How to Humanize AI Text: What We Learned from 19,804 Text Pairs

2026-06-22 · Source: Naturallanguageprocessing on Medium · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Data Science & Analytics · Depth: Intermediate, medium

Summary

ReText.AI conducted an internal study across 19,804 paired texts, evaluating how a specialized AI humanizer alters detection scores from a single AI detector. The research involved eight language models, including Llama-3.2–3B-Instruct and GPT-oss-120B, and covered 20 topic categories, with 80% of the dataset in Russian. The humanizer, built on Gemma-2–9B-IT and fine-tuned with SimPO, aims to rewrite text by adjusting sentence structure, rhythm, and vocabulary to sound more natural. Results showed a significant shift in AI-likelihood scores, with the median dropping from approximately 0.92 to 0.45 after humanization. More than 90% of texts received a lower score, and 14 of 20 categories achieved a classification flip rate above 50%, notably in recipes (66.7%) and legal texts (64.2%). The study acknowledges limitations, including its internal nature, the use of the same detector for training and evaluation, and the dataset's language skew.

Key takeaway

For content creators and editors aiming to produce authentic-sounding text, you should integrate AI humanizers as an editing tool, not a shortcut. Start by outlining your arguments, then use AI for a first draft. After humanization, meticulously verify semantics, data, and references, as the process may alter accuracy. Finally, cross-check with AI detectors for insights, but always complete the text by hand with personal insights and specific details to ensure genuine authorship and quality.

Key insights

Targeted AI text rewriting can substantially lower detector scores, but its effectiveness varies by topic and measurement.

Principles

• Humanization rewrites sentence structure, rhythm, and vocabulary.
• Lower detector scores do not prove human authorship.
• Structured content humanizes more effectively.

Method

A Gemma-2–9B-IT model, fine-tuned with SimPO, generated rewritten variants optimized to reduce AI-likelihood probability based on detector confidence.

In practice

• Generate AI first drafts, then polish.
• Verify factual accuracy post-humanization.
• Use AI detectors for cross-referencing drafts.

Topics

• AI Text Humanization
• AI Content Detection
• Language Model Evaluation
• SimPO Fine-tuning
• Text Rewriting
• Multilingual Datasets

Best for: Research Scientist, AI Scientist, NLP Engineer, Marketing Professional

Related on AIssential

• A 1B humanizer that matches human writing on an AI detector — A 1B-parameter humanizer can generate AI text that evades current AI detectors, matching human writing.
• Operation-Guided Progressive Human-to-AI Text Transformation Benchmark for Multi-Granularity AI-Text Detection — AI-text detectability is non-monotonic, influenced by edit operations and revision history, not just AI content proportion.
• How to Humanize Your AI Writing in 10 Steps — Humanizing AI writing requires addressing both AI's common errors and incorporating distinct human writing techniques.
• ‘Being human helps’: despite rise of AI is there still hope for Europe’s translators? — AI is rapidly transforming the translation industry, improving in capability but still lacking human nuance and contextual understanding.
• AI is teaching us to speak like bots, and it’s a problem — AI chatbots are subtly homogenizing human language, making our communication increasingly machine-like by nudging expression towards an average.
• SenFlow: Inter-Sentence Flow Modeling for AI-Generated Text Detection in Hybrid Documents — Modeling inter-sentence dependencies via structured prediction significantly enhances AI-generated text detection in hybrid documents.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Naturallanguageprocessing on Medium.