What is trading?
CROSS-CUTTING: COURSE 1 | LESSON 1
What AI can and can't do for traders
Learning objectives
Explain what a large language model (LLM) actually does — and why that means it can be confidently, fluently wrong.
List the four hard limits of AI assistants for traders: hallucination, knowledge cutoffs, no live data, no predictive power.
Identify the tasks where AI genuinely helps a trader (process support) versus the tasks it should never be trusted with (signals and forecasts).
The honest starting point
If you take one thing from this whole course, make it this: an AI assistant is a language machine, not a market oracle. Tools like Claude, ChatGPT and Gemini are large language models — systems trained on enormous amounts of text to predict what words plausibly come next. That training makes them remarkably good at explaining, summarising, structuring and questioning. It does not give them a view of the future, a feed of live prices, or any obligation to be correct.
That sounds like a limitation — and it is — but it's also a map. Once you know exactly what these tools can't do, the things they can do become genuinely valuable to your trading process.
The four hard limits
1. LLMs hallucinate — plausibly. When a language model doesn't know something, it doesn't say "error". It produces the most plausible-sounding answer it can, in the same confident tone it uses when it's right. Ask one for "EUR/USD's average daily range last month" and it may hand you a precise-looking number like "72.4 pips" that it simply invented. The danger is not that AI is wrong sometimes — every source is — it's that its wrong answers look identical to its right answers. There is no nervous cough, no hedging, no red flag.
Example prompt: "What was the RBA's cash rate decision at its last meeting?" Plausible AI answer: "At its most recent meeting, the RBA held the cash rate at 4.35%, citing persistent services inflation."
That answer might be right. It might also be last year's rate, or the wrong meeting, delivered with total confidence. You cannot tell from the answer alone. You verify it against the RBA website — every time.
2. Knowledge cutoffs. Every model was trained on data up to a certain date, then frozen. Anything after that date — rate decisions, elections, a broker changing its swap policy — simply doesn't exist for the model unless you paste it in or the tool has a live search feature (and even then, verify what it retrieved). If you ask "what's driving gold this week?" without giving it this week's news, the model can only guess from old patterns. It will still answer. That answer is fiction wearing a suit.
3. No live prices unless you provide them. A plain LLM cannot see your chart, your platform, or today's quote. If it tells you "USD/JPY is trading around 148", that's a memory from training data, potentially months stale. The correct workflow is always: you bring the data (paste the price, the calendar, the journal rows), the AI helps you work through it.
4. It cannot predict markets. Full stop. This is the limit people most want to argue with, so let's be blunt. If any system could reliably predict short-term price moves, its owners would not be selling chat subscriptions — they would be quietly running the world's most profitable fund. LLMs are trained on text about markets, much of it wrong, contradictory or outdated. Asking "will EUR/USD go up tomorrow?" gets you a fluent essay of generic factors, or worse, a confident guess. Neither is a signal. Anyone selling "AI trading signals" from a chatbot is selling you noise with good grammar.
So what is it actually good for?
The real value is process support — the unglamorous work that separates disciplined traders from impulsive ones, and that most of us skip because it's tedious:
- Journal analysis. Paste 30 trades from your journal and ask the AI to group them by setup, session and outcome, and to ask you pointed questions about the patterns. It's tireless and it doesn't flatter you. (Lesson X1.3 builds this into a reusable skill.)
- Research digestion. Paste a 4,000-word central bank statement and get a structured summary of what changed versus last time — then check the quotes yourself. (Lesson X1.4.)
- Articulating your plan. Explaining your strategy to an AI that keeps asking "what exactly triggers your entry?" forces a precision most traders never reach on their own.
- Stress-testing rules. "Here are my trading rules. Ask me ten hard questions a risk manager would ask." That's a genuinely useful sparring partner. (Lesson X1.5.)
- Learning concepts. "Explain swap rates like I'm new, then quiz me" — excellent use, low risk, still verify specifics.
Notice the pattern: in every good use, you supply the data and the decisions; the AI supplies structure, summarisation and questions. The moment the AI is supplying the facts or the trade direction, you've crossed into the failure zone.
The two habits that make AI safe to use
Verify every number and every fact against the source. Rates against the central bank's site, prices against your platform, statistics against your own journal export. Treat AI output as a draft written by a smart intern who never checks their work.
Never paste sensitive data. No account numbers, passwords, API keys, or personal identity details — yours or anyone else's. Trade sizes and instruments are fine; credentials never are. Lesson X1.6 covers this in depth.
And the regulatory framing, which we will repeat until it's reflexive: AI output is not investment advice. Not from the AI, not from us. It's unvetted text from a pattern-matching machine. Your decisions, your risk, your responsibility.
Key takeaways
LLMs predict plausible text; they have no market view, no live data and no crystal ball.
Hallucinated answers look exactly like correct answers — confidence tells you nothing.
Knowledge cutoffs mean anything recent must be pasted in by you, then verified anyway.
The genuine value is process support: journaling analysis, research digestion, plan articulation, rule stress-testing.
Verify every number against its source, keep private data out of prompts, and treat all AI output as research assistance — never as advice or signals.