Algospeak, Hiding in the Open: The Trade-off Between Legible Meaning and Detection Avoidance

Problem
This preprint addresses the gap in understanding the dynamics between linguistic evasion strategies, termed Algospeak, and their detectability by language models (LLMs). As LLMs are increasingly employed for content moderation, the coevolution of evaders and detectors necessitates a formalized understanding of how Algospeak affects both comprehensibility and detection rates. The authors introduce the concept of Majority Understandable Modulation (MUM) to quantify the trade-off between evasion and understandability, which has not been systematically explored in existing literature.

Method
The authors propose a reproducible framework for generating Algospeak variants that maintain meaning while varying modulation levels. This framework is based on a taxonomy of evasion strategies and allows for tunable modulation across five levels. The empirical evaluation utilizes a dataset of 700 modulated items derived from twenty base sentences, employing seven distinct evasion strategies. Two linked evaluations are conducted using seven different LLMs: one focuses on meaning recovery to assess interpretability, while the other evaluates disinformation detection through classification tasks. The study employs curve fitting to estimate the MUM threshold, facilitating sensitivity analyses across modulation strategies and models.

Results
The findings reveal a characteristic inverse relationship between understandability and modulation levels. Specifically, the MUM threshold is identified, indicating the point at which further modulation enhances evasion but significantly reduces comprehensibility for the majority of recipients. The results demonstrate that as modulation increases, the ability of LLMs to recover meaning diminishes, while detection rates for disinformation also vary across models. The study provides quantitative insights into the trade-offs involved, although specific numerical results and effect sizes are not detailed in the abstract.

Limitations
The authors acknowledge that their framework is primarily validated using COVID-19 disinformation as a case study, which may limit generalizability to other domains. Additionally, the reliance on a specific set of modulation strategies and the seven LLMs may not capture the full spectrum of linguistic evasion tactics or model capabilities. The study does not address potential ethical implications of using Algospeak in real-world applications, nor does it explore the long-term evolution of evasion strategies in response to detection advancements.

Why it matters
This research has significant implications for the fields of natural language processing and content moderation. By formalizing the dynamics of Algospeak and providing a framework for its analysis, the study lays the groundwork for future investigations into linguistic evasion strategies. Understanding the MUM threshold can inform the design of more robust detection systems and enhance the interpretability of LLMs in moderation tasks. Furthermore, the insights gained could influence policy-making regarding the dissemination of disinformation and the development of countermeasures in digital communication.

Authors: Jan Fillies, Ronald E. Robertson, Jeffrey Hancock
Source: arXiv:2605.06619
URL: https://arxiv.org/abs/2605.06619v1