What using AI taught me about reading a book

For the longest time I thought it was being smart while using AI. I’d ask a question, get a few answers, and then went on an endless cycle of checking sources, understanding what it’s saying, checking the logic, and comparing it with what I already knew.

It seemed like the right thing to do . Especially if you work in finance. But eventually I noticed something uncomfortable. I was only trusting AI when it told me something I already knew. If AI’s answer matched what I had already experienced, I would say “yes, it makes sense”. If it didn’t I would just ignore it.

I had in front of me one of the most powerful tools for thinking and learning but I was limiting myself to my existing knowledge. It’s like reading a book and saying “I’m happy to learn from this book as much as it tells me things I already believe”. That’s not really learning. Let’s just confirm what I already know.

Why reading a book feels different?

If you pick up a novel, a fiction novel, you know, the things there didn’t happen. All in there is just the imagination of the author. You know that because the book is classified as fiction.

If you get a non-fiction book, you see it differently. There is usually research, an editor, a publisher, sources, a review, and some process around checking the facts. Of course it’s never guaranteed that everything in the book is correct. As you’re reading you are still questioning. But your mind gets a little bit more open to absorb new ideas. That’s because you trust the process. You now can relax and open your mind to an idea that you did not already know. You can explore and challenge it. Even when you compare to your own experience, used to absorb new ideas.

With AI it is different. AI can give you a fact that is true but you never know if it’s an inference, an interpretation, or something completely fabricated. Both facts and hallucinations arrive beautifully written in paragraphs with the same type of confidence. Unless this is something that you already know, it’s hard to trust.

The three forces that rules AI

There are three forces we can apply on any AI interaction that helps shaping the output.

1. The probabilistic force

We all know AI is a machine learning system. That means it answers our questions based on pattern recognition. It recognizes the patterns and infers the answer. And it does that based on what it learned from gigantic source data. It’s like a super intelligent person that has read all content in the world and can remember everything.

But any knowledge has gaps. And AI knowledge base has gaps as well. The issue is when AI fills those gaps with something plausible. Let’s call hallucination.

The beauty of the probabilistic forest is that it takes me to places I have never been. Shows me things I have never learned. But if it’s not just giving me true facts, it will be hard to trust.

2. The Deterministic force

To put boundaries around the probabilistic force, we can create structure. The things we know that can be a gap in the knowledge base need to be defined previously. For example there might be a hundred ways to calculate an indicator. And AI can create 50 more that could be wrong. So if I want this indicator to be calculated according to my rules I have to tell AI how to do it. And once I define the way I calculate that indicator now I can have AI bring all of the possible analysis to it. It will have the bike number. It will have the fact.

That goes the same way for templates, reports, scripts, specific formats, defined structure and anything that will limit what AI will answer. It’s a hard shell around the creativity. AI will still bring the probabilistic force but now within limits.

3. The human Judgment

The human interaction with AI will bring context out of the systems. Things that AI has no need to know. This enriches the analysis with facts that are out of its knowledge base. It fills the gaps, an also gives humans more control over the outcome. The human interaction becomes an exchange. AI will take that context, fill the gaps, and become more powerful. At the same time the human will take advantage of the probabilistic analysis and expand its area of expertise.

The human improves because AI exposes them to more possibilities we see things we have not considered. If we disagree we are forced to explain why and this sharpens our thinking.

The interaction is probably one of the greatest values of using AI.

Trust is not the starting point

I used to think the question was, “Can I trust AI?” . I don’t think that’s the question anymore. My answer will be no. I cannot trust automatically. And you probably shouldn’t either.

A better question would be: “What structure do I need to create so that I can trust this particular output enough to learn from it?” . That distinction changed the way I work with AI. Because they are two extremes. At one extreme you trust blindly anything they read from it. That’s clearly dangerous. At the other extreme you distrust anything you can’t independently prove from your existing knowledge. It sounds less risky but it creates a different problem.

If you never can trust anything that comes from AI, then you don’t already know, you limit your learning. The opportunity sits in the middle. We allow probabilistic intelligence to explore and bring new possibilities. But we create the boundaries and controls around where accuracy matters.

You have become the editor

This is why I keep coming back to the book analogy. When you read a nonfiction book someone has created a structure around the ideas before they reach you. It is a process around that will check facts and set a structure. Cited work from trusted sources serve as credibility enhancers. You let go of doubts and open your mind to expand your knowledge. Even if you disagree you are still learning, because you twist the sources .

With AI we increasingly have to build that editorial structure ourselves , deciding which sources it can use, what there is allowed to infer, and how it’s allowed to calculate. We decide upon its validation and what can be challenged. We decide where human judgment enters the process.

Once those boundaries exist something opens up. You can stop AI from just confirming what you already know and you let it challenge you. You let us show you something unfamiliar, you can explore without immediately rejecting it. Now you have a process for distinguishing insights from inventions.

That’s when AI becomes far more than an efficiency tool. It becomes a way of expressing human capability.

Stop Building AI and start Building Intelligence

On September 24th I’ll be exploring this idea further in a session with the AFM called “Stop Building Models and Stop Building Intelligence.” Finance has spent decades becoming exceptionally good at building models. Models to calculate, organize assumptions, and help us see scenarios. In the last few years AI has become more and more capable of building those models for us. But these models exist to drive decisions. Decisions are taken by humans. Humans are prepared to let AI explore with the right boundaries around it.

AI gives us the opportunity to build something around the model. Something that can explore, question, interpret, challenge, explain, and recognize patterns, but interact with the human using it. That’s intelligence.

Intelligence without boundaries is not enough for finance. We need controlled intelligence. Systems where probabilistic AI can explore, the deterministic process can create reliability, and human judgment can continually improve the outcomes.

That’s the shift finance leaders need to stop preparing for now. This is perhaps the biggest change in my own thinking about AI. I started out believing I needed to decide whether I could trust it. Now I know this is the wrong goal. The goal is to create boundaries that allow me to trust enough to go beyond what I already know. Because if AI can only tell us what we already know, that is not intelligence. That’s confirmation.

Your opportunity in front of us is much bigger than that.