Research Prompts That Separate Evidence From Inference

Research answers become dangerous when a model blends what a source says with what it guesses. I want those categories separate, especially if I plan to publish. Before sharing research, Spot a Confident Wrong Answer Before You Share It gives you a quick audit of numbers, quotations, dates, and links.

Suppose I am comparing two game accessibility features. I ask: Using only the linked official design documentation and the supplied playtest notes, make three sections: verified facts, inferences, and open questions. For each verified fact, cite the specific source and date. For each inference, explain which facts support it. If a source does not answer a question, say so rather than filling the gap. This is not a magic anti-error formula. It is a way to inspect the answer. I open each source and see whether it supports the claim. If the link is broken or the cited section is unrelated, the claim stays unverified.

For a longer article I may use columns: claim, source, source date, confidence reason, and publication decision. The last column is my call: publish, revise, or remove. The model can help organize, but I decide after checking.

For product behavior, use official documentation. For research findings, use the original study and inspect the sample and method. For news or time-sensitive facts, verify the date and event. Google’s prompt design guidance encourages clear instructions; it does not replace checking whether evidence is real. Try it: Ask AI one factual question. Request verified facts, inferences, and unknowns separately. Open two cited sources and mark whether the exact claims are supported. Rewrite or remove anything unsupported. Keep learning: Spot a Confident Wrong Answer Before You Share It, Privacy Checks Before You Paste Anything Into AI, and How to Tell Whether an AI Answer Is Any Good.