Discernment: reading and evaluating AI output

• Check AI-generated work for accuracy by breaking claims into pieces you can verify one at a time, numbers you can recalculate, facts you can look up, and citations you can find in the actual source.
• Recognize the common ways AI output goes wrong: invented citations that sound real, patterns that fit the data but miss the real cause, and claims that aren't supported by the source even when the source exists.
• Read an output through three practical lenses: is the content accurate, is your back-and-forth with the AI actually improving it, and will it behave correctly when real people use it.
• Judge whether your back-and-forth with the AI is actually improving the output, and change your approach when it isn't, sometimes the problem is your prompt, sometimes it's the task you chose.
• Build a personal checklist you can run before you release AI-assisted work, so nothing goes out with your name on it until you have checked it.

Key concepts · tap to reveal

• “If an AI output reads fluently and cites sources, it is probably accurate and can be trusted without verification”

  Fluency and citation format are not reliable indicators of correctness. Current large language models fabricate metrics, invent citations, and miscalculate derived quantities even when the output appears authoritative. Discernment requires decomposing the output into atomic claims and verifying each claim against authoritative sources, regardless of how confident or well-formatted the response appears.