How to reduce AI detection score in n8n content generation workflow?

Help me Build my Workflow
1

## My Problem
I’m building a content generation workflow in n8n that creates blog articles.
Currently getting 60-95% AI detection scores. Need to reduce to under 20%
using only FREE tools (no budget for paid humanizers).

My Current Setup

  • Groq (8 chat models)
  • Google Gemini 2.0 Flash
  • JavaScript code nodes for cleanup
  • HuggingFace (tried, unreliable)

Workflow Structure

  1. Article draft (Groq llama-3.3-70b)
  2. Humanizer Pass 1 (Gemini)
  3. JS chaos injector
  4. Final cleanup (JS)
  5. Send to Telegram

What I’ve Tried

  • Custom humanizer prompts (academic + casual tone)
  • Word replacement (utilize->use, etc.)
  • Sentence burstiness
  • Multi-model rewriting
  • Adding hedge words

Current Results

What I Need

Free workflow techniques or templates that consistently achieve under 20% AI score
while maintaining professional quality.

Questions

  1. Has anyone achieved low AI detection with free tools?
  2. Are there specific prompt techniques that work?
  3. Any free APIs I’m missing?
  4. Should I add specific node types?

Thanks in advance for any help! :folded_hands:

2

Hey sarathkumar,

interesting workflow setup. I’ve dealt with this topic before and came to an uncomfortable conclusion.

The real problem isn’t in your workflow, but in the tools you’re using to test it. I fed the same completely AI-generated text into 5 different detectors. Result: 33% to 100% on the identical text. ZeroGPT, which you’re using, was the most unreliable in the test. So you’re optimizing for a metric that isn’t reproducible itself.

What actually makes the difference: feeding your own data, concrete examples, and personal experience directly into the prompt, not tinkering in post-processing. A model that writes based on real observations produces fewer recognizable patterns from the start. Prompt architecture is more effective than any humanizer pass afterward.

Regarding your setup: The Groq+Gemini approach with multiple passes is fundamentally solid. I’d drop the JS Chaos Injector—it makes the text worse without reliably lowering the score.

I’ve done a test with concrete numbers on this, if you’re interested, happy to share via DM.