AI in Finance · Lesson 01

The Two AIs

Almost every "AI in finance" headline collapses into one of two very different technologies. Tell them apart and you can read the news critically today — before you know a single application in detail.

The win By the end of this page you'll be able to take any AI-in-finance claim, sort it into predictive ML or generative AI, and name the risk that comes built in with that kind. That one sort is the backbone of the whole course.

One word, two machines

"AI" in a finance headline is doing one of two completely different jobs, and they fail in completely different ways. The confusion — and most of the hype — comes from a single word covering both.

Predictive ML · old · works

It scores or forecasts

Learns patterns from historical numbers in tables to answer a narrow question about something new.

Is this transaction fraud?
Will this borrower default?
Where is this price heading?

Decades old. Behind most AI already running in banks. Output is a prediction with an error rate, not a fact.

Generative AI · new · shaky

It produces language

Predicts plausible next words to generate text or code. The post-2022 wave — ChatGPT and kin.

Summarise this 200-page filing
Draft this client report
Answer this customer in chat

New and powerful. Output is fluent but sometimes confidently false — it can hallucinate.

The FSB draws exactly this line: firms have long used AI "to enhance internal operations and improve regulatory compliance" — that's the predictive kind — while generative AI "has given rise to new use cases, such as document summarisation, information retrieval, and code generation" FSB 2024. Two waves, two risk profiles.

Why the split is the whole game

Because the failure modes don't transfer. Ask the wrong question and you can't judge the claim:

	Predictive ML	Generative AI
Eats	Tables of numbers	Text (and produces it)
Gives	A score / a number	Words, code, a summary
Age	Decades, mature	Since ~2022, immature
Signature failure	Bias, overfitting, black-box	Hallucination — confident and wrong
Right question to ask	"How often is it wrong, and is it fair?"	"Who checks the output before it's used?"

A fraud-detection story and a chatbot story are both "AI in finance," but a smart question about one is a nonsense question about the other. Sorting first is what lets you ask the smart one.

The hype tell When a pitch says "AI" but won't tell you which — what it learns from, what it outputs, how often it's wrong — that vagueness is usually the product. The term for dressing ordinary software up as AI is AI-washing, and "won't name the kind" is its clearest symptom.

The systemic twist

There's one risk that appears because the tools are shared. When thousands of firms lean on the same handful of AI providers, GPUs, and cloud platforms, and train on the same data, two dangers stack up: concentration risk (one provider fails, everyone's hit at once) and correlation risk (everyone's models behave alike, so a wobble becomes a stampede). The FSB warns that common models and data "could lead to increased market correlations, amplifying systemic risks during crises" FSB 2024. Keep this one in your pocket — it's the risk the excited coverage almost always skips.

Check yourself

Answer from memory. A miss here is worth more than a scroll back up.

Now do it to real headlines

This is the skill, not the definitions. Find three recent "AI in finance" headlines — a news site, a vendor's homepage, a LinkedIn post, whatever. For each, answer three questions out loud:

The three questions — try each headline before opening

Which kind? Does it score/forecast from numbers (predictive ML), or produce language/code (generative)? If you genuinely can't tell from what's written, that itself is the finding — note it.
What's the built-in risk? Bias / overfitting / black-box for predictive; hallucination for generative; concentration/correlation if it's about scale.
What are they not saying? How often is it wrong? What does it train on? Who checks the output? The gap is where the hype lives.

Did it work? Bring your three headlines to your teacher and say how you sorted each one — especially any you couldn't sort. That "couldn't sort" pile is the most useful thing you'll produce this lesson; it's exactly where the vague claims hide, and reading a few together is worth more than another definition.

Read this next

Primary source: FSB — The Financial Stability Implications of Artificial Intelligence (2024). The best single landscape document, written by a body with nothing to sell. Read the executive summary (a few pages) — it lays out the use cases and the systemic risks in exactly the frame this lesson uses.