← All notes AI & data

AI in research: the useful bits

Jul 28, 2026 · 6 min read · The C·tradis desk

A grounded look at what machine learning does well in financial research, and the failure modes worth knowing before you trust it.

The genuinely useful jobs

Machine learning earns its place on well-scoped, high-volume tasks: standardising inconsistent data, extracting facts from long documents, translating and summarising, and flagging unusual patterns for a person to examine. These are jobs where scale and consistency beat intuition, and where a human still reviews the output.

The narrower the question, the more reliable the tool tends to be.

How it goes wrong

Trouble starts when a tool is asked to do more than it can justify. Models overfit historical noise, mistake correlation for cause, and can be quietly poisoned by data leakage that makes backtests look far better than reality. Many also cannot explain themselves, offering an answer with no checkable reasoning behind it.

An unexplained result is a starting point for investigation, never the end of one.

Human judgement stays central

The responsible pattern is simple to state and harder to practise: treat model output as a claim, trace it to a source, and test it against how the world actually behaves. Incentives, context and scepticism remain human responsibilities that no current tool removes.

Teams that respect these limits get a genuine edge; teams that ignore them get confident mistakes at scale.

Key terms

This note is general educational information only and is not financial, investment, legal or tax advice, and not a recommendation to buy, sell or hold anything. See our Risk Disclaimer. Have a correction or a topic to suggest? Write to the desk.