Honest Numbers: What Our Blind Tests Found, and What Is Still Weak

In blind tests, judges scored Humanish drafts between 3.2 and 3.5 out of 5 for sounding human, and the share of drafts with invented details fell from 27% to 14%. That is good progress and not perfection. Jokes and sarcasm remain our weakest area. Here is how we tested and what we are still fixing.

How did we test it?

We wrote drafts across eight posting styles (builder, coach, analyst, commentator, creator, craft, local business and professional) and six tones (casual, formal, funny, sarcastic, direct and warm), for both replies and original posts. We used invented source posts, never real people's. Separate judges then scored each draft without knowing which settings produced it.

What were the results?

Invented details are the failure we care about most: a made-up client, number or story. Halving them is not the same as removing them, which is why you read every draft before it goes out.

What is still weak?

Humor is the clearest gap. Funny, sarcastic and warm drafts landed less often than casual and direct ones, so treat those three as a starting point you will edit, not a finished joke.

What do we not claim?

• That drafts are undetectable.
• That any tone works every time.
• That growth is guaranteed.

Why publish the weak spots?

Because a tool that only shows its best numbers is hard to trust. You are going to post these under your own name, so you should know where they stumble. The story behind the quality check is in I got called out for sounding like AI, and the habits it looks for are in nine tells.

Test your own post · See plans

FAQ

Are the test posts real people's posts?

No. They were invented for the test.

What does 3.5 out of 5 mean in practice?

It reads human most of the time, with some drafts that still need an edit.

Can I change the tone?

Yes. Tones and voice settings let you steer a draft, and your edits teach it.

Does Humanish ever post for me?

No. You read, edit and send every reply yourself.