What AI Can and Cannot Do in Financial Research

Analytical software, including systems marketed as artificial intelligence, can retrieve text, cluster paragraphs, and draft a summary of a table. Those are reading aids. They are not a warrant that the draft is true, complete, or suitable as a household instruction. This figure is about that gap. North Peak Capitals does not use a model as an adviser, and it does not ask readers to automate a buy or sell decision from a prompt.

The educational temptation is to treat fluency as authority. A fluent paragraph about a filing can still invent a number, drop a qualifier, or mash two years together. Literacy after 2020 includes this new genre of error alongside the old ones.

A draft, not a proof

A model that predicts the next word can produce a sentence that looks like a citation. Looking like a citation is not the same as pointing at a page you can open. Research, even beginner research, still needs a source that exists outside the chat window. If the source cannot be opened, the sentence is a hypothesis about a source.

This is not an argument against using tools. It is an argument against retiring the habit of Fig. 04 — units, windows, authors — just because the restatement arrived in a second. The faster the restatement, the more the checking has to be explicit, because there is less time for doubt to happen by accident.

What tools can support

They can help you find a document you then read. They can list section headings so you know where a risk discussion sits. They can translate a dense paragraph into simpler English that you then verify. Those uses stay on the educational side of the line. They become something else if the tool is asked “what should I do with my savings?” — a question this site will not answer and a question a model is not entitled to answer for you.

Provenance problems

Training data are a mixture of dates, qualities, and incentives. A model may not know when a page was updated. It may mix a textbook definition with a marketing page. It may omit the jurisdiction that makes a sentence true in one country and false in another. Canada, for example, has its own regulatory vocabulary; a fluent US-centric paragraph can sound universal while being local.

Provenance means being able to say where a number came from. If the answer is “the model,” you do not yet have provenance. You have a generator. Put the generator beside the table. If they disagree, the table wins until you have a reason to distrust the table — a reason that is also a document, not a vibe.

Hallucinated citations are a special case of the same problem. A title that sounds like a working paper is not a working paper until you can open it. Beginners should treat unmatched citations as missing sources, not as extra reading they have not found yet.

Speed as a new opacity

Speed hides labour. A human summary takes time; during that time a reader can notice a missing axis. Instant text removes that pause. Opacity is not only “we do not know the weights.” Opacity is also “we did not sit with the doubt.” Educational use of AI should restore a pause on purpose: a checklist, a second source, a printed page.

• Does every figure appear in an openable source?
• Is the window stated?
• Is the jurisdiction stated?
• Has a qualifier (“estimate,” “unaudited,” “backtest”) been dropped?
• Has the draft turned a description into an instruction?

If the last box is ticked, stop. Instruction is outside this volume even when a human writes it. It is more outside when a model writes it, because there is no one to hold to a licence.

A checking drill

Take a public statistical table. Ask a tool to summarise it. Then, without looking at the summary, write three facts you can see with your eyes. Compare. Circle mismatches. That circle is the lesson. Repeat with a methodology page, which models often flatten. The drill teaches distrust that is specific — distrust of uncited numbers — rather than a vague fear of computers or a vague faith in them.

Another drill: ask the tool for a definition of a term used in this volume, then compare with a textbook or a regulator glossary. Notice the extra adjectives. Adjectives are where advice likes to hide (“attractive,” “resilient,” “compelling”). Educational prose can live without them.

A third drill is to ask the same question twice, hours apart. If the numbers move without a new source, you have watched a generator, not a database. Stability is not truth, but instability without a citation is a warning light.

Conclusion

AI can support financial research as a librarian and a drafter. It cannot stand in for a source, a licence, or a household decision. Speed is useful only if the pause is put back by hand. The next figure, Fig. 06, gathers the checking habits from this whole volume into a single way of comparing claims — including claims that arrive with a machine sheen.