Why You Can't Just Act on the Trading Advice an AI Gives You
Read it, don't act on it. The trading advice a language model gives you is a reorganization of the enormous quantity of text it has read, not a judgment made about your account. This isn't a question of how well one product has been tuned; it falls straight out of three structural limits that models of this kind have: the knowledge has a cut-off date, it doesn't know your positions, and it will describe things that don't exist in convincing detail. Not one of the three can be worked around. Below, each of them taken apart, and then a dividing line between what to ask and what not to ask, plus two checks that take under two minutes.
First, draw a line: it explains well, it decides badly
In one sentence: anything of the "what does this mean" kind it handles better than most tutorials do; anything of the "what should I do right now" kind is not something you can lean on.
Explaining is work like turning an English API error into plain language, making "mark price", "maintenance margin ratio" and "slippage" clear and then giving a counter-example, or restating a strategy you have only half thought through. Questions like that have settled answers, written down long ago across countless documents and textbooks; the model has seen them many times over and restates them fairly reliably.
Deciding is work like "can I add to this position now", "is this line about to break out", "how much have I actually lost on this trade". It answers those three as fluently as it answers the others — and that is exactly the problem. Fluency has nothing to do with correctness. It does not change its tone of voice because it doesn't know.
I started out using it as an adviser too: asking about direction, asking about levels, and the answers all sounded like the real thing. What actually changed how I use it was the time it handed me a very specific set of steps, I tapped along with them, and the menu simply wasn't there — while it had described the screen as if it had just been looking at it. After that my usage became: let it be the one asking the questions, not the one making the call. It lists ten things worth confirming, I go and confirm them one at a time, and that combination is far more useful than asking it for the answer directly.
Limit one: the knowledge stops at some point in time, but it won't say "I don't know"
Every model's training data has a cut-off, and it doesn't know what happened afterwards. That in itself isn't a defect; any source of knowledge has a shelf life. The problem is that it doesn't fall silent because of it — it takes information from before the cut-off and answers questions about after the cut-off in exactly the same confident tone.
And the things that change fastest in trading are precisely facts of that kind: an entry point for a feature has moved, a fee tier has been adjusted, a product has been withdrawn in certain regions, a trading pair has been delisted. Changes like these don't come with a press release; the page just quietly becomes something else. In the old "snapshot of the world" the model is holding, they are all still where they were.
Cut-offs differ from model to model, and vendors update them, so don't go memorizing a particular date — anything you memorize goes out of date soon enough. What you want to build instead is an unconditional habit: whenever it is telling you "here is what the rule is now", treat that as information that may already have expired.
Web search relieves part of this, but relief isn't a cure. What it retrieves may be a third-party article that is equally out of date, and once that has been paraphrased you can't even see where a given sentence came from any more. To confirm something properly you still have to go back to the Help Center or the fee page of the platform you are actually using.
Limit two: it doesn't know what you are holding
"You could consider adding a little here" is two opposite pieces of advice depending on whether it is said to someone holding nothing or to someone already fully committed with leverage on. The same sentence: acted on by one person it opens a position, acted on by the other it pushes them towards liquidation.
The model has none of your account context. It doesn't know your cost basis, how much margin you have left, how long you can stand to be down on this money, or where next month's rent is coming from. What it gives you is a general formulation aimed at the "average reader", and in trading there is no such thing as an average reader.
Even if you type your position data in, what it receives is a static snapshot. The market has moved on by the time you finish typing, and it won't know. Never mind the dimension of what you can absorb — "a 30% drawdown is within the normal range" may be statistically unobjectionable, but to somebody who put their living costs in, that sentence means nothing, and it has no way of telling which of the two you are.
· If anybody else asking the same question would get the same answer, then it isn't advice for you.
· An answer genuinely aimed at you has to have your position, your cost basis and the loss you can absorb in it as variables. The model has none of those variables, so it cannot produce an answer of that kind.
· You, however, do have them. So the correct division of labour is: let it lay out the options and what each one costs, and you put your own numbers into it.
Limit three: the more specific the answer, the more likely it was invented
The dangerous error isn't it saying "I don't know", it is it inventing on the spot — and inventing in great detail. The three most common shapes:
| What gets invented | What it looks like | Why you believe it |
|---|---|---|
| Menu paths | "Open the app → Trade → More → Strategy Center" | The hierarchy looks real and so does the wording; it is only the final entry point that doesn't exist |
| Parameter and field names | An API field that looks perfectly standard | The naming style is identical to the real documentation, and nobody questions a field that "just looks right" |
| Numbers | A fee, a cap or a threshold given to several decimal places | Precision is itself a hint that "this was looked up"; if it were vague, it wouldn't dare say it |
There is something counter-intuitive here: you are on your guard against vague errors and not against precise ones. Tell you "roughly a tenth of a percent or so" and you will probably go and check; hand you a number to three decimal places and you will probably take it straight into your calculation. To a human brain, precision is a signal of reliability; to the model, precision is just having filled the slot where a number belonged with something that looks like a number.
