In traditional content moderation, toxicity is often treated as a surface-level feature. A keyword filter flags a slur; a computer vision model flags graphic violence. However, internet memes break this paradigm through a phenomenon known as emergent toxicity. As explored in Safe + Safe = Unsafe? by Chenhang Cui and colleagues, a meme can be composed of an entirely benign image and a neutral caption, yet become hateful or harmful when the two are combined.
This research direction focuses on bridging the "semantic gap" between modalities to understand how meaning is constructed not within a single medium, but in the interaction between them.
The Paradox of Individually Benign Modalities
The fundamental challenge of meme detection is that the harmful intent is often "hidden" in the alignment of disparate signals. Yi Zhou 0042 notes in Multimodal Learning for Hateful Memes Detection that hate speech in memes frequently lives in the "gap" between the image and text.
Consider a meme featuring a picture of a common household appliance paired with a caption about "returning to your natural habitat." In isolation, neither component is offensive. However, when paired, they may evoke a sexist stereotype. Standard unimodal filters fail here because they lack the cross-modal reasoning to identify the emergent message. This is what Gaurav Mittal and the team behind AM3 describe as "hidden" meaning—toxicity that exists only at the intersection of vision and language.
Disentangling Targets and Causal Concepts
To move beyond simple binary classification (hateful vs. not hateful), researchers have turned to disentanglement and causal reasoning.
In DisMultiHate, Roy Ka-Wei Lee and Rui Cao propose forcing models to explicitly identify the target of the hate (e.g., race, religion, or gender). By disentangling who is being attacked, the model gains a more structured understanding of the meme's intent.
Building on this, Dibyanayan Bandyopadhyay and Asif Ekbal used Structural Causal Models in Seeing Through VisualBERT to identify high-level concepts—like "antisemitism"—that actually cause a model to flag a meme. This shifts the focus from identifying "bad words" to identifying the underlying social aggression or normalization of violence that Thamar Solorio identifies as a core component of online harm.
Knowledge Infusion and Reasoning Distillation
Because memes are deeply rooted in cultural context, models often require external knowledge to "get the joke." Recent work has moved toward neurosymbolic and catalyst-based approaches:
- Knowledge Graphs: Hemang Jain introduced KID-VLM, which infuses explicit relational facts from Knowledge Graphs into Vision-Language Models. This helps the model understand that a specific historical figure or symbol carries a weight that isn't apparent from the pixels alone.
- LLM as a Catalyst: In OSPC, Jingtao Cao and colleagues use Large Language Models (LLMs) to generate natural language explanations of a meme's context before classification. This "reasoning distillation" allows smaller, more efficient models to benefit from the vast cultural knowledge stored in models like GPT-4V.
- Multimodal Knowledge Enhancement (MKE): Junyu Lu and Bo Xu demonstrated in TOXICN MM that explaining the cultural context of a meme significantly improves detection performance, especially in linguistically complex environments like Chinese social media.
The Challenge of Cultural Safety
As these models become more sophisticated, the research is expanding into multimodal cultural safety. Haoyi Qiu and Nanyun Peng introduced the CROSS benchmark to measure how well AI avoids symbolic harm, such as violating religious taboos.
The consensus across papers like MOMENTA and MultiHateClip is that detection is no longer just a classification task; it is a reasoning task. To truly identify harmful memes, a system must synthesize text, image, and cultural metadata to decode the emergent intent that lives between the lines.