The Universal Problem with No Universal Solution

One of the most significant challenges in building effective hate speech detection systems is that hate speech itself defies universal definition. What constitutes hateful or harmful content varies dramatically across cultures, communities, and contexts—yet most AI systems attempt to apply one-size-fits-all solutions.

The Context Problem

Consider these examples:

• A phrase that's deeply offensive to one religious community might be completely benign to another
• Historical references that trigger trauma in one culture may be unknown in another
• Slurs that have been reclaimed by some communities remain harmful when used by outsiders
• Satirical content that's acceptable in one context becomes problematic in another

Traditional AI approaches often fail to capture these nuances, leading to either over-censorship of legitimate speech or under-moderation of genuinely harmful content.

Why Current Approaches Fall Short

Most large-scale content moderation systems rely on:

1. Binary classifications (hate speech vs. not hate speech)
2. Western-centric training data that doesn't represent global perspectives
3. Static definitions that don't evolve with communities
4. Lack of community input in the model development process

This results in systems that may work reasonably well for dominant groups but fail marginalized communities – the very people who most need protection from hate speech.

A Community-Centered Approach

At definehate.org, we're exploring a different path. Instead of imposing universal definitions, we're working with affected communities to:

• Document their specific experiences with hate speech
• Understand their cultural context and historical trauma
• Capture the evolution of harmful language over time
• Include their voices in defining what constitutes harm

The Technical Challenge

Implementing community-centered hate speech detection requires fundamental shifts in the machine learning approaches that have been historically used:

Multi-Stakeholder Training Data

Rather than training on generic datasets, we need training data that reflects the experiences and perspectives of different communities.

Context-Aware Models

AI systems need to understand not just what was said, but who said it, to whom, and in what context.

Dynamic Definitions

Models must be able to evolve as communities' understanding of harmful speech changes.

Transparent Decision-Making

Communities need to understand how and why content moderation decisions are made – both the targeted community, as well as those who are learning about that community.

Moving Forward

The path forward may not be direct. We do not yet know the unknowns that lie ahead of us. We do, however, have principles that will guide us through this process. Building hate speech detection systems that truly serve all communities requires:

• Sustained engagement with affected communities
• Investment in diverse perspectives throughout the development process
• Understanding unique community needs rather than seeking simple solutions
• Commitment to ongoing improvement based on community feedback

AI and Data Science Researchers

If you're working on content moderation, natural language processing, or fairness in AI systems, we would love to connect.

Community Advocates

Your lived experiences and deep understanding of how hate speech impacts your communities are invaluable to this work. We are actively seeking partnerships with advocates, community leaders, and organizations who can help to accurately label hate speech from your unique perspective.

An accurately labeled dataset of hate speech is an asset that could be used for tremendous good. It can be leveraged, scaled, and used in ways we have not yet envisioned. In a world increasingly driven by algorithms, we recognize the human vulnerability to sensationalized hate that drives engagement, the monetary, political, and personal gains that can result from exploiting this vulnerability, and the necessity of equipping data scientists, researchers, and platforms with tools they can use to counter harmful narratives and build more inclusive online spaces.