10 AI Research Prompts That Actually Work

AI has a research problem. Not a capability problem. A confidence problem.

Ask it a question and it answers immediately, fluently, and with the quiet certainty of someone who has never once said "I don't know." That's useful when you want a first draft. It's a trap when you need the truth. The difference between a good AI research prompt and a bad one isn't about clever wording. It's about whether you're using AI to find answers or to confirm the answers you already wanted.

This guide gives you 10 copy-paste AI research prompts for desk research, plus the safety rules that stop you from publishing something embarrassing. The prompts are organized in order: from clarifying your question all the way through to building a decision memo. Use them in sequence or grab the one you need.

Before we get into the templates, one rule that runs underneath all of them.

The core mistake people make with AI research

They treat AI like a search engine that gives better answers. It isn't.

A search engine returns links. You decide which ones are credible. AI returns synthesized prose that sounds like the conclusion, skips the sources, and buries any uncertainty in confident-sounding language. That's great for organizing material you already have. It's dangerous when you're trying to discover something true.

The mental model that actually works: AI is a fast, tireless intern. Exceptional at summarizing, comparing, formatting, and drafting. Terrible at knowing when it's wrong, citing real sources unprompted, or telling you that your original question was bad. What AI can and can't do is worth reading if you want the full breakdown. The short version: the judgment is yours. The legwork is theirs.

Every prompt below is built on that split. You bring the sources, the context, and the question. AI organizes, compares, and drafts. You verify, decide, and sign your name to the conclusion.

What not to paste into AI before we go any further

This matters and people skip it.

Do not paste: confidential company strategy, private customer data, unreleased product plans, proprietary research, personal medical or financial information, regulated data of any kind. That means no client lists, no patient records, no legal case details, no HR notes, no NDA-covered materials.

Anonymize before you paste. "Our client, a regional hospital" becomes "a mid-size healthcare organization." Specific revenue numbers, internal projections, personal names from private contexts: leave them out or genericize them.

And one more thing. AI will invent citations if you let it. It will give you a convincing author name, a plausible journal, a real-looking year. None of that means the source exists. Verify every citation, date, statistic, and quote against a primary source before it goes anywhere near something you're publishing or presenting.

If you're unsure whether your workplace allows AI tool use at all, that's a separate conversation worth having. Whether to tell your boss you use AI has become a real workplace question, and the answer depends on your organization and what you're working on.

The 10 AI research prompts, ready to copy

Prompt 1: Turn a vague question into a research plan

You know the general topic. You don't know what you're actually trying to find out. Use this first.

Prompt: I'm researching [topic]. My goal is to [decision I need to make / output I need to produce]. Here's what I already know: [brief summary]. Help me define a focused research question, identify 5-7 subtopics I should investigate, and flag 3 assumptions I might be making that I should test. Don't answer the research questions yet. Just help me structure the inquiry.

The last sentence is important. If you don't add it, the AI will just answer the questions. You want the plan first.

Prompt 2: Build a source checklist

Prompt: I'm researching [topic] for [purpose]. What types of primary and secondary sources should I look for? List them by category: official data, industry reports, academic research, expert commentary, practitioner examples. For each category, give me 2-3 specific places or organizations I might look. Flag which source types are likely to be biased toward a particular conclusion and why.

This gives you a structured sourcing plan before you touch a single document. Run it, then go find the actual sources yourself. Don't ask AI to find sources for you without verifying they exist.

Prompt 3: Summarize a provided article

Prompt: Here is the full text of an article I'm researching: [paste text]. Summarize it in 150 words. Then list: the main claim, the evidence used to support it, the methodology or basis for the evidence (survey, study, expert opinion, case study), and any notable limitations or caveats the author mentions. Don't add any information from outside this article.

That last instruction is critical. Without it, AI will happily blend the article with everything else it knows, and you'll lose track of what the source actually said.

Prompt 4: Extract claims and evidence

Good research separates claims from evidence. Most documents blur the line.

