Research used to be the work of "typing keywords into a search box and reading and comparing the links that came up, page after page." With AI, this flow changes dramatically. Ask what you want to know in words, and AI reads multiple pages, sums up the key points, and even shows you the sources — you have AI do the prep of the research itself, and the human moves to the side of verifying "is that answer really correct?" That said, this chapter has the single biggest caution found in none of the others: "fact-checking." AI will happily mix in plausible-sounding lies. That's exactly why you should read this chapter with correctness, not speed, as your axis.
The goal: research fast, and always verify
Researching in the era of "delegating search to AI"
Traditional search was a stack of manual steps: "think of keywords → search → open a link → read → open yet another page." Research using AI is fundamentally different in that it rolls this whole flow into a single question. Ask "Name three video-editing apps for beginners, including whether they're free to use. Add a one-line note on what each is and isn't suited for," and AI finds the candidates, organizes them, and returns an answer. What you do narrows to two things: posing a good question, and verifying the answer that comes back. Still, this doesn't mean "search has become unnecessary." If anything, you'll use traditional search alongside more often, to confirm AI's answers. The two aren't competitors — they divide the labor.
AI search and summary — pose a question, get key points and sources
Even "researching with AI" involves several tools. Broadly, there are two kinds: the web-search feature of chat AI, and AI tools specialized for search. Both share the same trait: "pose a question, it goes and looks at the web, sums up the key points, and returns them with sources attached."
ChatGPT, Gemini, and the like answer by searching the web themselves as needed. Their strength is that you can continue with "look this up too" within the flow of conversation. You can research seamlessly alongside deeper discussion.
Tools built around search from the start. Each sentence of the answer carries a number showing which page it referenced, so the sources are easy to trace. Suited to research where fact-checking matters.
The knack for choosing between them is simple. For light research as an extension of discussion or brainstorming, use chat AI; for research where you want to verify the sources properly, use the search-specialized kind. That said, at first just trying the web search in the chat AI you already use is plenty. Which tools are free and how far they go is collected in how to use free AI tools.
An actual question looks like this. The key is to explicitly ask it to "show the sources."
"Does remote work raise or lower employee productivity? Sum up the key points in 5, based on recent studies. For each, attach a source link to which study or article you referenced. If there are points where opinions are divided, show both sides."
Just adding "attach source links" and "show both sides" greatly changes how verifiable the answer is. Without sources, there's no way to check the answer.
💡 Traditional search isn't going away. AI summaries are superb as an "entry point." You can grasp the big picture in tens of seconds. But the accuracy of the details is another matter. For information you'll use in an important decision, actually open the source links AI shows and read the originals — this one extra step decides the quality of research in the AI era.
Compare and organize — put options into a criteria table
The most time-consuming part of research is comparing multiple options. Tools, services, methods — with three or four candidates, just going back and forth between their official pages and lining up the criteria can eat half a day. This is AI's forte. Ask "put these into a table with the same criteria aligned," and it organizes the scattered information into a single comparison table.
"Pick three task-management tools for small teams and put them into a comparison table. Columns: 'rough price / free plan available / team size it suits / features / cautions.' At the end, in one line with reasons, recommend which a team adopting one for the first time should choose. If numbers or pricing are uncertain, write 'needs checking' and don't state them as definite."
The trick is to specify the columns (criteria) yourself. Leave it to AI and important angles get missed. The line "if uncertain, write needs checking" makes the later fact-checking much easier.
Listing pros and cons follows the same idea. Ask "for these three options, give the merits, demerits, and best-fit situations of each in bullets," and options that were fuzzy in your head get visually organized. But — the pricing, figures, and feature availability that come out here must not be believed as-is. The next section is the single most important part of this chapter.
Organizing scattered information into the same form. Filling gaps in the angles. Making a first draft in an instant.
The exact figures of the latest pricing and specs. Use it on the premise that you'll always confirm these on the official page.
Fact-checking and hallucination defense (most important)
This is the one thing to take away from this chapter even if you forget everything else. AI can answer as though something that doesn't factually exist were true. This is called "hallucination." A nonexistent book title, a wrong statistic, a feature that doesn't exist, a made-up court case — and its tone is utterly confident. So the trick of "only doubt the parts that sound unsure" does not work.
🚨 The core principle: AI's answer is a "draft," not a "conclusion." In particular, numbers, proper nouns, dates, and the latest information are the four things AI gets wrong most easily. If you'll use AI's answer for a work decision, an external communication, or a money-related choice, always verify against primary sources (official sites, originals) first. Skip this and you'll spread AI's lie in your own words.
