People send me screenshots of stock analysis their chatbot wrote and ask whether it looks right. I can usually recognize the shape of the answer before I finish reading it. The model produces a confident block of text in about ten seconds, names a couple of opportunities and a couple of risks, recommends watching the next earnings report, and closes on a gentle word about diversification. It reads the way a careful junior analyst would write if you gave him an afternoon. The judgment it appears to contain was never actually formed.
What comes back from a blank chat is a polite room with no memory of how the work is actually done. Every question gets a courteous answer, and not one of the answers was built in the order that would let it mean something.
Stock research is not a conversation where each reply stands on its own, it is a sequence in which the answer to one question decides what the next one should be. The order is most of the work. Everything starts with the business and what it really sells. Competitive position comes next, then the balance sheet, and valuation arrives last, because a price tells you nothing until you know what is being priced. The bear case earns its place only once the bull case has been stated in plain terms. A general model inverts all of this. It writes the conclusion first and backfills reasons that sound as though they came earlier, and the paragraph passes a glance because it has the surface texture of analysis. What was supposed to produce that conclusion was never there.
The same room has no sense of which kinds of companies fail in which ways. Biotech lives and dies on trial risk, cash runway, and binary readouts. Cyclicals turn on normalized margins and where you stand in the cycle. Microcaps punish anyone who treats liquidity as an afterthought. A general model reaches for one template and lays it over all of them, and that template is shaped by whatever dominated its training, which is American large-cap. Point it at a Japanese bank or a Brazilian utility and the analysis is wrong before the first sentence, not because the model is stupid but because it has no doctrine for the company in front of it.
Real research starts wide and narrows. You throw out what does not matter until one driver is left standing, the one that actually moves the outcome. A blank chat does the reverse, spreading itself evenly across every topic because nothing in it has any reason to commit.
A workflow changes the result because the model stops starting from zero. Something outside the chat decides what gets gathered, in what sequence, and what has to be attacked before any conclusion is allowed to stand. The same model that writes fluent, average answers in a blank box can produce something sharper once it is forced to compress and argue first. This matters more in finance than almost anywhere else, because the dangerous failure here is not the obvious error. It is the fluent one. A confident paragraph with a wrong figure buried inside it does more damage than a visibly broken one, because the polish is doing the job that scrutiny was supposed to do. OpenAI's own researchers argued last year that models hallucinate in part because training rewards confident guessing over an honest admission of uncertainty. Regulators have arrived at a related worry from another direction, cautioning retail investors that an AI label is no mark of quality and is increasingly used to dress up thin or fraudulent offerings. Finance runs on dated numbers and context-bound claims, which is precisely where confident guessing is most expensive, and where the reader is most likely to mistake smoothness for rigor.
Someone who knows the field will reply that a skilled prompter can pull genuinely good research out of a blank chat, and in narrow cases that is true. A user who already knows what good looks like asks sharper questions and catches weak answers on the way back. The discipline in that case is living in the user, though, not in the model, and that user is absorbing the full cognitive cost the workflow exists to carry. A blank chat rewards its most expert visitor and offers almost nothing to the one who arrived unsure of what to ask, which describes nearly everyone who reaches for it. Selling a tool that only works for people who do not really need it is a strange way to claim you are helping.
This is the gap I built Tesseract Stock Agent to close, which makes me the wrong person to tell you whether it succeeds. What I can describe is the intent, which was to move the discipline out of the user's head and into the process itself, so the order holds whether or not the user remembers to impose it.
The deeper error is treating the model as a clever employee who only needs to be asked nicely. Better results come when you hand it an edge before you expect one back, whether that edge is a source library, a research framework, or a chain that walks it through the steps in sequence. A model can multiply structure it has been handed. It cannot reliably invent the right structure out of a lazy question, and it will never warn you that the question was lazy.
So the thing missing from the empty answer was never intelligence in the first place. A blank chat will always answer, and it will answer well enough to survive a quick read, which is exactly what makes it risky. What it withholds is the order of the questions, and the order is the part that decides whether anything it told you was worth keeping.
Sources:
OpenAI, Why Language Models Hallucinate
FINRA, Artificial Intelligence (AI) and Investment Fraud
Author: Benet Bani
Top comments (0)