Better Use of AI by Knowing Its Limits

I recently considered handing my journal — hundreds of pages of daily reflections on work, relationships, inspirations, frustrations — to an AI model, expecting it would bring forth enlightening insights on the essence of who I am, the arc of my own thinking. Maybe it could be a highly impactful coach, by "knowing" me so well.

The AI talked me out of it. It warned that its pattern-matching would flatten months of my inconsistent stream-of-consciousness writings about fleeting circumstances into a tidy formula. It could produce something directionally correct, but it would be over-simplified, misleading, even if presented in a convincing summary. The machine’s warning no doubt reflected classic observations about the inadequacies of summarizing a life such as Heraclitus famous line "no man ever steps into the same river twice, for it's not the same river and he's not the same man.” Leveraging AI as I thought I might in this case would have been a misunderstanding of its utility.

That interaction sparked this piece. We are learning to live in the new reality of AI, which means embracing its utility and knowing its limitations. Specifically, I want to address AI in the realm of financial decisions. The tool is extraordinary, but so are the risks of misunderstanding its application. In this context, I consider four distinct roles AI might play, and the corresponding limitation of each.

1. The answer machine. Something like the sum of recorded human knowledge, a few keystrokes away. This is new in history and profoundly empowering. But a library is not wisdom, and an infinite library is not infinite wisdom. All the answers in the world are worth precisely as much as the question you bring to them and preparation of your mind. An investor fearful of government instability asks about buying gold. He gets back beautifully crafted, intricate answers — answers that play into the very bias that produced the question. Bias and the hunger for validation are human flaws, and AI magnifies them. The more complex the field and the higher the stakes, the more cautious we must be in use of the infinite answer machine. We have to know when we lack context to ask the right questions or to recognize bias — that humility is itself rare.

2. The power tool. Like modern construction equipment, AI lets one person do the work of forty. But put me at the controls of an excavator or a crane without training, and the results will not be good. AI now lets anyone produce institutional-looking analysis in minutes. That output is the costume of expertise, and it can mask substantial misunderstanding or even purposeful manipulation.

I recently spent hours on a client situation involving a partnership split — poorly drafted documents, a possible multimillion-dollar exit, adjacent to my work but outside my expertise. I asked good questions of the AI tools, prompted carefully, and wrote up recommendations I was frankly proud of. Then I took them to an attorney who specializes in these disputes. In minutes, he identified flawed framing and showed me exactly where my proposal would fail to produce a resolution — not factual errors, but the kind of thing you only see after sitting in those rooms and watching proposals like mine fail. The tool had made me capable enough to be dangerous, and had I passed that well-crafted proposal to the client, it would likely have been accepted as expert advice. AI's productivity shortcuts will produce more output at lower quality. An eighth grader can “write” a convincing book of marriage advice. We can be the suckers consuming that material, but we might also be the “eighth grader” confidently generating it. Ease of production does not mean what is produced is good.

3. The consultant. AI plays the advisor fluently — it breaks down goals, sequences steps, even encourages you warmly. Here is what it does not carry. When I give advice, I bear a legal burden of proving it serves my client's interest. Beyond the law, I share a community with the families I serve, my name is on the door, and one careless recommendation can end a reputation built over decades. The burden is what makes the advice trustworthy. The machine carries no duty, no downside, no name. In my role, I take risk when I recommend investments. To evaluate that risk — say, when I review a private real estate development fund — I want to know how much personal capital the sponsors have at risk. Do they have skin in the game? AI tools do not have skin in the game. We have a word in markets for advice with the consequences removed: moral hazard.

4. The probability machine. In assessing odds, AI is genuinely superb: the analytics, the base rates, the historical frequencies, instantly and without arithmetic. But odds are not the future. When results depend on how millions of unpredictable human agents react to hundreds of unpredictable variables, none of that processing capability closes the gap. In complex adaptive systems — markets, economies, lives — the decisive variables are the unexpected twists, and no dataset holds them.

Which brings me to the error investors will repeat with the probability machine at their fingertips. They mine all of market history and build a brilliant thesis on the data. They mistake the quality of the presentation for the reliability of the prediction. Consider planning an outdoor wedding. We instinctively know how much money, stress, and emotion is at stake if weather ruins the event, so we might ask AI for an analysis based on the date and local climate. But if the wedding is a year out? Any understanding of complexity in weather tells you that prediction is not possible, no matter how much historical data and computing power we feed it. If there is a 10% chance of rain, the wedding needs risk management — the prediction is neither reliable nor even useful under the circumstances. The same limitation governs investing and financial decisions.

Every day, people use AI in these four roles: the answer machine, the power tool, the consultant, and the probability machine. I certainly do, and we benefit tremendously from its capabilities. But we must recognize the limitations. Infinite knowledge is not wisdom, the ease of production creates a mask of competency and false confidence, the machine has no stake in your outcome, and there is no such thing as a prediction machine.