The old research risk was incompleteness. You might miss a source, overlook a counterargument, fail to find the relevant study. That risk still exists. But the new risk is different, and it's harder to manage because it doesn't look like a risk at all.
The new risk is confident incorrectness. The model produces summaries that are accurate in broad strokes and wrong in details that matter — and it produces them in the same assured, readable voice it uses for everything. There is no signal that distinguishes a reliable summary from an unreliable one. You have to find out by checking.
My research loop has three beats: narrow, wide, narrow again. I skip any of them at my peril.
Narrow — define the question
Start with one precise question. Not a topic. A question. "Why did [specific company] fail in [specific year], and what role did [specific factor] play?" beats "history of [company]" every time.
A topic invites a survey. A question invites an argument. Surveys are easy to generate and hard to evaluate — they produce lots of sentences that sound true without committing to much. Arguments make specific claims that can be checked against evidence. The model is much better at generating coherent surveys than coherent arguments, which is another reason to force the question form.
The narrowing at the start is about choosing what you are actually trying to find out, not what you are vaguely interested in. Those are different. The vague interest is where bad research begins.
When I get this step wrong — when I start with a topic instead of a question — the output I get is impressively comprehensive and useful for nothing. It maps a space without locating me in it.
Wide — map the disagreement
Give the model the question and ask for the shape of the answer: competing views, key disagreements, loudest voices on each side, what is actually settled versus what is still contested. Ask specifically about what people disagree on, not what they agree on. Consensus is cheap to summarize. Disagreement reveals where the real complexity lives.
Speed is a feature of the loop, not a replacement for it.
The model has read more than you have. Let it use that. At the wide stage, you are not looking for the answer — you are looking for the map of answers. Where are the fault lines? What does one camp say that the other rejects? What do they both assume that might not be true?
This stage produces a research agenda. A set of specific claims to check, specific sources to find, specific questions to take to primary material. Without this stage, the third step has no direction. With it, you know exactly where to look and what to look for.
Note what you are not doing at this stage: taking the model's word for any of it. You are using its synthesis to understand the terrain. The terrain might be wrong in places. That's what the next step is for.
Narrow again — chase one claim to its source
Pick one claim from the wide stage — one specific, checkable assertion — and chase it to a primary source. Not the model's summary of the source. The source.
This step is slow by design. It is supposed to be slow. The loop is not intended to eliminate reading — it is intended to tell you what is worth reading. The wide stage compressed hours of survey work into minutes. The second narrow stage puts that time back in where it actually matters: at the point of specific, consequential claims.
The model's summary of a source is almost always accurate in broad strokes and wrong in the details that matter. The details that matter are usually why the claim is interesting in the first place. The nuance, the qualifier, the specific context in which the finding holds — these are what the summary drops, because they make it longer and harder to read. They are also what changes the meaning.
Read the source. Read the part the model summarized. Then read the paragraph before and the paragraph after. Context is almost always partially missing from AI summaries.
The honest limitations
Some research questions require sources the model hasn't seen or can't reliably access. Recent events, proprietary data, paywalled specialist literature — these are the cases where the model is most likely to confabulate. It will still answer. It will answer confidently. That confidence is not a signal of accuracy.
Know which kind of question you're asking before you start. Is this a well-documented historical question where good sources are abundant? The loop works well. Is this a question about something that happened last month, or in a specialized domain with limited public literature? The loop is a starting point at best and a liability at worst.
For those questions, use the model to build your search strategy and then leave it behind while you execute. It can tell you where to look. It cannot always tell you what you'll find.
The loop is not a shortcut to answers. It is a shortcut to good questions — and that is more useful, in the long run, than most shortcuts.