Read On

One paragraph was written by a person before language models. Two were written by a model. Can you tell which, and can a detector?

Made with pollinations.ai

    Check your own text

    Write a new round with your Pollen

    Detector scores weighted habits per 1000 words, recorded paragraphs

    After the slop score in "Qwen3-4B with the AI writing tics trained out": human writing there usually scores under 15 and the original model 121. A regex can't tell good writing from bad, which is why the reader plays too, and why damage (length, ", which" joins, invented studies) is shown beside the score.

    How it works

    The person

    Ten explainer paragraphs written by US federal agency staff (NOAA, USGS, NPS, NIH), public domain, each with an archived copy from before 2019.

    The model

    Each model gets the same request for every topic: one paragraph for general readers, about as long as the person's. It answers twice in separate calls: once as it is, once told to write plainly. Long replies are cut at the last full sentence, and the card says so.

    The two readers

    You pick the person, blind. The detector picks the lowest slop score. Martin Fowler wrote in LLM Voice that he distrusts both detector tools and his own ear. Here both get measured on the same paragraphs.

    The two requests a model gets
    What the detector counts