Table of Contents

AI is weirdly good at writing.

You paste a messy thought into a model and it comes back clean. Organized. Confident. The kind of thing you could drop into a doc, a Slack update or a LinkedIn post.

Then you reread it and feel that tingling problem.

It still reads fine. But it doesn’t feel anything like you.

It feels like something that could have been written by anyone.

That’s the ‘AI smell’. Not bad writing. It is unlived, hollow writing.

If you want the automated version of what I am about to explain in this post, I built De-AI that helps you recognize when your writing looks AI-generated in the How to fix it section at the bottom of this post. But, I would recommend you to read this post first - to understand why do these fixes that I recommend work.

What ‘AI smell’ usually looks like

1. AI skips what happened and jumps straight to what it means

AI loves lines like:

“This highlights the importance of clear communication.”

That sentence isn’t wrong. It’s just floating.

A human version usually has gravity to it because it has an event:

“We shipped the incomplete feature because no decision was agreed upon in writing and many assumptions were made by those involved in the project.”

The same lesson feels different, doesn’t it? One is a moral. The other is a possibly relatable moment.

Do this: If you used ‘highlights or reflects or demonstrates’, rewrite it as what happened instead. If you can’t name the event, add a placeholder and come back:

  • [Who] decided [what] on [when], and it caused [consequence].

2. AI uses impressive words to hide the lack of specifics

AI is great at professional-sounding fog:

“This initiative leveraged cross-functional alignment to optimize workflows.”

It sounds confident. It also says almost nothing. That’s the fog. It sounds like me at an exam when I did not prepare well.

A human version sounds like this:

“We saw support tickets 2x right after the launch, so we listened to five customer calls every Friday and changed onboarding the next week. This activity done over 4 weeks finally got the support ticket count down to what it was pre-launch.”

A quick test is to ask ‘How do you know?’

If your sentence cannot answer that question, it’s best to drop it.

One clean way to add proof is to insert just one of these:

  • a number: [metric], [delta], [time]

  • a constraint: “We only had [X] days”

  • a tradeoff: “We chose [A] and gave up [B]”

  • a before/after: “Before we did [X], [Y]. After, [Z].”

  • a mistake: “We had to undo [thing] because [reason]”

You do not need all of them. One is enough.

3. AI relies on sentence templates people now associate with AI

Some structures are fine. The problem is they’ve been spammed by AI so much that readers notice them faster now.

Examples:

  • Not only X, but also Y

  • If X, then Y

  • This enables teams to…

  • There are three key reasons…

When readers see these, they stop listening for your idea and immediately label it as AI slop.

A simple swap that keeps your meaning: report what happened instead of predicting a principle.

Instead of: “If teams document decisions, productivity increases.”

Write: “When we documented decisions, rework dropped. When we didn’t, it spiked.”

Now it sounds like someone who has watched the pattern play out, not someone giving advice from a podium.

4. AI over-explains everything like a tutorial

AI loves explaining:

“To understand why this matters, let’s break it down. First… Second… Finally…”

It reads like a course module.

Sometimes structure is useful, most of the times it is not needed. Your writing does not need to be some course module.

Humans do something simpler than AI when it comes to communicating - writing and speaking.

They say the point. Then they say the reason. Then they move on.

If your draft feels too polished, the fastest fix is usually to delete the coaching lines and keep the text the actual way you will say it.

5. AI uses em dashes too often

This one was an unnecessary, because em dashes are fun. They have been part of human writing for long - back when no one would even notice it.

The issue is the overuse. Models use them constantly and that creates a signature cadence that reads ‘machine-generated’ sometimes even without being one.

If you want an easy rule, stop using em dashes. If a dash can be a period, make it a period. You want to win the war, not a battle.

The real tell

AI writing often sounds like someone who knows the rules but little about actually writing from scratch.

Human writing sounds like someone who paid a price for the lesson having actually lived the experience.

That price can be subtle. A confusing meeting. A wrong assumption. A messy handoff. A tradeoff you still disagree with.

That’s where voice comes from.

Here’s a gut-check line that helps: What would I add here if I had to prove I lived this?

If the answer is ‘nothing’, you’ve found the smell.

How to fix it

1. The human pass

If you do one pass before you hit send, do this. Five moves. Ten minutes.

  1. Add one real thing that happened
    A decision. A mistake. A consequence. Even one sentence changes everything.
    Template: “[I/we/they] did [action], and [consequence] happened.”

  2. Remove one ‘fog’ sentence
    Find the line that sounds the most “professional.” The one you would not say out loud.
    Delete it or replace it with one concrete detail: a number, a constraint, or an action.

  3. Convert one template sentence into an outcome
    Trade prediction for observation.
    Swap “If X, then Y” with “When we did X, Y happened.”

  4. Delete one tutorial sentence
    Cut the “I will now explain” lines. Keep the actual explanation.

  5. Remove em dashes
    If they’re decorative, replace them with periods or rewrite the sentence.

Just these changes are enough to even make completely AI generated text to sounds so much more like you, clearly saying that you know the details, you lived the experience.

2. I made a Custom GPT that does this automatically

I got tired of running this checklist by hand, so I built a CustomGPT that:

  • flags the exact “smell points” (a lower score out of 10 indicates the content looks more like AI generated)

  • suggests specific point-by-point replacements (with placeholders instead of made-up facts)

  • helps you rewrite without turning your voice into jargon-y mush

You can access it here:

Reply

Avatar

or to participate

Keep Reading