The Model We Didn't Choose

They grew up in the same house.

Same parents. Same neighborhood. One year apart in age. Similar schools, similar academic paths, similar circumstances by almost any measure visible from the outside.

By the time they were adults, you would not have guessed they came from the same family.

One was brilliant and isolated. Deeply suspicious of others. Ultra-conservative in almost every decision, financial and otherwise. Money sat untouched rather than deployed. Plans were started and abandoned. Trust was withheld even where it would have helped.

The other was equally brilliant. Social, a connector, a planner. Trusted others without being naive. Made balanced decisions across every domain. Built things deliberately and followed through.

Same household. Completely different people.

I knew both of them well. And I still cannot fully explain the divergence.

But that, I eventually realized, was the more interesting part.

The siblings did not have identical inputs. They had the same household. That is not the same thing.

Birth order. Temperament. Friends. Teachers. Chance events. The way each interpreted the same conversation. What each made of the same parent on a different day. Experiences that one remembered and the other barely noticed.

Much of what shaped them was invisible to me.

The puzzle wasn't that the same inputs produced different outputs.

The puzzle was that I could never see the full input set.

I'm not sure anyone could.

I carried that puzzle into my work without quite realizing it.

Every client who walked into a conversation was carrying a model they had not consciously designed.

Not a financial model. A decision model.

It had been built from everything that happened before the conversation started: the family they grew up in, the first time they lost money, the first time someone they trusted with money let them down, the culture they absorbed around wealth — whether it was something to be protected, spent, feared, displayed, or never discussed.

That model produced outputs I could observe.

The client who couldn't hold cash without immediately investing it. The one who couldn't sell a position even when the numbers argued for it. The one who asked the same question seventeen different ways because they needed to hear the answer seventeen times before anything shifted.

I could see the outputs clearly. The inputs, partially. The process connecting them, almost never.

And over time I noticed something that complicated the usual language of bias.

The model was often not wrong.

The client who couldn't sell a concentrated position had sometimes built significant wealth through exactly that kind of conviction. The instinct that later made diversification difficult was the same instinct that had produced extraordinary returns.

The behavior that looked irrational in one context had been adaptive in another.

The problem wasn't necessarily that the model was defective.

Sometimes the world had changed. The model hadn't.

And when that happens, the model doesn't announce it. It keeps running.

Often, the reasoning comes after.

We experience a decision as the conclusion of analysis, but much of the direction may already have been shaped by the model underneath. The explanation arrives afterward, giving the decision a coherent story.

This isn't a defect unique to investors. It is one of the ways human cognition makes the world manageable. We learn from experience. We recognize patterns. We build shortcuts. We predict what is likely to happen next.

Speed requires shortcuts. Shortcuts require tradeoffs.

The question that stayed with me about the siblings was not why they were different.

It was how little of the explanation was visible even to someone who knew them well.

I began to recognize the same problem in client conversations — and then, years later, AI gave it a name I couldn't ignore.

Interpretability.

AI researchers can inspect what goes into a model and measure what comes out. They can sometimes identify which features influenced an output, trace certain internal patterns, find particular behaviors. But a complete account of why a model produced this output, in this situation, remains extraordinarily difficult.

We spent years calling AI systems black boxes.

Then we began to realize something uncomfortable.

We are too.

Not in the sense that human brains are literally neural networks, or that human cognition works exactly like machine learning. It doesn't.

The parallel is simpler.

Both are systems whose past shapes how they interpret what comes next — and both can encounter situations where those patterns stop working.

The human black box built the machine black box.

And in doing so, reproduced some of the properties that trouble us about our own cognition: learning from experience, pattern recognition, internal representations we cannot easily inspect, imperfect generalization when circumstances change.

We built systems in our own image without quite knowing it.

A decision model can be adaptive when it is formed and quietly maladaptive later.

The investor whose refusal to sell once protected them from panic may eventually mistake persistence for discipline.

The model keeps producing internally consistent answers.

That is precisely the problem.

Internal consistency is not evidence that the model is still describing the world accurately.

The model rarely tells you.

The most dangerous model isn't the one that was wrong. It's the one that was right for a long time.

Perhaps this is where self-knowledge becomes less about understanding yourself completely and more about becoming suspicious of your own certainty.

You don't need to reconstruct every experience that shaped a decision. You probably can't.

You don't need a complete causal account of why you respond to one situation with confidence and another with fear. You may never have one.

But you can notice when the predictions keep failing. When the same decision produces the same disappointing outcome. When evidence that should change your mind somehow never does. When the explanation you give yourself remains remarkably stable while the world around it changes.

That may be the closest thing we have to auditing the model.

Not opening it up completely.

Watching what it predicts.

And noticing when reality keeps disagreeing.

I still think about those two siblings.

Not about what went wrong or right.

About the puzzle.

The same household, the same parents, the same neighborhood — and yet two completely different ways of moving through the world.

I still cannot fully explain it.

Not because I lacked access.

Because a complete explanation would require seeing everything that mattered: every formative experience, every interpretation, every weight placed on every event, every connection between them.

That information was never available to me.

It may not be fully available to them either.

And perhaps that is what makes the comparison with AI uncomfortable.

We are discovering that, in trying to understand machines, we have built a mirror for a problem we have never completely solved about ourselves.

We can observe the output.

We can sometimes reconstruct the inputs.

But somewhere between the two is a process we may never fully see.

We built systems in our own image, it turns out.

More than we realized.

Diptes Basu writes about investing, behavior, and decision-making, drawing on twenty-five years in global financial markets.

This essay reflects the author's personal views and is intended for educational purposes only. It does not constitute investment or financial advice.