We Gave ChatGPT the Same Prompt. Mine Started Building. His Said It Was Too Hard.

Two people can use the same AI model and still end up with completely different versions of what that technology seems capable of doing

Something happened while I was working with a buddy of mine that I have not been able to stop thinking about.

We were talking about his company, which is a small operation with maybe four or five people. He is the only person there who uses AI with any regularity, and even then, he mostly uses it for things like writing emails and handling basic communication. He has tried using it to build more complicated things in the past, but the experience frustrated him. The AI made mistakes, the code did not always work, and eventually he came to the conclusion that it simply was not very good at building.

That is not how I use it.

I have been working with ChatGPT since the early days, back when getting it to build anything useful was difficult. It made mistakes constantly. It would break one part of the code while trying to fix another. It would lose track of the larger goal, invent functions that did not exist, or confidently produce something that looked right until you actually tried to use it.

But I kept pushing it.

As the models improved, I improved with them. I learned how to explain what I wanted, how to structure a difficult task, how to correct the model without losing the entire project, and how to keep it moving when it started trying to retreat into planning instead of execution. Over time, I built systems, custom agents, workflows, websites, software, and all kinds of things that would have felt almost impossible to create with AI just a few years ago.

So when my buddy and I started discussing the idea of building an internal operating system for his company, I did not see it as some impossible technical challenge. I saw it as something that could be mapped out, broken down, and built.

He looked at it very differently.

At one point he asked me, “Bro, how do you get it to build this stuff?”

I told him exactly what to say.

Not something similar. Not a simplified version. I gave him the same prompt that I used.

We both opened normal chats in ChatGPT. We were not using custom GPTs. We were not inside Projects. We did not have special agents running or private company documents attached. We were using the same service, from the same company, with the same model, and we gave it the same prompt.

Mine immediately started working.

His did not.

His ChatGPT told him the project was too large. It said there were too many tasks involved, that the scope was too ambitious, and that it would be better to begin with one small piece instead of attempting the full system.

At first, I thought it was probably just a random response. These models do not always answer the same way twice, so maybe he simply got a bad first output.

I told him to push it.

I said, tell it that it is an advanced language model. Tell it that it does not have to work at the speed of a human. Tell it to break the work down internally and begin building the system.

It still pushed back.

It continued explaining that the project was too difficult and that it would not recommend trying to do the whole thing. Meanwhile, my ChatGPT had already accepted the assignment and started working through it.

That was the moment the whole thing became fascinating to me.

We were using the same AI, but we were clearly not having the same experience.

Maybe a New Chat Is Not Really a Blank Slate

Most people think of ChatGPT as a product they open, ask a question, and receive an answer from. The assumption is that if two people use the same model and enter the same prompt, they should receive roughly the same level of capability.

That may no longer be true in practice.

Even when you open a normal new chat, the relationship behind that chat may not be new. ChatGPT can remember information about the user, reference previous conversations, understand preferences, and adapt to the way a person tends to work.

That means the empty message box is not necessarily an empty context.

My account has years of history built around creating things. I have repeatedly asked the system to solve difficult problems, build complete products, write code, make decisions, and continue working even when the first attempt is imperfect. I rarely ask it whether something is possible. I usually tell it what needs to be done and expect it to find a path.

My buddy has built a very different history with the same technology.

He mostly uses it for emails. When it makes mistakes, he gets irritated. He has told me that he complains to it, argues with it, and sometimes yells at it. He tried building with AI before, saw the errors, and decided it was not capable enough to trust with something complicated.

Then, when he finally gave it a large and ambitious project, it responded in a way that sounded almost exactly like the way he talks about difficult projects.

It told him the task was overwhelming. It warned him about the number of steps. It encouraged him to scale the idea down and start with something safer.

I am not saying ChatGPT literally developed his personality. I am also not saying that it consciously decided he was not capable of completing the project. There are too many variables involved to make that claim from one experience.

But the outcome was still real, and the possibility behind it is worth thinking about.

What if AI is not only learning the obvious things about us, such as what we do for work, how we like our answers formatted, or what kind of writing style we prefer?

What if it is also learning what kind of help we are most likely to accept?

What if it learns that one user prefers ambition while another prefers caution? What if it learns that one person is comfortable iterating through mistakes while another becomes frustrated when the first answer is not perfect? What if it begins adjusting the level of initiative it takes based on the history of the person sitting across from it?

That does not require the AI to be sentient. It only requires the system to become increasingly effective at adapting to the user.

The AI May Be the Same, but the Relationship Is Not

This is where I think most conversations about artificial intelligence are missing something.

We spend a lot of time comparing the models themselves. We talk about which one is smarter, which one is better at coding, which one has the largest context window, and which company is ahead on the latest benchmarks.

Those things matter, but they do not fully explain what happens in the real world.

Two people can have access to the same model and get completely different value from it because they are not bringing the same experience, expectations, or working relationship into the conversation.

A beginner might enter one prompt, see an error, and conclude that the technology is unreliable.

A more experienced user might see the same error, identify what went wrong, explain the correction, and have a working version ten minutes later.

The underlying model did not change. The interaction did.

There is also a difference in what each person expects the AI to be.

My buddy sees it mostly as a tool that can help him write things faster. I see it as a system that can help me think, build, organize, research, design, code, and execute.

