What Good AI Output Looks Like (It Is Rarely What You Expect)
After two years of reading AI drafts, the tells of quality are not polish or grammar. They are three quieter things.
When people evaluate AI output, they evaluate polish: grammar, flow, structure. Polish is table stakes and nearly meaningless. After two years of working with generated text daily, I grade on three quieter signals, and they predict usefulness far better.
The first is specificity that could not have been generated without my input. If a draft contains a detail that traces back to something I told it, the model is working with me. If every sentence could appear in any document on the topic, I am reading the average of the internet, and the average of the internet is worthless by definition.
The second is the handling of what it does not know. Good output marks its uncertainty: 'I am not sure about the timeline here' or an explicit gap. Confident output that fills every hole is worse than honest output with visible holes, because the holes I can fill and the confident errors I have to hunt. This signal appears in the first paragraph, which makes it a fast test.
The third is resistance. The best outputs I have received argued with me: pushed back on my framing, flagged that my structure buried the point. Sycophancy feels pleasant and produces nothing. A draft that disagrees well has already done the thinking I was trying to avoid.
What I no longer grade: grammar (trivially fixable), length (a trim away), and tone (one instruction). People reject drafts for surface reasons and accept drafts that are confidently hollow. Flip that attention around and your hit rate with the same tools roughly doubles.