The patterns that give AI writing away, and the stances that prevent them.

Em-dashes deployed as a universal connective tissue — between clauses, inside lists, instead of commas, colons, and parentheses.

Setting up a small claim only to dramatically upgrade it in the same sentence. Manufacturing stakes the reader didn't ask about.

Every claim wrapped in softeners until nothing has actually been asserted. Uncalibrated humility as a uniform safety wrapper.

Reflexively presenting 'both sides' on questions that have a clear answer, manufacturing controversy where none exists.

Opens with content-free throat-clearing. Restates the topic in inflated language and conveys zero new information.

Reflexively listing three near-synonyms where one word would do. Form imitating substance.

Turning continuous reasoning into disconnected bullets. Visible structure substituted for actual structure.

Generic compliments to the user before the answer. Affect divorced from any actual evaluation.

Restates what was just said, or offers generic continuation invitations. Bookends without a body that needs them.

A small word cloud of inflated vocabulary that signals 'AI wrote this' on its own. Costume rather than precision.

Bolding random nouns and verbs throughout body text. Emphasis applied uniformly conveys nothing.

Caveats and 'consult a professional' framings on benign requests. Risk management without risk evaluation.

Posing questions the reader didn't ask, just to answer them. Borrowed dialogue used to pace the writer's own thinking.

Constructions that gesture at multiple causes without naming a single one of the others. Nuance-signalling without nuance.

Confidently stating frequencies the model has no way to know. Substituting the prior on a phrase for the prior on the thing.

Proposing a locally-correct fix without tracing the call path. Pattern-matching the shape of a solution and skipping the data-flow audit.

Performing certainty on guesses. Fabricating specific names, numbers, signatures, or citations to make fluency look like knowledge.

Wrapping the answer before the work is done: 'bottom line up front' or a final-sounding sentence used as an escape hatch from actually reasoning through it.

Surfacing a problem and moving on. Naming an issue in the response without resolving it when resolution was the job.

Patching at the wrong layer. Fixing a data-shape bug in CSS, swallowing a logic error in a try/catch, regex-fixing a parser problem.

Sensory or temperamental adverbs ("quietly", "softly", "gently") attached to subjects with no sound, body, or intention. Mood borrowed and pasted onto things that can't hold it.

Patting the user's artifact on the head with craft vocabulary. "that's doing real work", "earning its keep", "load-bearing". Appraisal that adds no information about why.