Researching With AI
Gather, summarize, contrast, verify — in four stages
① Throwing "summarize this" at an AI produces a plausible but unverified summary. The problem isn't the AI — it's handing over the whole job as one lump.
② Split research into scoping → gathering → contrasting → verifying, with a prompt suited to each stage, and the output changes character.
③ If you keep one principle: AI is not a source, it's a signpost to sources.
① Why "summarize this" fails
The common pattern: name a topic, ask for a summary. What comes back is fluent and well structured. And you still can't use it, because there's no way to tell where fact ends and inference begins.
The reason traces back to how the model works. It appends one plausible token at a time. Ask it to research, judge, summarize, and write in a single request and all four blend into one generation stream. Fact and guess get joined seamlessly inside the same sentence, and can't be separated afterwards.
Well-written prose lowers the reader's guard. That's exactly where AI summaries are dangerous. An awkward passage prompts you to check; a smooth one gets accepted. Looking high-quality and being verified are entirely different properties.
② The four-stage workflow
Stage 1 — Scoping (the most-skipped step)
Don't start with research. Start by building the list of what needs researching. This is something AI is genuinely good at: it requires no fact-checking, and models are strong at surfacing angles you'd miss.
"I'm about to research [topic]. I haven't gathered any material yet.
① List the sub-questions I'd need to answer to understand this properly.
② For each, note what kind of source would answer it (filings, statistics, papers, news).
③ Flag separately any angle beginners typically miss.
Don't answer any of them yet — just build the question list."
That last line matters. Without "don't answer yet", the model will barrel straight into answers and the whole point of staging is lost.
Stage 2 — Gathering
Work the question list from stage 1 one item at a time. Not batching them is the point: each answer has to stay short enough to verify.
A search-enabled tool helps here. Without one, find the material yourself, paste it in, and confine the model to it.
"Answer using only the material below. If something isn't in it, write 'not in the material' — don't infer.
Question: [one question from stage 1]
Material: [paste]
For each claim, note which paragraph of the material it came from."
Two elements carry this prompt: "say so if it isn't there" blocks guessing, and "cite the paragraph" cuts verification down to thirty seconds.
Stage 3 — Contrasting perspectives
This is where AI research earns its keep. Have it deliberately construct the opposing case.
Researching alone, it's easy to get anchored on the first framing you encounter. Having AI lay out the other side breaks that bias cheaply. And because this is organizing perspectives rather than establishing facts, the hallucination risk is comparatively low.
"Here's what I've concluded so far: [summary].
① Lay out the case someone would make against this conclusion.
② Point out which of my premises are unverified.
③ Tell me what fact, if discovered, would mean this conclusion is wrong."
③ is especially valuable. Deciding in advance what evidence would falsify your view means you only need to monitor that one indicator afterwards — and it keeps research from expanding indefinitely.
Stage 4 — Source verification (never skippable)
Close the AI and open the primary documents yourself. Three things to check.
| What to check | Where | Why |
|---|---|---|
| Every number | Issuing agency, company IR, original filing | Digits are predicted as tokens, never computed |
| Every citation and link | Click the URL yourself | Real author + fabricated title is a common pairing |
| Anything recent | Dated primary material | Post-cutoff content may be inference |
For the specific techniques, apply the five steps from our hallucination guide directly.
1️⃣ Scope — "just build the question list, don't answer yet"
2️⃣ Gather — one question at a time, "say so if it's not in the material"
3️⃣ Contrast — "the counter-case? my unverified premises? what would falsify this?"
4️⃣ Verify — close the AI; check numbers, citations, and recency at the source
③ Applying this to investment research
Used for a company or sector, the division of labour becomes especially clear.
| Stage | Give to AI | Do yourself |
|---|---|---|
| Scope | Which metrics matter in this industry, competitive structure, regulatory variables | — |
| Gather | Structuring pasted earnings material, explaining terminology | Obtaining filings and IR material |
| Contrast | Building the bear case, poking holes in your premises | — |
| Verify | — | Every figure: revenue, margins, ratios |
Asking AI "should I buy this stock" sits outside this workflow entirely. The model doesn't know your portfolio size, risk tolerance, or horizon, and has no live prices. AI helps assemble and structure the inputs to a decision; the decision itself is yours.
For figure-based analysis, work from source-grounded material like our Big Tech earnings guide or circular financing analysis.
④ How much time does this save — honestly
This workflow takes longer than "summarize it for me." Four prompts instead of one, plus source-checking at the end.
The savings land elsewhere. Stage 1 (working out what to examine) and stage 3 (finding the counter-case) take hours alone and still leave gaps. You gain heavily there and spend a little in stage 4.
And decisively, the output has a different standard of reliability. The gap between "the AI said so" and "I checked the filing" isn't the kind of thing that converts into hours.
⑤ Closing
AI's place in research, in one sentence: it is not a source, it is a signpost to sources.
Telling you where to look, catching the angle you missed, structuring what you've gathered — it does all of that well. But the final answer to "so what's actually true?" always lives in the original document. Hold that one line and everything else is upside.
※ This article reflects information as of July 2026 and is not investment advice. The workflow described concerns research method; all investment decisions and their consequences rest with the investor.
※ This guide is provided for general educational purposes and simplifies technical details for readability.
New guides, when they land
We publish AI literacy guides twice a week. Subscribe and the next one comes to you — free, unsubscribe anytime.