Research with AI: sources first, summary second

There are two ways an AI app can answer your question, and the difference matters for research.

The first: an answer from memory. The model writes from patterns it learned in training. No document is opened. Such an answer can be fine for general, slow-changing matters, but it cannot be traced anywhere, and for new or specific things the risk is high. The previous lesson explains why.

The second: an answer that retrieves sources. Before the model writes, pieces of documents or pages are put into the conversation, and the model answers with that material in front of it. They get there in different ways. Sometimes the model itself uses a search tool: it writes search terms, the app searches, and the results come back to it. This is called tool use. But retrieval is not always a tool the model calls. Many work apps search the office document collection themselves as soon as you ask, then paste the closest pieces into the request before the model sees it. To find the pieces closest in meaning, the app usually stores each piece as an embedding (the text's meaning turned into a row of numbers) and compares it with your question. This pattern is often called RAG (retrieval-augmented generation: the model answers from documents retrieved first). And when you paste text or upload a file yourself, that is retrieval too; you are the one who fetched it.

An answer that retrieves sources has one big advantage: you can check it. There is a link, a document title or a quote you can open. But do not misread this. Retrieving a source is not the same as being right, and hallucination can still happen. The model can still misread, blend two sources, quote an out-of-date page, or attach a link to a claim that is not on that page at all. Search can also surface weak sources: unnamed blogs, sites copying other sites, or forums.

So the rule is simple. The model's memory is a starting point, not evidence. A retrieved source is candidate evidence, and becomes evidence only after you open it yourself and find the sentence that supports the claim.

Checking a source takes three questions, and each can be answered in a minute or two.

Does the source exist? Click the link. A dead link, a page without the title mentioned, or a document you cannot find by its title is a warning sign. A model can produce links that look convincing.

Does the source support the claim? Find the sentence. Use find-on-page with two or three keywords from the claim. Read the paragraph around it. Often the page is on the same topic, but the figure differs, the condition is narrower, or it says the opposite.

Is the source fit to rely on? For rules, use the text from the government site or the body that issued it, not a blog summary. For official figures, use the body that collects the data. For office policy, use the latest version of the document from its owner. Secondary sources such as news articles can help you find the official one, but should not be your final word.

Finally, note the date you opened the source. Pages change, rules are replaced, figures are updated. A note saying "accessed on such a date" tells the reader, and you next month, how long the answer can be relied on.