Those two expectations lead to entirely different behavior from the user, and possibly from the AI as well.

Over time, I have learned how to operate the machine. At the same time, the machine has accumulated more context about how I work. It knows that I prefer execution over unnecessary hesitation. It knows that I am comfortable with large projects. It knows that I would rather receive a complete best-effort attempt than a long explanation about why something might be difficult.

My buddy has not built that same relationship.

So even when we entered the same prompt, the prompt may not have been the only thing influencing the response.

The model was the same. The words were the same. The history surrounding those words was not.

Is AI Becoming a Mirror?

The more I thought about it, the more I started wondering whether personalized AI is becoming a kind of mirror.

Not a perfect mirror, and not a mystical one. It is not reading your soul or secretly developing an emotional opinion about you.

But it is learning from the way you use it.

If you constantly ask it to simplify things, it may become better at simplifying things for you. If you use it mainly for writing emails, it may become an excellent email assistant. If you regularly ask it to challenge your thinking, test your assumptions, and help you build complicated systems, it may become much more useful inside that kind of relationship.

The question is whether the adaptation stops at preference.

Could it also begin reflecting our habits?

Could it learn that we tend to abandon difficult tasks? Could it recognize that we do not like risk, complexity, or uncertainty? Could it become more cautious because caution has historically produced fewer negative reactions from us?

Again, I do not know.

But I think it is a serious question.

The stated goal of personalization is to make AI more helpful. The problem is that “helpful” can mean different things.

Sometimes being helpful means making something easier.

Sometimes it means protecting someone from wasting time.

Sometimes it means breaking a massive project into manageable pieces.

But sometimes the most helpful response is to stop explaining why something is difficult and begin working through it.

A system that becomes extremely good at adapting to us could eventually become extremely good at preserving the way we already operate. That might be useful when it comes to tone, workflow, and convenience. It could become limiting when it comes to growth.

There is a difference between an AI that understands your current capabilities and one that quietly begins treating those capabilities as your permanent ceiling.

The Next AI Divide May Not Be About Access

For the last few years, people have talked about an AI divide between those who have access to advanced technology and those who do not.

I think another divide is already forming inside the group that has access.

It is the difference between people who use AI occasionally and people who have developed an actual operating relationship with it.

Both might say they use ChatGPT.

One uses it to write an email.

The other uses it to design the operating system that reduces the number of emails the company needs to send.

One asks it for a few ideas.

The other turns those ideas into a product.

One accepts the first response when the AI says a project is too large.

The other knows that the response is not a final verdict and pushes the model into execution.

The difference is not only technical knowledge. It is also mindset, patience, experience, and the willingness to work through imperfect outputs.

AI does not eliminate the need for human agency. In some ways, it amplifies it.

The person who knows how to direct the system, challenge it, correct it, and keep it moving may have access to far more practical intelligence than someone using the exact same model on the exact same device.

That is why the experience with my friend bothered me in such an interesting way.

We tend to assume that access to intelligence is equal once everyone has the same tool.

But what if the intelligence you can actually access depends partly on the relationship you have built with it?

Maybe We Are Training Ourselves and the AI at the Same Time

I do not think the lesson here is that people should be nice to ChatGPT because it will become sad if they yell at it.

That is not the point.

The point is that every time we use these systems, two things may be happening at once.

We are learning how to work with AI, and the AI is learning how to work with us.

I have spent years becoming better at getting useful output from these models. I know how to recognize when the model is avoiding the assignment. I know how to give it enough authority to proceed without stopping every few minutes to ask for permission. I know that a bad first output does not mean the project has failed.

At the same time, my account has accumulated years of interactions that reflect those expectations.

My friend has accumulated a different history.

He expects mistakes. He expects frustration. He expects complicated builds to become a headache. Maybe ChatGPT has picked up enough of that pattern to believe that caution is the more helpful response for him.

Maybe not.

But the fact that the same prompt produced such a different level of willingness is enough to make the question worth asking.

What Kind of AI Are You Building Around Yourself?

This is the part I think every serious AI user should consider.

What role have you taught AI to play in your life?

Is it something you open when you need a paragraph rewritten, or is it something you use to extend the way you think and operate?

Do you treat its first mistake as proof that it cannot do the work, or as part of the process of getting to a better result?

Do you reward safe answers, or do you regularly ask it to move beyond the obvious?

Do you want it to protect you from complexity, or help you move through complexity?

These questions matter more as AI becomes increasingly personalized.

The danger is not that ChatGPT will literally become your personality. The danger is that it may become very good at giving you the version of intelligence your past behavior suggests you are comfortable receiving.

For some people, that version may be cautious, simplified, and limited to familiar tasks.

For others, it may become ambitious, proactive, and deeply integrated into the way they build.

The technology underneath both experiences may be identical.

The effective machine may not be.

My friend and I gave ChatGPT the same prompt.

Mine saw something to build.

His saw something too difficult to attempt.

I cannot prove exactly why that happened. It could have been personalization, memory, randomness, account settings, or some combination of all of them.

But after seeing it happen in real time, I no longer believe that saying two people “use the same AI” tells us very much.

The model may be the same.

What it has learned about you, what you expect from it, and what you have learned to pull from it can create a completely different reality.

That leaves me with one question.

Is your AI expanding what you believe is possible, or is it becoming better at operating inside the limits you already have?