CR4T: Rewrite-Based Guardrails for Adolescent LLM Safety

TL;DR

This paper introduces CR4T, a rewrite-based framework that enhances adolescent LLM safety by transforming unsafe outputs into age-appropriate guidance, reducing refusals and improving conversational quality.

Contribution

CR4T offers a novel, model-agnostic approach that reconstructs unsafe responses into developmentally suitable guidance, addressing limitations of traditional refusal-based safety methods.

Findings

CR4T significantly reduces unsafe outputs and refusals.

Targeted rewriting preserves benign interactions and improves user experience.

Experimental results validate the effectiveness of CR4T in adolescent LLM safety.

Abstract

Large language models (LLMs) are increasingly embedded in adolescent digital environments, mediating information seeking, advice, and emotionally sensitive interactions. Yet existing safety mechanisms remain largely grounded in adult-centric norms and operationalize safety through refusal-oriented suppression. While such approaches may reduce immediate policy violations, they can also create conversational dead-ends, limit constructive guidance, and fail to address the developmental vulnerabilities inherent in adolescent-AI interactions. We argue that adolescent LLM safety should be framed not solely as a filtering problem, but as a socio-technical, developmentally aligned transformation problem. To operationalize this perspective, we propose Critique-and-Revise-for-Teenagers (CR4T), a model-agnostic safeguarding framework that selectively reconstructs unsafe or refusal-style outputs into ageappropriate, guidance-oriented responses while preserving benign intent. CR4T combines lightweight risk detection with domain-conditioned rewriting to remove risk-amplifying content, reduce unnecessary conversational shutdown, and introduce developmentally appropriate guidance. Experimental results show that targeted rewriting substantially reduces unsafe and refusal-oriented outcomes while avoiding unnecessary intervention on acceptable interactions. These findings suggest that selective response reconstruction offers a more human-centered alternative to refusal-centric guardrails for adolescent-facing LLM systems.