I "detect" them through overuse of some patterns, like "It's not X. It's Y."

This is an artifact of the default LLM writing style, cross-poisoned through training on outputs -- not an "universal" property.