Spot a Confident Wrong Answer Before You Share It

The most convincing wrong answer is the one that arrives in a confident tone with a believable link. I do not judge factual work by tone. I check the pieces that could mislead a reader. For a deeper research workflow, Research Prompts That Separate Evidence From Inference separates what the source says from what you infer.

Names and dates: Are they current and correctly spelled? Numbers: Does the cited source actually contain the number and its denominator or sample? Quotations: Is the wording in the original source, not a paraphrase presented as a quote? Links: Do they open, and do they support the adjacent sentence? Inference: Is a conclusion labeled as an interpretation rather than a reported fact? Suppose the model says “most developers use this prompt pattern” and cites a survey. I open the survey and inspect how many developers were sampled, what they were asked, and whether the cited percentage actually refers to that pattern. If not, I remove the claim rather than rewriting it to sound safer.

I can ask: List the five claims in this draft most likely to require verification. For each, give the exact sentence, a proposed primary source, and what the source would need to show. Do not claim verification has happened. That makes a check queue. It is not the check itself.

When I cannot verify a specific statistic or release date, I either find an authoritative source or remove the detail. A useful article does not need invented numbers to feel alive. Try it: Pick one AI-written paragraph. Check every number, quote, date, and link. Mark each supported, unsupported, or uncertain. Publish only after correcting or removing the latter two. Keep learning: Research Prompts That Separate Evidence From Inference, How to Prompt Across Models Without Pretending They Work the Same, and How to Tell Whether an AI Answer Is Any Good.