So how do you verify? It's not hard. Just make it a habit, in order.
Actually click and read the references AI showed. If a link is missing, won't open, or its content contradicts the answer, doubt that information.
Prices, statistics, people's names, product names — confirm once more with traditional search. The key is whether you can reach a primary source (official or the announcer).
Add to the prompt: "If you're not certain, answer that you don't know." Just not forcing an answer visibly reduces the lies.
Point ③ especially is a knack that pays off big to know. AI tends to get pulled toward "I have to answer something," and the result is a plausible fabrication. Just giving it permission in advance to "say you don't know if you don't know" puts the brakes on this runaway.
"Tell me what's known about the following topic. But for anything that isn't certain, clearly write 'this is uncertain.' Don't phrase things that can't be stated as fact as if they were definite. When giving numbers or proper nouns, add sources if possible. And if you don't have enough information, don't force an answer — say 'I don't know.'"
It's handy to remember this line as a "set phrase for research." The reliability of the answer goes up a notch. For the basics of asking AI well, also see things to watch when entering input into AI.
💡 We doubt not because "AI is bad." Hallucinations are, given how today's AI works, unavoidable. Even as performance improves, they won't hit zero. So "using it while doubting smartly" is the right way to get along with it. For what to do when things go wrong, see common AI usage troubles and how to handle them.
Market, competitor, and industry research
For research that goes a bit deeper — say, "I want to know an industry's trends" or "organize the features of competing services" — AI shows its strength in making the first draft. It drafts in minutes the work of gathering material from zero and putting it into a table. The human can spend their time on scrutinizing, fleshing out, and judging it.
"Give an overview of the online-learning-service market, understandable to a beginner. In 3 items: ① what kinds of players there are ② what users seem to value ③ the direction of recent change. When giving numbers, add sources, and for anything unconfirmed, mark it as 'conjecture.' Don't mix fact and conjecture."
The overview you get this way is, at best, a map for starting your research. Only after a human confirms each point — "is this number real?" "is this player still a major one?" — does it become material you can use.
✅ Be conscious of the division of labor. AI is good at organizing "broad, shallow, fast." The human's job is to judge "narrow, deep, correct." Have AI sketch the big picture and a human fact-check the key spots — this combination makes research the most efficient, and the safest.
Pitfalls — cutoff, plausible lies, source quality
Finally, let's lock in the three pitfalls that most easily trip you up when researching with AI. Know the mechanism and you can avoid them.
AI only has knowledge up to when it was trained. It's weak on very recent events, prices, and new products. Insist on "latest info with web search," and always confirm the real thing for anything where freshness is critical.
It can create nonexistent sources, numbers, and names with full confidence. Smooth prose is no guarantee of correctness. Always fact-check numbers and proper nouns.
What AI references is a mixed bag. An individual's guess and an official announcement can get blended as equals. Look at "whose information, and from when" before deciding whether to accept a source.
All three are prevented by the single point: "don't turn AI's answer into your conclusion." For the cutoff, use the latest primary sources; for plausible lies, confirm the sources; for low-quality sources, apply the "who's the source?" lens. Once you're used to it, it's tens of seconds of effort. And that effort is exactly what turns AI from a "handy but risky tool" into a "partner you can trust."
⚠️ Don't settle for "I researched it fast." The real goal of AI research is "fast and correct." Take only the speed and skip the fact-check, and you'll use wrong information with confidence — the worst-case pattern. Aim to put the time you saved on speed into fact-checking, and you'll have it about right.
- Research enters the era of "pose a question, get key points + sources." Choose between chat AI for research as an extension of discussion, and search-specialized for source-focused research.
- Have multiple options summed up into a comparison table with aligned criteria. Specify the columns (angles) yourself, and treat numbers on a "needs checking" basis.
- Most important is fact-checking. ① Open the sources ② re-search numbers and names ③ ask it to "say you don't know if you're not certain." Numbers, proper nouns, dates, and the latest info are especially dangerous.
- For market and industry research, use the division of labor: AI drafts, a human scrutinizes.
- Three pitfalls — knowledge cutoff, plausible lies, source quality. All are prevented by "don't turn AI's answer into your conclusion."
With the previous chapter, Chapter 5, "Data, spreadsheets, and analysis," you handle numbers, and this chapter gives you the power to gather information. Now for the grand finale. In the next chapter, Chapter 7, "Delegating with AI agents," let's move from one-off instructions to a way of working that delegates a whole sequence of work.