The question
This note grew out of a question about whether recurring stylistic features—predictable phrasing, polished structure, or punctuation habits—could meaningfully distinguish machine-generated prose from human writing.
To explore it, I asked multiple research agents and their consumer-facing counterparts—including Kimi, Z.ai, ChatGPT Deep Research, Claude Deep Research, Qwen, and other systems—to examine the problem from different angles. This note brings those findings together and considers what they can actually support.
What the research suggests
The most important conclusion is that AI-generated writing cannot be reliably detected in the way I first imagined. A single signal may be interesting, and a cluster may justify a closer look, but neither is proof. Human writers can use the same patterns, AI systems can avoid them, and editing makes the boundary even less clear.
Why the framing matters
That conclusion changed the purpose of the note. Rather than presenting a checklist that claims to identify AI with certainty, it examines the patterns that can make writing feel AI-generated and the limits of using those patterns as evidence. The goal is a more honest understanding of what stylistic analysis can and cannot tell us.