There is a difference between using a model to read your documents and letting it decide your orders. The first is time saved on a tedious task. The second changes the nature of the risk, and the problems it raises are not teething troubles that a better version will fix.
This article does not say AI is useless in finance. It says autonomous order execution is a category of its own, and why.
The reason cannot be audited
A generative system produces a plausible justification after the fact, not the actual cause of its decision. A convincing explanation and a true one are different objects.
It learns the past
Any rule drawn from history describes a market regime. Regimes change, and it is exactly at the change that the rule fails - that is, at the worst moment.
It acts faster than you can correct
A human mistake is noticed between two decisions. An automated one executes hundreds of times before anybody opens the screen.
Responsibility does not automate
The order is placed in your name. The loss is yours, the tax return is yours, and the explanation to give is yours.
It spends the only edge you had
A private investor has neither speed nor information. They have patience and no obligation to act. A bot consumes precisely that edge.
None of these five is a bug. They are properties of what the machine is being asked to do, and improving the model does not remove them.
The audit problem, which is the deepest
When a generative system explains why it did something, it produces plausible text. That text is not a record of its computation: it is a generation, exactly like the decision was. The two can diverge with nothing to signal it.
That is very different from a written rule. "Sell if the position exceeds 20% of the portfolio" is a sentence you can read, test, and prove wrong. You can accuse it of being false. A model answering "I judged the macro context justified trimming" cannot be proven wrong, because the sentence is not the cause.
That has an uncomfortable practical consequence: you cannot tell whether the system was lucky or right. Over a few months, the two look identical. And the period during which they look identical is also the period in which you decide to trust it with more.
The regime problem, which is the most classical
Any strategy learned on past data learns a regime: a set of conditions — rates, volatility, correlations, liquidity — that held during the training period. A regime is not a law. It changes, usually fast, and often precisely during the episodes when you would most need the strategy to hold.
That is not a criticism of machine learning; it is its definition. A system trained on ten years of market has seen ten years of market. What it will do facing a configuration it has never seen is not predictable from its past performance, and that is exactly when the question arises.
The speed problem, which is the most underestimated
A bad human decision is a one-off. You make one, you regret it, you correct the next. The rate of your mistakes is bounded by the rate of your attention.
An automated system has no such bound. A faulty rule applies at every occurrence, without fatigue and without doubt, between the moment it starts misbehaving and the moment somebody notices. The gap between those two moments is the real risk, and it is measured in hours or days, not seconds.
It is the same mechanism as the silent software bug: what costs is not the error, it is how long it goes unnoticed.
The responsibility problem, which is the most concrete
If a system places orders in your name, you hold the positions. You declare the gains, under the rules that apply to your situation. You explain, if explaining is required.
"The bot did it" is not a defence, not an explanation, and not a line on a tax form. Technical delegation comes with no delegation of responsibility, and that asymmetry is rarely highlighted by those selling the tool.
The edge problem, which is the most ironic
A private investor has no speed advantage: they will never have better latency than a trading desk. No information advantage: they read what everybody reads, after everybody. No cost advantage: they pay more per trade than an institution.
They have exactly two things. Patience — nobody obliges them to produce a quarterly performance. And the right to do nothing — no mandate requires them to be invested.
A system that acts continuously spends both. It converts the one ground on which a private investor is not disadvantaged into one where they are. That is the point that should give pause before all the others, because it does not depend on the quality of the model: even an excellent model, by acting often, trades a real edge for a competition already lost.
Where a model is useful, and why that is a different activity
Nothing above says a model has no place. The distinction is between reading and deciding.
Reading a forty-page bank statement and pulling out the lines is tedious, checkable, and harmless if redone. You see the result, you correct it, nothing is committed. That is the kind of task where a machine genuinely helps.
Placing an order is an irreversible action with real money, whose justification is not auditable and whose responsibility stays entirely with you. It is not the same category, and treating them as the same is the mistake this article is trying to name.
The question to ask before delegating anything is not "does it work?". Over a short sample, almost everything works. The question is: can I check why it worked, and will I know when it stops? If the answer is no, what is being delegated is not a task. It is a risk.