AI for Research

The biggest reason AI research disappoints is a lazy question. "Tell me about the project management software market" produces a bland essay you could have guessed. A framed question names four things: the decision you're making, the scope (who's in and out), the format you want back, and the depth you need.

> I'm deciding whether to build a scheduling feature or integrate
> a third party. Research the three most common ways software
> products handle scheduling. For each: typical approach, when
> teams choose it, main downside. Keep it to a short table.
> Flag anything you're unsure about.

You named your decision, bounded the scope, fixed the format, and asked it to flag uncertainty.

Research goes faster when you stop inventing structure each time. Ask the AI to build a framework first, then fill it — this separates "what should I look at?" from "what's the answer?", and the framework is reusable forever.

> Design a competitive-analysis framework for comparing
> customer-support software. List the 6 to 8 dimensions a buyer
> should compare on. Don't fill them in yet — just give me the
> reusable checklist. Then compare [Tool A], [Tool B], [Tool C]
> against it in a table, one row per dimension. Where you lack
> reliable information, write "needs verification", not a guess.

The "needs verification" instruction is gold: it gives the AI permission to admit ignorance instead of inventing a plausible figure — the behaviour that gets professionals into trouble.

Research isn't reading one article — it's reconciling several that don't fully agree. AI is genuinely excellent at this, if you give it the sources rather than relying on its memory.

Paste in two or three articles, then ask for synthesis:

> Here are three articles on remote-work productivity trends.
> [paste articles]
> Synthesise them: where do they agree, where do they disagree,
> and what's the single most important takeaway for a team lead?
> Attribute each claim to the source it came from.

"Attribute each claim to the source" anchors the AI to the material you provided instead of blending in half-remembered facts. When every claim is tagged, you can verify it — and you can see which claims rest on nothing.

A hallucination is when AI states something false with complete confidence — a made-up statistic, a fake quote, a study that doesn't exist. It's the biggest risk in AI research, and it never sounds wrong, which is what makes it dangerous. Learn to feel the difference between three kinds of output:

The tells of a likely hallucination: a suspiciously precise statistic ("73.4% of buyers"), a named study with no findable author, a confident answer about very recent events, or a direct quote. When you spot these, probe the AI directly:

> For each claim in that summary, tell me your confidence and
> whether it comes from the documents I gave you or your general
> knowledge. Mark anything you can't source as "unverified".

A well-behaved AI often backs down when challenged — "On reflection, I can't verify that figure." That climbdown just saved you from quoting a number that doesn't exist.

You don't verify everything to the same depth — that would be exhausting. Verify in proportion to risk: the more a claim could embarrass you or steer a decision, the harder you check it.

The loop for any claim that matters: isolate it (a name, number, quote, or date), find the original (the actual report, not a summary), match it (does the source really say this?), and note where you confirmed it. Let the AI build the checklist for you:

> Pull out every factual claim in that briefing I should verify
> before sharing externally — numbers, quotes, dates, and named
> sources. List them so I can tick each off against a primary
> source.

This turns a vague worry ("is any of this wrong?") into a concrete, finite list — instead of staring at a wall of text wondering what to trust.

Strong research isn't just collecting answers — it's noticing what's missing. AI is a good partner here, because it has no ego about admitting the picture is incomplete. Once you've gathered findings, turn the AI into a critic:

> Here's the research I've assembled for a market-entry decision.
> [paste your findings]
> Act as a sceptical reviewer. What's missing? What would a
> cautious executive ask that this doesn't answer? Where am I
> relying on a single source or an untested assumption?

This "devil's advocate" move catches the blind spots confirmation bias hides. When you've spent an afternoon building a case, you stop seeing its holes — a fresh pass surfaces them. A useful output looks like this: