ChatGPT Trading: What a Language Model Helps With, and What It Does Not Do

ChatGPT trading works for a narrow set of jobs and fails at a wider set, and almost every expensive mistake here comes from mixing the two up. A general purpose language model is good at explaining a prop firm rule in plain language, turning a risk plan into concrete numbers, reviewing the logic of an expert advisor, structuring a trade journal and drafting the questions to send your firm's support desk. On its own it does not place an order, see your balance, track your live distance to the drawdown limit, or know whether the strategy it described survives real trading costs. The first list is analysis, the second is execution, and treating the first as the second is how traders burn evaluation fees.

That answer sits between the dismissive version, which says a chat model is useless for trading, and the promotional one, which says you can prompt your way to a funded account.

One disclosure first, because it should change how you read the last section. This blog is published by PraxAI and we sell trading automation to prop firm traders, so we have a commercial interest in where you draw that line. Judge the argument, not the conclusion. For the wider survey, start with our overview of the [best AI trading bots for prop firms in 2026](/blog/best-ai-trading-bot-prop-firms-2026).

These are not consolation prizes. Each addresses a real reason people fail evaluations, and each is a language task rather than a market task. The common thread is that you can check the output in a minute, against the firm's own page or your own arithmetic. The trouble starts when you cannot.

This describes what a chat window is connected to by default, not a permanent verdict on any product. Vendors add tools, data feeds and broker integrations, so check what a specific product is wired into before relying on it. The structural point underneath does not move: something has to be connected, running and accountable at the moment a limit matters, and a conversation is not that thing by itself.

Together they reduce to one sentence: a chat model can describe a trade and cannot be responsible for one. If it suggests a size that leaves you a fraction of a percent from the daily loss limit, nothing objects, because nothing knows. Whether automation clears an evaluation is a separate question, answered in [can AI pass a prop firm challenge](/blog/can-ai-pass-prop-firm-challenge).

This is the costliest error in the ChatGPT trading category. You ask for a strategy and get an entry rule, an exit rule and a risk figure. Then you ask how it performed and receive a win rate, a profit factor and a maximum drawdown. Those numbers were not measured: no data loaded, no trades simulated, no spread applied. They appeared because your question implied numbers should follow.

They fool people because they look plausible rather than absurd. A fabricated backtest rarely claims a 99 percent win rate, it claims 61 percent with a profit factor of 1.8, and nothing marks it as unmeasured.

The response is a hard rule: any performance figure that did not come from a testing engine or a live account statement does not exist. Run the test yourself with realistic spread and commission, then treat even that with scepticism, because the gap between a tuned historical run and a real account is its own problem, laid out in [why backtests and live EA results diverge](/blog/backtest-vs-live-ea-results). Ask directly whether anything was run.

Set the invented numbers aside and there is a second problem in the strategy itself. A model produces the most probable continuation of your prompt, so what you get is the average of everything ever written about trading: moving average crossovers, standard indicator combinations, round number targets. Generic rather than wrong, and generic ideas are already priced into the cost of trading.

Cost is where most of these ideas die. A rule with a small average win per trade can look profitable on paper and be negative once spread, commission, swap and slippage are charged against every entry. That is often the entire edge. Nothing in a chat window knows your spread on gold at the London open or your slippage during news, so it cannot say whether the idea survives your costs, a gap covered in [slippage and execution on a prop firm EA](/blog/slippage-execution-prop-firm-ea).

An evaluation is also not won by the best expectancy. It is won by the run that reaches the target while staying far from the limits, so hand the model your constraints first and make headroom the goal.

Using AI to write an EA is now common, and the sensible position splits by which part of the code you are asking for. For structural code it is a real time saver: the boilerplate of an MQL5 file, iterating over open positions, formatting a log, handling a symbol suffix. That code either compiles and behaves or it does not, and the strategy tester tells you quickly.

For the parts that decide risk, the same tool is dangerous, because the failures are silent. Sizing that mixes up points and pips, or forgets the contract size on gold, produces an order that executes perfectly and is five times too large. A daily limit check that resets on the wrong timezone passes every casual test, then fails once with money behind it.

So the workable practice inverts how most people use it. Have the model review risk code you wrote or bought, and be far more careful about shipping risk code it generated. If you do, test it deliberately: force the daily limit condition and confirm the software stops. The wider framework is the [checklist for automating an FTMO account safely](/blog/automate-ftmo-safely-checklist).

Put the two lists side by side and the architecture is obvious. Use the language model where the output is text you can verify, and deterministic software where the output is an order or a limit, because that layer has to behave identically every time, whether or not anyone is watching. It is also why the phrase AI trading misleads here. Much of the software sold under that label is fixed rules in code, and for risk enforcement that is the better design. A cap that decides whether today is an exception is not a cap, a distinction set out in [whether any trading bot is actually AI](/blog/is-any-trading-bot-actually-ai).

The same split explains why passing is the easier half. An evaluation is a closed problem with a published target and published limits. A funded account has no finish line, and it is commonly cited that only 1 to 3 percent of funded traders keep the account long term. The practical version of that is in [how to keep a funded account](/blog/how-to-keep-funded-account). A chat model helps you understand the problem. It is not what solves it at three in the morning.

Since we sell into this category, here is our position plainly. PraxAI is the executor half, not the analyst half. It runs on MetaTrader 4 and MetaTrader 5, with a cTrader cBot that is new and in validation with the first clients, and NinjaTrader 8 for futures, where whether you may automate at all depends on each firm's policy and must be confirmed with that firm in writing. PraxAI GUARD holds the daily loss and drawdown limits you set as code level locks, and shuts the system down before the firm's limit rather than at it. That part is deliberately not AI, because for a hard limit you want predictable ahead of clever.

For evidence, we publish 24 approved challenge accounts at praxai.io/results, each with the interactive equity curve and the real maximum drawdown displayed next to the gain, across firms including FTMO, The5ers, E8 Markets, FundedNext and FundingPips. One named example: an FTMO 100K at plus 10.28 percent with 1.96 percent maximum drawdown over 12 days. The figure we point at is that the highest maximum drawdown across all 24 accounts is 3.35 percent, because that describes the room left when the target was reached. Those are our own accounts published on our own results page, not a third party audit, and you can open each one and read the curve. None of it promises that any particular challenge will be approved.

Keep using the chat model for the analyst work, where most traders underuse it. Just do not ask it to hold a limit, because it is not in the loop when the limit matters. Whatever software you put in that loop, ours or not, judge it by one standard: passed accounts with the maximum drawdown shown beside the gain, and limits you set yourself.