I Got Called Out for Sounding Like AI, So I Built a Gate for It
AI reply????
Photo by Mikhail Nilov on Pexels
On September 12, 2026 a reader answered one of my replies with a few question marks and "AI reply?", and they were right. The reply had already passed a rulebook I wrote the day before. What gave it away was a shape, not a phrase. Here is the reply, what was wrong with it, and how being called out twice turned a list of banned words into a gate.
What was the reply?
I'm building a tool to make replies sound human, and I was flagged as a bot. That was the first time a reader caught a draft before I did. The reply was:
Why did it pass my rules?
On September 11 I had written a rulebook for my own drafts: banned phrases, quotas and a self-audit step. This reply went straight through it, because none of the words were on a list.
What was actually wrong?
The shape. It opens by validating a concrete detail from their post ("$3000 a month is a real number"). Then it turns into a broad, reflective question the poster would never answer naturally ("what actually got lost in that trade").
The first half sounds sincere, and that makes the second half more convincing than it deserves to be. The tacked-on probing question is the tell, and "curious" is its usual signature. The fix is one move per reply: agree, disagree, add a detail or joke, never two of them.
Why wasn't a list of banned phrases enough?
Because each rule only covered the last mistake.
| Date | What happened | What changed |
|---|---|---|
| Sep 12 | The compliment-then-big-question shape got me flagged | It became a rule |
| Sep 22 | I caught myself echoing the other person's own words back (an internal catch, not public) | If all you can say is their own sentence, skip the post |
| Sep 25 | A thread of "AI comment tells" went around ("the scary part is...", "it's not X, it's Y", "this is gold") | I turned the tells into an automatic check, and a draft that trips it is discarded |
| Sep 28 | Words weren't the problem, shapes were | I added questions about the whole draft, like "does this read like a bot?" and "does it engage with the post?" |
Did it happen again?
Yes, on October 2. A second reply was publicly flagged as AI. It had passed every check, because it was built like a quotable one-liner, a tidy punchy contrast. So I added a check for something a person would not say: a slogan or aphorism dressed up as a reply.
What does a reply that passes look like now?
Short. One move. A small, specific detail. Sometimes a dry joke, often no question at all. For the reply above, here is how I'd write it today:
| Before | After |
|---|---|
| "$3000 a month is a real number to walk away from, curious what actually got lost in that trade besides the bill." | "that's a big number to walk away from, hope it was worth it" |
| Same post | "$3000 a month, i'd have a hard time walking away from that" |
For the longer list of habits behind this, see nine tells that make posts sound like AI.
What did I take from it?
- A draft that fails is thrown out and rewritten, never patched.
- If good drafts are thin, show fewer rather than padding.
- Every caught mistake becomes a permanent rule or check.
- I send every reply myself, and nothing is posted for me.
Is the problem solved?
No. This is a gate, not proof. Readers will find the next shape, and I would rather be told. Paste a post into the free checker and tell me what it misses. The story of why I started building this is in Why I built Humanish.
Try the checker ยท Start my free trial
FAQ
How can I tell if a reply was written by AI?
Look for a compliment followed by a broad question, a tidy slogan at the end, and words like "curious". Real replies usually make one small point and stop.
Why do AI replies get called out?
They tend to be too smooth: validating, certain and quotable. People notice the smoothness even when they can't name it.
Does a banned-word list fix it?
Not by itself. In my case the reply that got flagged contained no banned words.
Can this still happen to me with Humanish?
Yes. It lowers the odds, and it can't promise anyone will never suspect a reply.
Try it on your own account
Free for 7 days, no card needed. You send every reply yourself.