This Free AI Chatbot Was Trained to Talk Like a Real Person

A developer who goes by kvyb got annoyed at something many of us have felt: talking to an AI that always sounds like an AI. Too helpful, too wordy, always starting with "Certainly!" and ending with an offer to help further. So he did the unreasonable thing — he collected 125,217 real human-to-human chat messages from 1,396 conversations, stripped out identifying details, and used them to retrain an existing open-source model. The result, called Qwen3.8-27B-Humanlike-Chat, is meant to sound like a person texting, not a customer service desk.

In practice, that means noticeably shorter replies, less polish, and fewer forced attempts to keep the conversation alive. The developer says it feels less like an assistant even without giving it any special instructions. If you use AI for emails, brainstorming, or just chatting, this hints at where things are heading: people don't always want a helpful robot, they want something that doesn't sound like one.

There are real catches. An earlier version of this model scored five percentage points lower on a standard test that measures how well AI follows instructions — the developer hasn't rerun that test on the released version, so the exact tradeoff is unknown. Coding performance was never tested at all. The base model it's built on has had its safety refusals stripped out, which means it may answer things a normal chatbot would decline. And this is one person's hobby project, not a company product with quality control.

You can try it free through a demo page on Hugging Face, or through a free but rate-limited API endpoint. The bigger story: the gap between "AI that sounds like AI" and "AI that sounds like you" is closing fast — and it's being closed by hobbyists, not just big labs.