How to Use AI for Research Without It Making Things Up
TL;DR
- Give it the sources. Do not ask it to remember them.
- A citation you have not opened is not a citation.
- Ask the same question two ways. Two different confident answers means neither is reliable.
- It is excellent at digesting material and unreliable at recalling facts. Use it for the first thing.
The failure everyone has heard about is AI inventing a source. It happens because these tools generate plausible text, and a plausible-looking citation is easy text to generate.
The fix is simple and almost nobody does it: stop asking it to know things, and start giving it things to read.
Source-grounded beats memory every time
There is a large difference between “summarize the current research on X” and “here are four papers, summarize what they say about X.”
The first asks it to recall, which is where invention lives. The second asks it to read, which is what it is genuinely excellent at. Same tool, completely different reliability.
Tools built for this
NotebookLM only answers from documents you upload and points at the page it took each answer from. If making something up would be a serious problem, that constraint is the feature.
The general assistants can do it too if you upload the files rather than asking from memory. The discipline matters more than the tool.
Ask for the quote, not the summary
When something matters, ask it to give you the exact sentence from the source that supports the claim.
If it cannot produce the sentence, the claim is not in the document. This one habit catches most problems, and it takes five seconds.
Cross-examine it
Ask the same question in a fresh conversation, phrased differently. Consistent answers are a decent sign. Two different confident answers means you need a real source.
You can also just ask “what would make this answer wrong?” It is often useful about its own limits when you invite it to be.
Where it genuinely shines
- Digesting a 90 page report into the six things that affect you
- Comparing several documents and surfacing where they disagree
- Turning research notes into a structured outline
- Explaining a dense passage in plain language
- Being the sceptic: “argue against this conclusion”
Notice these all involve material you supplied. That is the pattern.
Writing the report
Draft from your verified notes, not from a blank prompt. Feed it what you established, tell it who is reading and what they need to decide, and ask for a structure before any prose.
Approve the structure, then get the draft. Reports that go wrong usually went wrong at the outline, and fixing an outline costs a minute.
The one rule
Never publish or act on a fact you have not seen in the source yourself. Not the summary of the source. The source.
AI removes the reading time. It does not remove the responsibility, and that trade is still overwhelmingly worth it.
Want the source-grounded approach in depth?
The NotebookLM guide covers loading the right sources, asking questions across everything at once, getting cited answers, summarizing and organizing, generating audio overviews, and being honest about the limits of a specialist tool.