Paper Type
Complete
Abstract
Toxic algospeak- coded language conveying harm while evading detection-enables hate speech to bypass content moderation. This paper investigates how audience expertise moderates algospeak effectiveness through dual cognitive mechanisms. Analyzing 13,780 high-evasion posts (TES ≥ 0.60) from 721,236 Reddit climate discussions, we find metalinguistic awareness (recognizing euphemistic labeling) enables universal detection without harm discounting, while epistemic sophistication (evaluating scientific jargon misuse) enables expert-specific detection with 16.5% harm reduction. LLM-based harm assessment achieved substantial agreement with expert consensus (κ = 0.73, r = 0.66), validated against hate speech baseline (r = 0.991). Uniform governance fails: metalinguistic patterns benefit from user education; epistemic patterns require domain experts. This dual-mechanism framework advances information systems research on human-AI collaboration in platform governance, applicable across various domains, such as health and politics.
Paper Number
1875
Recommended Citation
Agarwal, Nitin and Falade, Tope, "Beyond Algorithmic Detection: How Community Expertise Moderates Toxic Algospeak via Dual Cognitive Mechanisms" (2026). AMCIS 2026 Proceedings. 9.
https://aisel.aisnet.org/amcis2026/sig_odis/sig_odis/9
Beyond Algorithmic Detection: How Community Expertise Moderates Toxic Algospeak via Dual Cognitive Mechanisms
Toxic algospeak- coded language conveying harm while evading detection-enables hate speech to bypass content moderation. This paper investigates how audience expertise moderates algospeak effectiveness through dual cognitive mechanisms. Analyzing 13,780 high-evasion posts (TES ≥ 0.60) from 721,236 Reddit climate discussions, we find metalinguistic awareness (recognizing euphemistic labeling) enables universal detection without harm discounting, while epistemic sophistication (evaluating scientific jargon misuse) enables expert-specific detection with 16.5% harm reduction. LLM-based harm assessment achieved substantial agreement with expert consensus (κ = 0.73, r = 0.66), validated against hate speech baseline (r = 0.991). Uniform governance fails: metalinguistic patterns benefit from user education; epistemic patterns require domain experts. This dual-mechanism framework advances information systems research on human-AI collaboration in platform governance, applicable across various domains, such as health and politics.
When commenting on articles, please be friendly, welcoming, respectful and abide by the AIS eLibrary Discussion Thread Code of Conduct posted here.
Comments
SIGODIS