It isn't lying to you. It is continuing along the probabilities of language: a sentence has reached the point where a field name belongs, so it generates a string that looks like a field name. Which makes the standard a blunt one — anything that could be falsified by looking at a specific page (steps, field names, numbers) should not be used directly.
So what is it actually good for? Two lists: ask and don't ask
Translate the three limits above into everyday use and you get the two lists below. The ones on the first list I use almost daily.
Safe to hand over to it:
- Explaining a concept. Have it make a term clear, then follow up with "give me an example where this concept doesn't apply". A working counter-example means it has caught the edges of the concept rather than reciting a definition.
- Translating error messages. Paste an error response in and ask what the line is saying and what usually causes it. It does this quickly and well, because the meaning of an error code is relatively stable and doesn't change when an interface is redesigned.
- Generating a checklist. Describe the operation you are about to perform and ask for "ten things to confirm before starting". You won't have thought of everything on its list, and anything it missed you will spot at a glance — the two complement each other neatly.
- Restating your thinking to find the holes. Talk it through your strategy logic and have it restate it, setting out each implicit assumption. The places it restates wrongly are usually the places you hadn't thought through. This is the use of it I find most underrated.
- Reading code and sketching a skeleton. Have it explain what somebody else's order-placing logic is doing and which edge cases have been left out. If it is really going to run, you still have to read it line by line yourself.
Not to hand over to it:
- Giving levels and calling direction. It has no real-time data and no reason at all to know better than you where the next candle goes.
- Working out your actual profit and loss. It doesn't know your fee tier, your real fill price, or whether the order filled partially. The figure it produces looks tidy and doesn't reconcile with your account.
- Confirming how a feature is reached right now. See limit one. The right destination for questions like that is the platform's own Help Center.
- Judging the rules, the fees or regional availability as they stand today. Same as above — and being wrong about these costs real money.
Both lists compress into one line: let it handle "what does this mean", and don't let it handle "what should I do now".
Two habits that take under two minutes
Together these two take under two minutes, and they block most of the holes.
- Paths and numbers go back to the official page for a check, without exception. If it gave you steps, open the Help Center of the platform you are going to use and walk through them as described; if it gave you a fee or a cap, turn to the fee page and compare. Note that it has to be the one you are actually using — entry points and parameters for features with the same name often differ between platforms, and taking one platform's description over to another leaves you tapping at nothing.
- Have it separate out the parts it isn't sure about. Add a line to your question: "list separately any parts you are unsure about or are inferring." This doesn't guarantee honesty, but in practice it does move some of the guesswork into that list — and that list is exactly what you need to go and check.
There is a more fundamental layer worth a sentence of its own: even if every rule it tells you is right, the price it is looking at may not be the same number as the one you are looking at. Where the data comes from, which definition of price is being used, and how long ago it was taken — those three differences turn "the market at one moment" into several versions of itself — and they are worth checking on whichever platform you actually use.
And if you intend to take a step further and wire the model's output into a program that runs by itself, there is a whole runtime layer to get through: disconnects, rate limits, parameters quietly going stale, none of which anybody will warn you about. See Three Ways Automated Trading Fails: Disconnects, Rate Limits and Parameter Drift.
One last thought
The three limits once more. The knowledge has a cut-off date, and it will carry on telling you expired things regardless; it has none of your account context, so it cannot produce advice aimed at you; it will describe paths, fields and numbers that don't exist in convincing detail, and "convincing detail" is exactly the form that gets past people most easily. These three aren't defects in one product, they follow from how models of this kind work, and switching vendor or version leaves them in place.
A serviceable mental model is this: treat it as somebody who has read an enormous amount, has never seen your account, and isn't sure what today's date is. Somebody like that can help you straighten out a concept, fill out a checklist, and question your ideas until the holes show; but you wouldn't have them place your orders.
Read next: for where the numbers a model reads actually come from, see Data Sources, Latency and Price References; to hand a strategy over to a program, start with Three Ways Automated Trading Fails: Disconnects, Rate Limits and Parameter Drift; for something more hands-on, there is Can You Buy US Stocks with USDT? and No Brokerage Account? How to Buy Apple and Tesla Stock With a Crypto Wallet.
This page contains no referral links and no invite code; where it says "the platform" or "the one you use", that is generic and doesn't point at any particular one. For this site's disclosure policy and risk disclaimer, see Disclosure and Risk Notice.