DEV Community

Scrap Labs
Scrap Labs

Posted on

Free AI-text detectors failed our paraphrase test. Here's the eval we use now

Before writing a single guide at Pastagi, we ran a humbling benchmark: six free AI-text detectors against progressively paraphrased model output.

The setup

Take 100 GPT-generated paragraphs. Pass each through a light paraphrase (synonym swaps, sentence reordering). Ask the detectors to classify both versions. No fine-tuning, no tricks - the same text with its surface scrambled.

The results that stung

Detectors that scored 90%+ on raw model output dropped 20-40 points after paraphrasing. One popular tool flipped to calling paraphrased human text "AI" more often than actual paraphrased AI text. The lesson isn't "detectors are bad" - it's that detector scores are meaningless without knowing the surface-form robustness of the specific detector.

What we do differently now

Every generation feature we ship gets an eval harness with three suites: raw output, paraphrased output, and adversarial near-duplicates. Regression means the harness fails, not vibes. And RAG answers must carry citations or they don't ship at all - attribution does more for trust than any detector.

The full benchmark methodology is on the site. Reproduce it before you buy any detector's enterprise tier.

title:
published: false
description:
tags:

cover_image: https://direct_url_to_image.jpg

Use a ratio of 100:42 for best results.

published_at: 2026-09-04 20:00 +0000


Top comments (0)