Using AI to Research and Synthesize

The concept

AI is a powerful research accelerant, but it is not a research replacement — the distinction matters more than almost anything else in this module. An accelerant means the AI does the fast, mechanical parts of research (reading a pile of material quickly, structuring it, spotting patterns across sources) while you still own the judgment of which sources matter and whether the claims in them are actually true. A replacement would mean trusting the AI's synthesis as the final word without checking it — and because of how LLMs generate text (see the hallucination lesson), that's exactly the failure mode to avoid.

The most useful research prompts fall into three categories. Synthesis prompts ask the model to find structure across material you provide: "what are the 3 main arguments across these sources," "summarize the areas these two reports agree and disagree on," or "produce a comparison table of price, features, and limitations from these product pages." Gap-finding prompts ask what's missing: "based on these five articles, what questions about this topic are left unanswered?" or "what would a skeptic say is missing from this argument?" Outline-generation prompts turn raw research into a starting structure before you write anything: "organize these notes into a logical outline for a 1,000-word brief."

The one rule that makes all three of these safe to use is: feed the AI the actual source material (paste the articles, documents, or transcripts into the prompt, or clearly name specific sources) rather than asking it to research a topic from its own training data and report back. When the AI is synthesizing text you gave it, it's doing a much more reliable job — pattern-matching and summarizing what's in front of it — than when it's trying to recall facts from training, which is where hallucination risk is highest. And even with provided sources, treat every specific factual claim, number, or quote in the synthesis as something to verify against the original — synthesis can still misattribute or overstate a point even when the underlying material is real.

It also helps to separate two different research jobs the AI can do. The first is compression: taking a pile of material you've already decided matters and making it faster to digest. The second is triage: helping you decide what's worth reading in the first place, by skimming a long list of headlines or abstracts and flagging the handful most relevant to your specific question. Triage prompts ("given these 20 article titles and one-line summaries, which 5 are most relevant to a decision about switching CRM vendors, and why") are a quieter but very high-leverage use case — they save you from reading everything, not just from re-reading what you already gathered.

A final habit worth building early: keep a running list of every source you fed the AI, in order, alongside the synthesis it produced. When you go back to verify a claim weeks later, or when a colleague asks "where did this come from," you want to be able to point at the original material instantly rather than reconstructing which of twelve browser tabs was the source.

Worked example

A learner researching a laptop purchase for work gathers three competing product reviews. They paste all three into an AI assistant with the prompt: "Based on these three reviews, produce a comparison table with columns for price, battery life, build quality, and reviewer recommendation. Then list any points where the three reviews disagree." The AI returns a clean table and flags that two reviews praise the battery life while the third reports it as mediocre — a genuinely useful synthesis that would have taken much longer to extract by re-reading all three reviews manually.

Before acting on it, the learner spot-checks two things: they open the original review that claims the best battery life and confirm the AI's number (quoted as "14 hours") actually appears in that review (it does, word for word) — and they check the review that disagreed, confirming it really does report "around 7 hours under heavy use" rather than the AI having invented a contradiction. Both check out. Had either number been wrong, the learner would correct the table manually before using it to make a purchase decision, and would be more cautious about the rest of that synthesis.

The learner also tries a gap-finding prompt on the same three reviews: "what questions about this laptop are these three reviews not answering?" The AI notices that none of the three mention the laptop's repairability or how easy the battery is to replace — a genuinely useful gap, since it's exactly the kind of long-term-ownership question reviews often skip. The learner adds "repairability" to their own follow-up search before finalizing the purchase decision.

Exercise

Pick a real research question from your job or a purchase decision you're actually facing. Gather three to five source articles, reviews, or documents — real ones, not a hypothetical. Paste them (or clear excerpts) into an AI assistant and ask it to produce a structured synthesis: a comparison table, a list of areas of agreement/disagreement, or an outline, whichever fits your question best.

Then verify at least two specific factual claims from the output by checking them against the original sources directly — not by asking the AI "are you sure," but by opening the source yourself and confirming the number, quote, or claim actually appears there. Finally, run one gap-finding prompt on the same material ("what's missing from these sources relative to my question?") and note whether it surfaced something genuinely useful or just restated what you already knew.

Recap

• AI excels at structuring and summarizing research material you provide — synthesis, comparison, and gap-finding are strong use cases.
• Feeding the AI your actual source material (not asking it to recall facts from training) makes the synthesis far more reliable, but doesn't eliminate the need to verify.
• Always spot-check specific claims, numbers, and quotes in a research synthesis against the original sources before acting on it — never skip the check.

Resources

• Anthropic: "Prompting for research and synthesis tasks" — https://docs.anthropic.com/en/docs/build-with-claude/prompt-engineering/overview
• Perplexity: AI research/answer engine with cited sources — https://www.perplexity.ai/
• Poynter: "IFCN fact-checkers' guide to verification" — https://www.poynter.org/ifcn/