Prompting for research

A research assistant, not an oracle

AI assistants are great at helping you get oriented: explaining unfamiliar terms, mapping out the main questions in a field, summarizing long documents, and poking holes in your arguments. They are much less reliable as a final source of facts.

The reason is how they work. A language model generates text that is likely given its training and your prompt. Usually that lines up with reality. Sometimes it produces something that sounds exactly like a fact but isn't — a hallucination. Models also have a knowledge cutoff, so recent events may be missing unless the tool can search the web, and even then it can misread or misattribute what it finds.

The practical rule: use AI to find and understand; use sources to confirm.

The hallucinated citation problem

Ask a model for "five peer-reviewed studies that prove X" and it may oblige — with authors, journals, years, and page numbers that look perfect. Some may be real. Some may be real authors attached to papers they never wrote. Some may not exist at all.

You can reduce the risk with how you ask:

The improved version removes the pressure to "prove" a conclusion, asks for both sides, and makes uncertainty acceptable to express. That doesn't eliminate hallucinations, but it makes them less likely and easier to spot.

Summarizing documents you provide

The most reliable research use of AI is working with text you give it. When the source material is right there in the prompt, the model doesn't need to recall facts from training — it can read. This idea, grounding the answer in supplied documents, is the core of retrieval-augmented generation.

Grounding works best when you explicitly limit the model to the provided text:

Below is [a report / article / transcript]. Using ONLY this text:

1. Summarize the main argument in 3–4 sentences.
2. List the key evidence the author gives, with a short direct quote for each.
3. List any claims the author makes without supporting evidence.
4. If something I ask about isn't covered in the text, say "not stated in the document" rather than filling the gap.

Document:
"""
[paste text]
"""

Asking for direct quotes is a powerful habit: you can search the document for each quote and confirm it's there and in context. A paraphrase is much harder to check.

Asking for counterarguments

One of the most underused research prompts is asking the AI to argue against you. Models tend to go along with the framing of a question, so if you ask "why is X true?", you'll get reasons X is true. Flip it on purpose.

Here's my current position and reasoning:
[your claim and your reasons]

Act as a thoughtful critic who disagrees. Give me:
- The three strongest counterarguments, stated as fairly as possible.
- The evidence that would most weaken my position, if it exists.
- Any assumptions I'm making that I haven't defended.
Don't soften the critique, and don't invent statistics.

This is sometimes called "steelmanning" the other side: presenting the best version of an opposing view, not a weak caricature. Use it to stress-test essays, proposals, and decisions.

A research workflow that holds up

1. Orient — ask for an overview, key terms, and the main debates.
2. Find — use search engines, library databases, or AI tools with search to locate real sources.
3. Read with help — paste the sources in and ask for grounded summaries with quotes.
4. Challenge — ask for counterarguments and gaps.
5. Verify — check every fact and citation you plan to use against the original.

The AI speeds up every step, but the verification step is yours. That's what makes research trustworthy.