Let me start with the thing everyone who uses AI for research needs to understand, and most people find out the hard way:
AI models make things up, and they do it confidently.
Not out of malice. It's a structural feature of how they work. They predict likely text, and sometimes a confident-sounding false claim is more statistically "likely" to follow a given prompt than an honest "I don't know." The technical term is hallucination, which is a polite word for fabrication.
Knowing this doesn't mean AI is useless for research. It means you need to use it differently than you'd use a search engine or a library database. Here's how.
What AI Is Actually Reliable For (Research Edition)
Start with an honest map of where the model is likely to be right versus where it might be confidently wrong.
Generally reliable:
- Explaining established concepts and theories — things that are covered extensively in textbooks, Wikipedia, academic papers, and general reference material
- Summarizing or paraphrasing ideas you've already verified from primary sources
- Helping you understand jargon or technical terminology
- Identifying what questions to ask and what angles to explore
- Comparing frameworks, schools of thought, or competing perspectives
- Thinking through implications and connections
Unreliable and often wrong:
- Specific statistics and numbers (it will quote a plausible-looking figure that may be made up or outdated)
- Citations and references (it fabricates paper titles, author names, and journal names with stunning confidence)
- Recent events (training data has a cutoff; anything after it is unknown)
- Niche or obscure information (the less it's been written about, the more the model fills gaps with invention)
- Anything where the exact wording matters (legal, medical, regulatory)
That last point about citations is worth emphasizing. If you ask an AI to provide academic sources, it will give you something that looks like a real citation. It is often completely made up — title, authors, journal, year, DOI, all of it. This has tripped up students, journalists, and lawyers. Don't use AI-generated citations without independently verifying that the source actually exists.
The Framework I Use
Think of AI as a research thinking partner, not a research source. Here's what that looks like in practice:
Phase 1: Orientation — Use AI to get a quick map of the topic. What are the main concepts? What are the key debates? What's the established consensus versus what's contested? This gives you a framework before you start reading.
Phase 2: Question generation — Ask AI what questions you should be asking. What would an expert in this field want to know? What does this topic connect to? This often surfaces angles you hadn't thought of.
Phase 3: Primary research — Do actual research using sources you can verify. Use the map from Phase 1 to guide what to look for, not to replace looking.
Phase 4: Synthesis — Bring your verified findings back to AI. Use it to help you connect ideas, spot inconsistencies, write up summaries, or stress-test your understanding. This is low-risk because you're feeding it verified material rather than asking it to generate facts.
Prompts That Actually Work for Research
The topic orientation:
I'm starting to research [topic] and know very little about it. Give me:
1. A one-paragraph orientation — what is this topic and why does it matter?
2. The 3–5 most important concepts I'll need to understand
3. The main debates or disagreements in this field
4. What this topic connects to — what adjacent areas should I be aware of?
Note: I'll be verifying facts from primary sources. For this stage I just want a map of the landscape.
That last note is useful — it signals to the model that you're not asking for verifiable facts but for orientation, which tends to produce more appropriately hedged output.
The question generator:
I'm researching [topic] with the goal of [your specific purpose — writing an article, making a decision, understanding a concept].
What are the 10 most important questions I should try to answer? Prioritize questions where the answer would most change my understanding or decision.
The expert interview simulation:
I want to understand [topic] from multiple perspectives.
Simulate a discussion between three different types of experts who might have different views: [e.g., an economist, a sociologist, and a policy maker].
For each, present their likely perspective on [specific aspect of the topic]. Flag where they'd disagree with each other.
This is useful for understanding a topic's contested landscape before you go looking for real expert opinions.
The synthesis assistant:
I've done research on [topic] and collected these key points from verified sources:
[paste your notes, bullet points, or excerpts]
Help me:
1. Find the common threads
2. Identify any contradictions or tensions in what I've found
3. Suggest what might be missing from my research
Here you're feeding AI your research rather than asking it to create facts, which is much safer.
How to Verify What AI Tells You
When AI gives you a fact that matters, here's the verification hierarchy:
For statistics and data: Look for the original source. If AI says "40% of small businesses fail in the first year," ask it where that comes from — then look up that actual study. If it can't tell you or the source doesn't exist, treat the number as unverified.
For academic claims: Search Google Scholar or Semantic Scholar for the actual paper. If AI cited a paper, verify the title, authors, and year are correct and that the paper exists. Read the abstract at minimum.
For historical facts: Check primary sources (original documents, firsthand accounts) or established reference sources (Britannica, major newspaper archives, government records).
For recent events: Don't use AI at all. Use a search engine with date filters and check multiple sources.
The quick gut check: If a fact seems surprising or highly specific, that's a signal to verify. The more impressive or counterintuitive the claim, the more likely the model is confabulating a plausible-sounding thing it doesn't actually know.
The "Explain Your Confidence" Trick
This is one of the most useful prompts I've found for research contexts:
For each claim in your response, tell me: how confident are you in this, and what would I look up to verify it?
Models often respond with useful self-assessment: "I'm confident in the general concept but less certain about the specific percentage — you'd want to look for [organization/study name] to verify that." That's actionable. You know what to follow up on.
Alternatively:
After your response, flag any claims where you're uncertain about the specific details, and suggest what search terms I'd use to find authoritative sources.
This doesn't make the model perfect, but it surfaces its own uncertainty in ways that are useful for deciding where to verify.
The One Hard Rule
Never cite AI as a source.
Not for academic work, not for professional reports, not for journalism, not for anything where accuracy matters. AI is a thinking tool, not a reference source. Use it to understand, orient, and synthesize — then find the primary sources and cite those.
This isn't just about intellectual honesty (though it is that). It's practical. If someone asks where your data comes from and your answer is "ChatGPT," that's the end of your credibility on the topic.
AI-assisted research is great. AI-sourced research is a trap.
Curious about how AI models handle uncertainty and why hallucinations happen at a technical level? The Intermediate Track has a lesson on avoiding hallucinations that explains the mechanics and more prompting strategies for reducing them.