Research with AI without losing track of the evidence

Turn a broad topic into a research question

“Tell me about AI and jobs” can produce an attractive but untestable essay. “What evidence distinguishes task automation from job elimination in office work?” invites a more disciplined comparison. Specify the audience, geography, period, and decision the research should inform.

Begin with a question map: what is known, what is disputed, and what evidence would settle the decision. Ask the assistant for search terms and likely source types. A provider's product page can establish an advertised feature; it cannot independently establish that the product is best for everyone.

Build a small source collection

Collect a few strong sources before asking for a synthesis. Record title, author or organization, publication date, link, and the exact question each source helps answer. Include a contrary source when the subject is contested. Ten articles repeating the same press release are not ten independent pieces of evidence.

Document assistants are useful for working within a selected collection. Search assistants are useful for finding additional material. Use each for its purpose and check whether the system actually accessed the document or merely recognized its title. Images, tables, and inaccessible PDFs may require separate inspection.

Ask for a claim-by-claim comparison

Using these sources, compare the evidence for and against the claim. For each important statement, name the source and identify the supporting passage. Separate study findings, the authors' interpretations, and your own inference. List missing evidence and do not resolve disagreements by guessing.

Read the passages that drive the conclusion. Check whether a study measured a simulated task, real workplace performance, or a forecast. Notice who was studied and whether the result applies to your situation. An impressive percentage may conceal a small sample or a narrow outcome.

Write a decision-ready summary

Put the answer first, then explain the strongest evidence and the limitations. If two sources disagree, explain the differing definitions, dates, methods, or incentives. Avoid averaging incompatible numbers just to produce a single figure.

Finish with what would change the conclusion and which facts should be rechecked later. Preserve the source list with the summary. If you share it with friends, label your own interpretation and include enough context for someone else to challenge it. A good research note should help the next reader think, not require them to trust the assistant's confidence.