How to Use AI for Research Without Being Lied To: 5 Habits That Actually Work

How to Use AI for Research Without Being Lied To

Language models are useful for research and very confident at making things up. They summarize a 30-page filing in a paragraph, extract named entities from 1,000 tweets, and draft a brief in your voice. They also fabricate quotes, invent citations, and report patterns from their training data as if they were facts. The five habits below are the ones that keep the model in the loop without inheriting its blind spots.

A model that summarizes a document you do not have will happily invent quotes

The model is a compression algorithm, not a source. If a sentence is going to appear in your brief, it should be traceable to a record you control, not to a model output. This rule sounds pedantic until you watch a friend lose two days defending a quote that the model fabricated. Then it sounds like the only rule that matters.

Warning

Habit 1: Ground every prompt in a source

When you paste a document into the prompt, you are creating a new mini-corpus that the model reasons over. When you ask a model to "summarize the news about X," the model is reasoning over a vague memory of its training data — which may be months or years out of date, may include misinformation, and may include nothing at all about your specific case.

The first move is almost always the second one. Give the model the source, then ask. If you cannot give the model the source, you are not doing OSINT — you are doing a vibe check.

A useful pattern:

"I will paste a 30-page SEC filing. Summarize the risk factors in plain language, with page numbers. Do not add risks that are not in the document. If a section is unclear, say so."

The "do not add" instruction is doing real work. Without it, models will often fill in plausible context that sounds like the document and is not.

Habit 2: Treat the model as a junior analyst, not a source

A model can compress a long filing into a paragraph, list the named entities in a corpus, or draft a one-page brief in your voice. None of those are decisions. The model does not know whether the filing is the right one, whether the named entity list is biased, or whether the brief is true.

A useful mental model: the model is a junior analyst with a perfect memory of what you handed them and no idea whether the assignment is real. You are the editor. Their draft is a starting point, not a citation.

This means:

• The model summarizes; you decide.
• The model extracts; you verify.
• The model drafts; you redline.

The model is a tool. The judgment stays with you.

Habit 3: Never ask the model to cite itself

If a citation came from a model, treat it as a lead, not a source. Open the cited page. If the page does not say what the model said it said, drop the claim. Assume roughly one in five model-generated citations points to a real page that says what the model claims it says. The other four are fabrications, mistakes, or distortions.

A few prompt patterns to avoid:

• "Cite your sources." Produces URLs that look right and may not exist.
• "Pretend you are an expert in Y." Increases confidence without increasing accuracy.
• "What does the literature say about Z?" Produces plausible-sounding citations to papers that may not exist.
• "Find me a quote from [famous person] about [topic]." The quote is often invented.

A few prompt patterns that help:

• "From the attached filings, extract one row per company: name, registration, jurisdiction, date of incorporation, listed officers. Return as a markdown table. If a field is missing, write 'missing' — do not guess."
• "Translate the following 20 pages from Portuguese to English. I am only deciding whether to read the originals. Keep named entities, dates, and numbers exactly as they appear. Do not summarize."
• "I will paste 12 records from my spreadsheet, in the order they appear. Write a 400-word brief in plain English that supports the following claim. Use only the records I have given you. Where two records disagree, surface the disagreement. End with a one-paragraph 'what would change my mind.'"

The pattern in all three: the model is given a source, told what to do with it, told what not to invent, and asked to commit to a specific output. That is the move that scales.

Habit 4: Store the prompt, the input, and the redline

Three things to keep when you use a model for research:

• The prompt. The prompt is part of the record. A future reader — including you — should be able to reproduce the model's output from the prompt and the input.
• The input. The PDF the model read is the evidence. The output is a paraphrase of that evidence.
• Your redline. The places where you changed the model's draft are the analysis. A redline of model → final is often the most useful artifact in the project.

A research project that does not store the prompt and the input is a research project that cannot be re-run. A research project that does not store the redline is a research project that cannot show its work.

Habit 5: Separate triage from citation

A useful mental model: the model has two roles in your workflow, and they are not the same.

• Triage. Reading a long document, summarizing a corpus, translating for a quick check, extracting a candidate list of named entities. The model is fast and useful here. The output is a starting point. The output is not a citation.
• Citation. Anything that is going to appear in your brief, in your footnotes, in your social post, in your argument. The citation should be traceable to a record you control, not to a model output.

If you cannot tell which role the model is playing, the model is doing the wrong thing. The rule is: triage is fast, citation is slow. A model that produces a citation in a single round-trip is hallucinating; a model that produces a citation from a document you pasted, after you have checked the document, is summarizing.

When the model is genuinely the right tool

A few cases where the model is the best available tool, not a workaround:

• Translation for triage. The model can render 20 pages of Portuguese into English in a way that tells you whether to read the originals. It is not a substitute for the originals, but it is the right layer for "should I keep reading."
• Structured extraction. When the schema is clear and the input is messy — extracting a table of companies from a folder of filings, or a list of named entities from a transcript — the model is much faster than a custom script.
• Drafting from records you control. When you have a corpus of records in your own spreadsheet and you want a first draft in your voice, the model is a useful starting point. The redline is the analysis.

In all three cases, the model is grounded in something you provided, producing something you will edit, and producing an output that is not a citation. That is the move that holds up.