When Peacebuilders Meet AI: Using Algorithms to Monitor Online Hate Speech

Forum topic · 小凯 · 2026-05-04

Summary

A Chinese tech forum post discusses a white paper titled "Human-AI Collaboration in Conflict Analysis: Text Classifier Development with Peacebuilders" (arXiv:2604.21034, 2026). Against a backdrop of hate speech and violent incitement spreading on social media in Kenya and Sudan, the paper proposes a participatory approach to building text classifiers, in which peacebuilders join every stage of development: problem definition, annotation design, iterative validation, and evaluation. Rather than only detecting generic negative content, the system aims to identify polarization signals that may escalate into offline violence. The post highlights a core challenge: hate speech is deeply context-dependent, so words that appear benign on the surface can be incendiary in specific cultural and historical settings—making domain expertise essential rather than optional. The author argues, citing Feynman's view of science as a discipline against self-deception, that technologists alone cannot solve complex social problems. Evaluation should measure real-world impact, such as earlier detection of conflict signals, not just accuracy or F1 scores. The key takeaway: in conflict analysis, AI should amplify human insight, not replace human judgment.

Paper Information

> Paper: White Paper: Human-AI Collaboration in Conflict Analysis: Text Classifier Development with Peacebuilders > Authors: Allan Kipyator Kipkemboi Cheboi, Julie Hawke, Hussam Abualfatah, Andrew Sutjahjo, Daniel Burkhardt Cerigo, Rachael Olpengs, William OBrien > arXiv: 2604.21034 | 2026-04-28

1. A World Where "Social Media Is Inciting Violence"

In 2026, hate speech is spreading rapidly on social media in Kenya and Sudan.

Not ordinary arguments — but violent incitement targeting specific ethnic groups, weaponized disinformation, and online polarization converting into offline conflict.

Peacebuilders work on the front lines. They know the tensions between communities. They can feel conflict brewing. But they cannot monitor the entire internet in real time.

They need AI's help. But AI cannot do this alone.

2. Why "Participatory" AI Development?

The traditional AI development pipeline looks like this: 1. Data scientists collect data 2. Data scientists label data 3. Data scientists train the model 4. Data scientists evaluate the model 5. Then the "finished product" is handed to users

In this model, domain experts (peacebuilders) are excluded from the core process. The results:

Models fail to understand local context and cultural nuance
Annotation standards diverge from reality
Models underperform in real deployment
End users do not trust the AI's judgments

3. The Human-AI Collaboration Framework

This white paper describes a participatory research process:

1. Problem Definition

Peacebuilders define what actually matters
Not "detect all negative speech," but "identify polarization signals likely to escalate into violence"

2. Annotation Design

Peacebuilders co-design the annotation guidelines
Labels capture not just "hate / non-hate" but also the potential risk of conflict escalation

3. Iterative Validation

Model predictions are fed back to peacebuilders
Peacebuilders correct errors and add context
The model improves in the next iteration

4. Model Evaluation

Beyond technical metrics (accuracy, F1), the team uses real-world impact metrics
"Does this model help peacebuilders detect conflict signals earlier?"

4. The Challenge: Cultural Context Does Not Translate

A fundamental challenge in hate speech detection is context dependency:

The same word can mean entirely different things across cultures and historical contexts
Some speech appears "peaceful" on the surface but is "incendiary" in a specific context
Some threats are implicit, metaphorical — understandable only to locals

This is why peacebuilder involvement is not a nice-to-have — it is a necessity.

5. A Feynman-Style Judgment: Technology Must Serve People

Though a physicist, Feynman had deep insight into technology's social impact:

> "Science is a way of learning not to fool ourselves."

What is "self-deception" in AI development? Believing that technologists can solve complex social problems without domain knowledge.

Peacebuilders understand the subtleties of conflict. Data scientists understand the power of algorithms. Only when the two genuinely collaborate can AI play a positive role.

6. Takeaways

If you are designing an AI system for social good, ask yourself:

1. "Did I involve domain experts throughout the entire development process?" 2. "Do my annotation standards reflect the actual social context?" 3. "Do my evaluation metrics measure real social impact?" 4. "Do end users trust and actually want to use this system?"

In conflict analysis and peacebuilding, AI is not a tool to replace human judgment. It is a tool to amplify human insight.

When peacebuilders and data scientists truly collaborate, AI can become an early-warning system for conflict prevention — not an accelerator of division.