Stylistic Contrastive Learning for Human-Like AI Text Generation

Michael Bronikowski

Keywords: Stylistic Contrastive Learning, Human-like Text Generation, Stylometry, Style Control for LLMs, Large Language Models

Abstract: AI-generated text is often fluent yet stylistically off: it leans formal, repeats safe phrasing, underuses idioms, and exhibits templated discourse, making it detectably non-human to both algorithms and attentive readers. We synthesize recent evidence quantifying these gaps—lexical diversity, syntactic variety, idiomaticity, and discourse planning—and propose Stylistic Contrastive Learning (SCL), a training framework that learns a human-style embedding and pushes generations toward it via a supervised contrastive objective. We instantiate SCL on GPT-5 and evaluate across essays, newsy expositions, and dialogues. SCL reduces stylometric detectability against a GPT-5 baseline by 18–22 points (e.g., from 72% → 54% on essays), increases distinct-n and idiom use, and raises human “sounds-human” ratings while preserving topical fidelity. Ablations identify idiom frequency and discourse markers as the strongest perceptual drivers. We discuss implications for evaluation, alignment, and detection.

Submission Number: 344