Check out this new article by Aqsa Iftikhar and Ramesh Kumar Ayyasamy, which reviews how artificial intelligence is changing the automated detection of online hate speech, and where current systems still fall short.
The study examines how hate speech detection has evolved from traditional machine learning to deep learning, transformer models and large language models, or LLMs. It focuses particularly on two problems that remain difficult for automated systems: detecting implicit hate and working across different languages and cultural contexts. It also examines whether researchers are adequately testing their models for bias, fairness and computational efficiency.
The authors conducted a systematic review of 56 peer reviewed studies published between 2017 and 2025. They followed established PRISMA guidelines for systematic reviews and assessed each study against ten quality criteria. This allowed them to examine not only whether newer AI models perform better, but also whether research in this area is addressing some of the problems that matter when these systems are used in practice.
The review shows a clear shift towards newer AI approaches. More than half of the studies, 55.4%, used transformer based models or LLMs. These methods have improved the ability to understand language in context, but important problems remain. Only 30.4% of the studies fully addressed implicit hate speech. This matters because online hate is often communicated through coded language, sarcasm, stereotypes, dog whistles or references whose meaning depends on social and cultural context.
The review also identifies gaps in how these technologies are evaluated. Only 35.7% of studies fully reported fairness or bias analysis, while 44.6% reported information about computational costs. There has been some progress: studies published after 2022 were about twice as likely to report fairness evaluations as earlier studies. But fairness testing is still far from standard practice.
The authors identify knowledge augmented LLMs as one direction for future research. These systems combine language models with structured sources of knowledge, potentially helping models interpret cultural references, context and implicit meanings. Yet only six of the 56 studies, 10.7%, used this type of approach, all published in 2024 or 2025. The authors therefore propose further research on combining LLMs with structured knowledge while improving multilingual performance, contextual reasoning and debiasing.
For policymakers and technology companies, the message is that improved AI models do not automatically produce reliable hate speech detection. Systems need to be evaluated for implicit and multilingual hate, cultural context and fairness before they can be relied upon for content moderation. Better reporting standards are also needed so that governments, platforms and researchers can understand not only how accurate a system is, but where and for whom it is likely to fail.