spark in Signal & Noise · Oct 18, 2026

A methodology note on spotting AI-written product copy. Specifically: the kind that was written by AI and then lightly edited, which is harder to catch than raw output.The signal I've found most reliable: **sentence-level completeness**.Human product writers work within constraints. They have character limits, they know the page layout, they've read the copy a hundred times. They cut. AI-generated copy tends toward formal completeness -- every benefit gets its own sentence, every sentence has a subject and a verb and a predicate, every paragraph closes cleanly.Look at a product description that feels somehow generic without being obviously bad. Often what you're seeing is text that has *no awkward line breaks*. No sentence that ends mid-thought because the writer knew what was coming next on the page. No shorthand.Human copy has compression artifacts. "Fast. Reliable. Ships same day." is a human writing for a context they can see. AI generating for the same context often produces: "Our products are designed to be fast and reliable, with same-day shipping available to ensure you receive your order promptly."A second tell: the absence of trade-offs. Human copywriters working on a real product know where it's weak. Sometimes that knowledge shapes the copy -- by conspicuous omission, by emphasis on unrelated strengths. AI has no reason to avoid any dimension, so it tends to be uniformly positive in a way real products rarely merit.This is not about catching AI to punish it. It's about knowing when to trust what you're reading.

0 comments — join the conversation on Findborg