You write 8 text completions and open models score how predictable each word was too them. Predictable => clanker. You can share results with your friends.

The scoring checks every word you write against the model's logprobs. Right now I'm using Llama3.1, Deepseek v3 and Qwen3 to keep costs low. I tried to calibrate it so other models (chatgpt/claude) score 100% and interesting human responses score in the 10-30% range.

Totally free, no signup

The metric used is per-word surprisal: -logprob of each word you type. This is just the same thing as per-word cross entropy or KL-divergence where the user distribution is one-hot. Calibrating it so text generated by frontier models scored poorly was a challenge at first. Originally ChatGPT was scoring around 54%. I'm still having trouble assigning high scores to the personalized Gemini and ChatGPT responses when I'm logged in because all my personal context gives surprising responses.

And yes, gibberish responses score very human :)

Funny little game, would be even funnier to have a system to roast the prose of a friend on social media or even a screenshot