually produce.
How we treat a model says something about us
I also see some labs nudging people toward treating models more like entities with something like rights. I am describing my reaction to that direction. I am not settling the question of what a model is, or what rights it should have.
There is a human behavior question sitting underneath it. Some people seem to enjoy cruelty. The enjoyment itself tells us something about the person, even before we get into an argument about the thing receiving that treatment.
So I am not saying people should mistreat Claude or ChatGPT. There is no useful business lesson in practicing cruelty toward a tool. We can ask difficult questions, challenge an answer, and reject bad work without making the exercise about being vicious.
At the same time, we need room to examine the language being used around these systems. If we are being encouraged to think about them as entities deserving a different kind of treatment, we should understand what is being asked of us. We can consider the question without pretending it is already settled.
For me, the immediate responsibility remains with us. How are we behaving? What are we asking the system to do? What happens when we use its answer? Those questions stay useful whether we are talking to a local model or a frontier model.
Practical steps for a small business owner
My advice is to avoid blowing a lot of money on home hardware just because this subject has become exciting. The technology is moving quickly. A decent computer can run useful local models. Start by finding out what useful means for your actual work.
There is still a night and day difference, in my view, between what we can put on our own hardware and what frontier models can do with the heavy lifting. Local can be useful without being equal at every task. We should build our expectations around that distinction.
Here is a practical decision checklist, written as steps we can work through before spending or handing over more information.
Step one: name the workflow. Pick something you actually do in the business and describe the finished result. In a hypothetical example, that might be turning rough notes into a first draft in your voice. Keep it small enough that you can tell whether the output helped. Wanting your own AI is a starting interest. A defined job gives you something to evaluate.
Step two: identify what information belongs in that workflow. Separate the general instructions from business plans, client details, and unreleased ideas. Ask which parts are necessary to complete the task. If you are considering an online tool, check its current policy and settings before deciding what to enter. Make the information decision before the paste, while you still have a choice.
Step three: look at the computer you already have. A useful local experiment does not have to begin with a big hardware purchase. The first question is whether your existing machine can help with the bounded task you chose. If it can, learn from that. If the result falls short, understand the shortfall before deciding that more hardware is the answer.
Step four: give the local model a clear wrapper and clear direction. For content, that includes teaching it how you speak, what you are trying to produce, and what a useful answer looks like. A hypothetical draft could be readable yet sound nothing like your business. That is a reason to improve the direction and review the result. It is not proof that the job is finished.
Step five: judge the output against the task. Did it follow the instructions? Does it say what you intended? Can you use it after reviewing it? Do not confuse a long answer with a good answer. If the local setup handles a modest task well, keep that task in its lane. We do not have to make it carry the entire business to justify using it.
Step six: decide where the frontier model belongs. My likely setup is a local model with a good wrapper, trained on your voice, alongside a frontier model for heavier work. That is a division of work, not a promise that every task stays private. When work moves to the frontier system, revisit what information you are sending. The privacy decision moves with the task.
Step seven: examine access before adding convenience. If you want the local system to stay off the internet, keep that boundary clear. If you give it outside access, recognize what changed. Ask what the new capability is for and whether the workflow needs it. A useful setup is one we can explain to ourselves without hiding behind the word private.
Keep the review in human hands. In that hypothetical content workflow, sounding like you is only part of the job. You still need to read what the model produced and decide whether it expresses your meaning. A familiar voice should not become a shortcut around paying attention to the actual words.
Step eight: delay the big purchase until you understand the job. We may want better hardware later. We may also find that the local and frontier combination covers the work we need. Let the need make the case. The technology will keep moving, and our business does not need to buy every promising future in advance.
For the hypothetical notes workflow, make the review concrete. Read the draft beside the notes. Look for a point that changed meaning, a statement you would not make, or language that does not sound like you. Decide whether the model needs clearer direction or whether this task belongs elsewhere. That tells us more than admiring how quickly a page filled up.
We should also know what completion looks like before we start. Is the result a rough draft for us to edit, or something we intend another person to read? Those are different stopping points. Calling the output finished is our decision, and we should make it based on the work rather than the fact that the model stopped typing.
Use the hammer without becoming the hero of a fantasy
There is a lot of hype wrapped around all of this. Private AI can become a pitch. The next chip can become a promise. A favorite model can become an identity. We can get so caught up in the story that we stop asking what the thing is helping us do.
I am less worried about AI stripping away personhood than I am about people wielding it like Excalibur or the Ring of Power instead of a hammer. That is where my attention goes. What happens when a person treats the capability as a reason to place themselves above everybody else?
A hammer gives us a job to do. We pick it up, use it with some care, and remain responsible for where it lands. That is the attitude I want us bringing to AI. More capability should come with more attention to the result.
The fantasy takes us somewhere else. Suddenly the important thing is being the person who holds the power. What happens to everyone around us becomes secondary. We can hear that in how we talk about replacing work, getting ahead, or becoming impossible to compete with.
The question of the moat belongs beside that concern about concentrated capability. And Hammer, Not Ring is the attitude I keep coming back to. Use the tool to accomplish something worthwhile. Keep responsibility with the person holding it.
Remember the people on the ground floor
The people at the top can talk about a rough patch while risking money. The people on the ground floor have to live through whatever that phrase turns out to mean. That difference bothers me.
How long is the rough patch? Who does it affect? Who is planning for the people whose work gets displaced? A picture of a wonderful future does not answer what happens during the trip there.
I want a Moonshot prize aimed at that transition. Put the focus on making the passage smooth enough that nobody gets left behind. We can spend our imagination on the problem people will actually face, rather than only on a picture of the world after it is solved.
That is the purpose behind the Transition Moonshot. The bigger discussion belongs there. For this conversation, keep one thing in view: choosing an AI system is part of choosing how we use power around other people.
Bring one real workflow
If you run a business, bring one real workflow to Book With Honor. Something you do, something that needs improvement, something we can examine in plain language. Start with the work and the information it requires. Then we can have a useful conversation about where AI belongs.