Prompt: Here is a document I'm analyzing: [paste text]. Create a table with three columns: Claim, Evidence Provided, and Evidence Type (data/expert opinion/anecdote/analogy/assertion). List every significant claim the document makes. If a claim has no evidence, mark the Evidence Provided column as "None stated."

This prompt alone is worth the price of the article. It makes bad arguments visible immediately.

Prompt 5: Compare two sources

Prompt: Here are two sources on [topic]: [paste Source A] [paste Source B]. Compare them across these dimensions: main argument, evidence quality, areas of agreement, areas of contradiction, and what each source seems to be missing. Present your comparison as a table. Do not synthesize a conclusion. Just show me the comparison.

Again: no conclusion yet. You make the conclusion. The AI makes the table.

Prompt 6: Find missing viewpoints

This one catches the blind spot in your own research.

Prompt: I'm researching [topic]. My current sources all come from [describe the type of sources you've used: industry reports, academic papers, news coverage, etc.]. Based on the topic, what perspectives, stakeholder groups, or types of evidence am I likely missing? Be specific. For each gap, explain why that perspective would matter to a complete picture of this issue.

Dee covers the judgment problem in Don't Replace Me: the human work is deciding which sources matter, which assumptions are weak, and which conclusions are actually useful. Spotting your own gaps before someone else does is exactly that kind of work.

Prompt 7: Prepare interview or customer research questions

Prompt: I'm about to interview [describe person: role, context, what they know] about [topic]. My research goal is [what you're trying to understand]. Here's what I already know or suspect: [brief summary]. Write 10 open-ended questions for the interview. Prioritize questions that test my assumptions rather than confirm them. Avoid leading questions. Flag which questions might feel sensitive to ask.

This is good for customer discovery, stakeholder interviews, and expert conversations. It's not a replacement for doing the interview. It's preparation.

Prompt 8: Build a decision memo from research notes

You've done the research. Now you need to turn scattered notes into something a person can act on.

Prompt: Here are my research notes on [topic]: [paste notes]. My audience is [who will read this]. The decision we need to make is [specific decision]. Organize my notes into a decision memo with these sections: Background (what we know), Key Findings (what the research shows), Open Questions (what we still don't know), Tradeoffs (options and their implications), and Recommended Next Step. Mark any finding where the evidence is thin or uncertain. Do not add information from outside my notes.

The instruction to mark thin evidence is one you should never skip. It's the difference between a memo that helps people decide and one that overconfidently misdirects them.

Prompt 9: Fact-check a draft

This prompt works on your own draft or someone else's.

Prompt: Here is a draft document: [paste draft]. Identify every specific factual claim: statistics, dates, named studies, quotes, organizational claims, historical assertions. For each one, flag whether it's verifiable from context in the document, requires external verification, or appears to be an unsupported assertion. List them in a table. Do not verify the claims yourself. Just identify which ones need checking.

You then go verify the flagged items against primary sources. Don't ask the AI to verify them. It will often produce a confident-sounding wrong answer.

Prompt 10: Build an evidence table

Prompt: I'm writing a [report/article/presentation] arguing that [your thesis]. Here are the sources I've collected: [paste summaries or excerpts]. Build an evidence table that shows: Source, Claim from source, How it supports or challenges my thesis, Evidence strength (strong/moderate/weak, with brief reason). Order the table from strongest to weakest evidence. Then list any claims in my thesis that have no supporting evidence in my sources.

That last instruction will show you where you're arguing from hope rather than evidence. Fix those gaps before you present, not after.

Which prompts work best for which research tasks

Different jobs call for different prompts. Here's how they map to common work situations.

Use prompts 1 and 2 at the start of any unfamiliar topic. Use prompts 9 and 10 before anything goes to a decision-maker. Prompts 3, 4, and 5 are the workhorses you'll return to constantly.

A note on research versus writing

People conflate these two things and it causes problems.

