The common tells
Relentless balance. Every point gets a counterpoint. Every claim gets hedged. The result reads as thorough but says very little.
Empty scaffolding. Phrases like "in today's fast-paced world," "it's important to note that," and "let's dive in" that add words without adding meaning.
Uniform rhythm. Sentences of similar length, paragraphs of similar size, an evenness that human writing rarely has.
Summary-of-a-summary vagueness. Text that describes a topic accurately without ever getting specific enough to be useful or memorable.
Over-signposting. "First, we'll examine X. Then we'll explore Y. Finally, we'll conclude with Z" — announcing structure instead of just having it.
How to avoid them
Give it your actual details. Generic input produces generic output. Real specifics — a real number, a real example, a real constraint — force specificity into the response.
Ask for a point of view. "Argue for X" produces sharper writing than "discuss X." You can soften a strong draft; you can't sharpen a mush.
Name the tells you want avoided. Explicitly: "no phrases like 'in today's world', no hedging every claim, vary sentence length."
Give it a sample of your writing. Paste something you wrote and ask it to match that voice. This works better than describing your voice in adjectives.
Cut ruthlessly afterwards. AI drafts are almost always too long. Cutting 30% usually improves them, and the cuts are usually the padding phrases.
The most important habit
Rewrite at least some of it yourself. Not because AI writing is inherently bad, but because the parts you rewrite are where your actual voice and thinking enter. A draft you've meaningfully edited reads differently from one you've merely approved.
A note on honesty
If you're publishing AI-assisted writing in a context where the source matters — academic work, journalism, anything with disclosure norms — the ethical answer is to follow those norms rather than to get better at hiding the tells. This article is about writing well, not about deception.
Practical starting point
Our prompt for turning rough notes into a draft is built around keeping your original meaning and details rather than replacing them — which is the core of writing that still sounds like you.
The specific tells
People say AI writing is recognizable but rarely say what they are recognizing. These are the concrete patterns:
Openings that circle before landing. "In today's fast-paced digital landscape, it is more important than ever to..." Two sentences that could precede any topic. Human writers usually start closer to the point because they are impatient to make it.
Relentless balance. Every claim immediately softened by its counterpoint, every paragraph ending in an on-the-other-hand. Real writing commits somewhere.
Tricolon everywhere. Three-item lists in a rhythm, repeatedly, regardless of whether the subject has three parts.
Nothing specific. No names, no numbers, no particular Tuesday when the thing went wrong. Specificity is the fastest way to sound like a person, because only a person was there.
A conclusion restating the introduction. It signals nothing was learned in between.
An editing pass that works
Take an AI draft and do this in order. It takes about five minutes.
- Delete the first paragraph. Then read it. About seven times in ten the piece is better, because the first paragraph was throat-clearing.
- Add one thing only you know. A number from your own work, a mistake you made, a client who wanted something odd. One concrete detail does more than a page of adjectives.
- Cut every hedge you do not mean. "Arguably", "it could be said that", "many experts believe". If you believe it, say it.
- Break the rhythm. If three paragraphs run the same length with the same structure, merge two or split one. Uniformity reads as machine.
- Read one paragraph aloud. Anything you would not say out loud gets rewritten.
Where not to use it at all
Some writing should not be AI-assisted, and the line is not about quality. Condolences, apologies for a real mistake, performance feedback, a resignation, a message to someone going through something hard — in all of these the effort itself is part of the message. A perfectly worded condolence that took eight seconds to produce is worse than a clumsy one you wrote yourself, and if the recipient ever suspects, the damage is not recoverable.
The Using AI in Everyday Work course covers this line in more detail, and our rough notes to first draft prompt is built specifically to keep your own material at the center of the draft.