People can identify AI-assisted writing, and their objection is rarely that it is wrong. It is that it is weightless — fluent, balanced, agreeable, and saying nothing that could be disagreed with.

The cause is mostly not the model. It is how the model is being asked.

The tells

Symmetry. Every point balanced by a counterpoint, every claim softened. Real writing has a position and defends it.

Enumeration for its own sake. Three reasons, four benefits, five considerations — regardless of whether the subject actually decomposes that way.

The abstract noun stack. "Leveraging innovative solutions to drive operational excellence." Words with no referent.

Universal applicability. Copy that could describe any company in the category. If a competitor could put their logo on it unchanged, it says nothing.

The unearned summary. A closing paragraph restating what was just said, adding nothing.

Cheerfulness. An even, mildly positive tone regardless of subject. Real writing about a hard subject sounds like it.

Why these appear

Models are trained toward helpfulness and inoffensiveness, which pushes toward the median of everything written on a topic. The median is balanced, structured and unremarkable — by definition.

Asking for "a blog post about X" gets exactly that: the average of everything written about X. The output is not a failure. It is the correct answer to a question nobody should have asked.

What to do instead

Supply the position. Do not ask what to say. Decide what you think, then ask for help saying it. "Argue that monthly competitive reports are structurally useless because of latency, and address the objection that continuous monitoring creates noise" produces something. "Write about competitive intelligence" produces the median.

Supply the specifics. Real numbers, real constraints, real situations. A model cannot invent the detail that makes writing credible, and if asked to, it will produce plausible fabrications — the worst possible outcome.

Ban the summary paragraph. Explicitly. End on the last real point.

Demand asymmetry. Ask for the strongest version of the argument, not a balanced overview. Add the counterargument yourself, where you decide it belongs.

Cut 20% after. Almost always improves it, because models over-explain. The first sentence of most paragraphs is frequently throat-clearing.

Read it aloud. The AI register is much more obvious in the ear than on the page. Anything you would not say to a person, cut.

What models are genuinely good at

Structure, when you supply the argument. Variation, when you supply the original. Consistency of terminology. Compression. Adapting an approved piece across formats and lengths. Catching the sentence that contradicts an earlier one.

That is a substantial amount of the work of writing. It is simply not the part where the writing acquires a point of view.

The division that holds

The human decides what is true, what matters, and what to claim. The machine helps with everything downstream of those decisions.

Writing that sounds like nobody wrote it is usually writing where nobody made those three decisions before generation started.