Research is finding out what's true. Writing is communicating what you found. They're different jobs, and AI is better at one than the other.

For writing, AI is genuinely fast and useful. You paste your notes, it drafts a structure, you edit. For research, AI is a trap if you're asking it to generate the underlying facts. It doesn't have access to real-time data (in most configurations), it can't call an expert, and it has no way of knowing that the study it's confidently citing was retracted two years ago.

The prompts in this article are designed for the research phase, not the writing phase. They work on material you bring to the conversation. If you want prompts for the writing and communication side of work, the broader guide to AI prompts at work covers that territory.

The one place the line blurs is the decision memo (Prompt 8). That's a writing task built on your research. It works because you're pasting your notes and asking AI to organize them, not asking AI to know things.

Quality rules that apply to every single prompt

Verify before you publish. Every statistic, every date, every named study, every quote. If it came from AI, assume it might be wrong until you've checked it against the original source.

AI invents references. It doesn't mean to. It just pattern-matches toward what a citation in this context might look like. A realistic-sounding author, journal, and year doesn't mean the paper exists. Search for it. Find the primary source. If you can't find it, don't cite it.

Don't let AI write your conclusion. Use it to organize evidence. The conclusion is yours. You're signing your name to it.

Keep the "do not add outside information" instruction in any prompt where you're working from specific documents. Without it, the AI blends your sources with everything else it knows, and you lose track of provenance.

The actual skill here isn't prompting

It's asking good questions in the first place. Knowing which sources are credible. Spotting when evidence is thin. Understanding the context that makes a finding relevant or irrelevant to your specific situation. Recognizing the difference between "the data says X" and "this data, with these limitations, in this context, suggests X."

Those are the skills that make you hard to replace. AI makes a mediocre researcher faster. It doesn't make a fast researcher good. The human skills that matter in research, including judgment, source evaluation, and intellectual honesty, are exactly what AI can't fake consistently.

The prompts above are tools. The tool doesn't decide what question to ask. It doesn't decide what evidence is trustworthy. It doesn't take responsibility for the conclusion. You do.

Frequently asked questions

Can AI find sources and citations for desk research?

Not reliably. AI can suggest types of sources to look for and organizations that might publish relevant data, but it frequently invents specific citations, including plausible-sounding author names, journal titles, and publication years that don't exist. Always find and verify primary sources yourself before citing anything in work you'll publish or present.

What's the safest way to use AI for research without hallucinations?

Paste the source material yourself and instruct the AI to work only from what you've provided. Add "do not add information from outside this document" to any prompt where you're analyzing specific texts. Use AI to organize, compare, and summarize, not to generate facts from memory.

Should I use ChatGPT or Claude for research prompts?

Both work for the prompts in this article. Claude tends to be more cautious about flagging uncertainty. ChatGPT is slightly more fluent in formatting tables and structured outputs. The more important factor is how you write the prompt. A good prompt works in either tool. See the broader guide to AI prompts for work for more on this.

Can I paste company data or client information into AI research prompts?

No, unless your organization's policy explicitly permits it and you're using an approved enterprise tool with appropriate data handling guarantees. Strip out or anonymize any confidential strategy, customer data, financial projections, HR information, or regulated data before you paste. When in doubt, leave it out.

How do I know if the evidence AI finds in my documents is actually there?

Keep the original source open. When AI produces a summary or claim extraction, check it against the source text directly. If the AI attributes a claim to a source you pasted but you can't find that claim in the original, it may have blended your source with outside knowledge. The prompt instruction "do not add outside information" reduces this but doesn't eliminate it entirely.

Is researching with AI cheating?

That's the wrong question. Using AI to organize notes, summarize articles, or build comparison tables is the same category of work as using a spreadsheet to sort data. The question is whether the judgment, sourcing, and conclusions are yours. If you're letting AI decide what's true without verifying it yourself, that's not cheating. It's just bad research.