Asking an AI assistant one research question usually brings back one point of view, the most common one. Stanford researchers tested that habit and confirmed the problem. A confident single answer can still be tunnel vision dressed up as research.
Their system, STORM, was published as peer-reviewed research in 2024. It starts by working out which perspectives exist on a topic. It then acts out conversations, with a writer from each perspective questioning an expert who draws on real web sources. Because a first answer changes what someone asks next, STORM keeps asking follow-up questions. The answers from every perspective are pulled into an outline, then built out into a full article with citations.
The team tested STORM’s outlines against a strong comparison system that also searched the web. STORM’s outlines scored 25 percentage points higher for organization and 10 percent higher for how much of the topic they covered. Experienced Wikipedia editors who tried the system all agreed it helped their research before the writing started.
STORM needs its own web search and citation tracking, code most people will never install. The idea underneath needs none of that. Asking a general AI chatbot to argue with itself from several disagreeing angles gets a rough version of the same benefit. A guide for everyday users packs that into a short sequence of instructions that runs inside Claude, Anthropic’s AI assistant.
The sequence gives Claude five roles built to disagree. A practitioner explains what works day to day, and an academic reports where studies contradict popular belief. A skeptic makes the strongest case against the mainstream view. An economist follows the money and names who profits, and a historian points to the closest past example and how it played out.
Four typed instructions run the sequence in one chat. The first has Claude answer as all five experts. The second lists every point where two experts contradict each other and rates how much each one matters for a real decision. The third combines the views into a short brief that keeps the contradictions visible. The fourth has Claude critique its own brief harshly, flag disputed claims stated as settled fact, and rewrite it. Running all four takes about five minutes.
Only Stanford’s tool went through those tests. The four-instruction shortcut brings no web sources of its own, so it leans entirely on what the AI already knows.
