← Back to feed
2026-05-25data

Retrieval-Augmented Detection of Potentially Abusive Clauses in Chilean Terms of Service

Christoffer Loeffler, Tomás Rey Pizarro, Daniel Ignacio Miranda Vásquez, Andrea Martínez Freile

PDF preview unavailable
Read on arXiv →

Key claim

Automated detection of abusive clauses improves legal contract review.

This paper presents a framework for automatically detecting abusive clauses in Chilean Terms of Service, leveraging retrieval-augmented generation techniques. A key result shows that this approach allows local models to perform comparably to larger cloud-based systems while being more cost-effective.

In plain English

This paper presents a framework for automatically detecting abusive clauses in Chilean Terms of Service, leveraging retrieval-augmented generation techniques. A key result shows that this approach allows local models to perform comparably to larger cloud-based systems while being more cost-effective.

Novelty
7.5/10

The paper introduces a novel retrieval-augmented generation framework tailored for legal text analysis, which is a significant extension of existing methods.

Reliability
8.0/10

The study provides solid experimental validation with comparisons to both commercial and open-weight models, supporting its claims effectively.

Deep reliability assessment

The methodology supports the claim that retrieval-augmented generation (RAG) can improve the classification of potentially abusive clauses in Chilean Terms of Service, but the generalizability to other legal contexts or languages is not fully addressed.

Reproducibility

No open source code or dataset is mentioned, making reproducibility challenging.

Discussion questions

  1. 1.How well does the retrieval-augmented generation approach generalize to other legal systems or languages?
  2. 2.What are the practical implications for developers looking to implement this system in real-world applications?
  3. 3.What evidence would be needed to falsify the claim that RAG significantly improves classification performance over few-shot learning?

Key figure

Figure 1 illustrates the processing pipeline for detecting and classifying potentially abusive clauses, highlighting the user interface, detection, and classification modules.

Benchmark results

Chilean Abusive ToS ExtendedMacro-F1: 0.7316vs GPT 5.2 (Dense)+0.001SOTA
Retrieval-Augmented Detection of Potentially Abusive Clauses in Chilean Terms of Service — Frontier Papers