Writing with AI without sounding like AI

You can usually tell when something was written by AI. There is a smell to it. It is not bad writing — often it is technically correct — but it has a sameness, a smoothness, a kind of polite muzak quality that human writing does not have.

Most people can sense the smell but cannot name it. That is the problem with fixing it. This article names the smell, gives you six specific fixes, and walks through a workflow that produces AI-assisted writing that does not sound like AI wrote it.

What "AI writing" actually smells like

The six tells, in roughly the order people notice them:

1. Filler openings. "In today's fast-paced world," "It's no secret that," "Have you ever wondered," "In an era of...". The model has learned that these are common, polite, and safe ways to open. They are also the first three sentences of every blog post you have ever skipped.

2. The "I hope this finds you well" reflex. Even outside email, AI defaults to warm, generic preambles. "Great question!" "I'd be happy to help!" "It's important to note that..." Every one of these is a sentence that adds words and removes signal.

3. The em-dash plague. AI loves em-dashes. They are used to perfectly balance sentences in a way humans rarely do. If every paragraph has two em-dashes splitting it into three balanced clauses — like this — you are probably reading AI. (We use em-dashes here where they are natural; the tell is the cadence of two perfectly placed dashes per paragraph, not the dashes themselves.)

4. Perfect-rhythm sentences. AI-generated prose tends to have suspiciously even sentence lengths and consistent rhythm. Human writing has variation — a short punch followed by a longer wandering one, then back. AI rhythm sounds slightly metronomic.

5. The hedged conclusion. "While there are many factors to consider, it's important to weigh all the options carefully and make the decision that's right for you." The non-position. AI hedges to avoid being wrong, and the hedge is the most boring sentence in the article.

6. Vocabulary that signals nothing. "Leverage," "utilize," "synergy," "robust," "delve into," "navigate the complexities of," "in today's rapidly evolving landscape." These words exist in real writing too — but they are AI's default vocabulary, and overusing them is the surest sign.

Notice that none of these are wrong. They are just generic. AI writing fails by being safe, not by being incorrect.

Six fixes

A small set of moves that strip the AI smell out of any draft.

Fix 1: Strip the opening

The first paragraph of an AI draft is usually skippable. Cut it. If the rest of the piece doesn't make sense without it, cut the second paragraph too. Real writing starts at "here's the thing"; AI writing starts at "here's why I'm about to tell you the thing."

A useful prompt to do this in one pass:

Strip the opening of this draft so it starts at the first concrete idea. Remove anything that previews what's coming or sets up the topic generically. Keep everything else.

Fix 2: Replace generic transitions with specific ones

"Furthermore," "additionally," "moreover," "in conclusion." These signal the existence of a paragraph break without doing the work of one. Replace each one with either nothing (the paragraph break alone is enough) or a specific reference to the previous idea ("The same logic applies to..." "There's a related problem...").

Fix 3: Vary the sentence rhythm

After AI drafts something, read it aloud. If the rhythm feels metronomic, break it up. One short, sharp sentence. Then a longer, more wandering one that explores an idea you were not sure how to land at first. Then a short one. Done.

You can ask AI to do this for you, but it is often faster to do it yourself in five minutes. Just vary lengths deliberately.

Fix 4: Cut the hedges

Words and phrases to delete on sight:

• "It's important to note that"
• "It's worth considering"
• "One might argue"
• "It could be said"
• "While there are many factors"
• "Ultimately"
• "In conclusion"
• "At the end of the day"
• "In essence"

Each of these is the writer (or the AI) buying time before saying the actual thing. Delete them and the sentence almost always becomes stronger.

Fix 5: Replace the AI vocabulary

A targeted prompt that helps:

Edit this draft for word choice. Replace every instance of the following words with something more specific: leverage, utilize, robust, navigate, in today's, going forward, synergy, ecosystem, journey, holistic, paradigm. Do not change the meaning. If a word genuinely belongs, keep it.

You can also keep your own personal list of words you never want in your writing and add it to your custom instructions.

Fix 6: Add specificity

Generic AI sentences turn into good human sentences when you replace abstract claims with concrete ones. Numbers, names, examples, specific situations.

"Many companies face challenges with adoption."

versus

"Three of the five enterprise pilots we ran in Q2 stalled at the same point: the security review."

Same fact. The second sentence is human. The first is AI. The difference is one specific detail.

A useful prompt to push toward this:

Find every abstract or generic sentence in this draft. For each, either delete it or replace it with a concrete, specific version — with a number, a name, an example, or a particular situation. Do not invent facts; flag anywhere you'd need me to fill in a specific.

The workflow that doesn't produce AI smell

The structural fix — the one that prevents AI smell rather than removing it after — is to keep your voice at the centre of the draft. The pattern that works:

1. You think first. Before opening AI, jot down 5–10 bullet points of what you actually want to say. Your real take, in your real words. Even if it is messy.
2. You give the model your raw thinking. Paste in your bullets and say "expand these into a draft. Use my words and my framings — do not add new ideas or smooth out my edges. Keep specific examples specific."
3. AI drafts based on you, not from scratch. The result already has your voice baked in.
4. You edit aggressively. Cut what does not sound like you. Rewrite the parts that sound like AI by hand. Run the six fixes.
5. You read aloud. The final test. If a sentence makes you wince when you say it, fix it.

Compare this to the AI-first workflow: "ChatGPT, write me a 500-word post about X." That produces AI-smell-prose by definition. The model has nothing of you to work with, so it defaults to its average voice — which is everyone's average voice.

A specific pattern: the "speak it first" trick

If you struggle to write your bullets, use voice mode. Open ChatGPT, switch to voice, and just talk through what you want to say for two or three minutes. Then ask the model: "Take the transcript of what I just said. Turn it into a polished version of the same content, in my voice. Use my actual phrases. Tighten the rambling, but do not smooth out my edges or add things I did not say."

You will be surprised at how good the result is. The model is no longer writing for you — it is editing what you said. The voice is already yours.

A note on writing where AI smell is invisible

Some kinds of writing do not show AI smell strongly even when AI does most of the work:

• Boilerplate. Standard "thanks for booking with us" confirmations, polite cancellation notices, generic FAQ answers. These are supposed to be generic.
• Translation. A clean AI translation is fine if the source already had voice. The model is preserving voice, not creating it.
• Summary and extraction. Pulling facts out of a long document. Voice does not matter; accuracy does.
• Highly templated formats. Bug reports, invoice descriptions, status updates with a fixed structure. The structure is the voice.

For these, the AI workflow is fine as-is. Skip the voice-preservation moves; just optimise for accuracy and structure.

For everything else — anything that is meant to be read by someone who knows you, anything that represents you or your company's voice, anything where the who matters — keep your voice in the loop.

Two prompts to keep around

The two prompts I reach for most often when finishing an AI-assisted draft:

Edit this for AI smell. Find anything that sounds generic, hedged, or like every other AI-written piece. Quote each line you want to change and suggest one or two alternatives. Be ruthless. Cut, do not just polish.

Read this draft as if you were a skeptical reader who can spot AI writing. Where would you stop reading? Where does the rhythm feel artificial? Where does the vocabulary feel off?

Saved as snippets, these are the difference between sending a draft and sending a clean one.

The takeaway

AI smell is not a mystery. It is six specific tells that come from AI defaulting to safe, smooth prose. You can strip them with a small set of edits, or — better — prevent them by keeping your voice central from the first bullet.

Six fixes. One workflow. Two cleanup prompts. Use AI to draft what you have already thought, not to think for you, and your writing will keep getting better instead of slowly converging on everyone else's.