TL;DR
This paper explores the "conflict paradox" of Generative AI (GenAI): active warfare makes these tools incredibly valuable as a fallback for unreachable human experts, yet simultaneously undermines the very conditions—like safety, connectivity, and ground-truth verification—required to use them responsibly. By studying Sudan’s community-led Emergency Response Rooms (ERRs), the authors demonstrate that responsible AI in crisis zones isn't just about model accuracy; it’s about the physical safety of carrying a device and the fragile credibility of local responders in the eyes of international donors.
What Problem Does This Paper Address?
Most research on "Responsible AI" assumes a stable environment with reliable internet, institutional oversight, and users who aren't at risk of being arrested for carrying a smartphone. This paper addresses the massive gap in our understanding of how community-based humanitarian responders—specifically those in Sudan’s ERRs—actually use or avoid GenAI during an active armed conflict.
In Sudan, where 37% of health facilities are non-functional and millions are displaced, local volunteers have stepped in where international NGOs cannot reach. These responders face a unique set of challenges: they need rapid information (e.g., medical advice or technical solar repair steps) but operate under "digital traces" that can lead to detention or death if discovered by armed actors. The paper investigates how wartime conditions reconfigure the utility and risks of tools like ChatGPT and Gemini.
Methods
The researchers employed a two-stage empirical approach:
- Exploratory Survey: 89 ERR responders provided data on their roles, GenAI usage patterns, and perceptions of risk.
- Semi-Structured Interviews: 15 in-depth interviews (9 users, 6 non-users) were conducted to explore the nuances of why and how they integrated (or rejected) these tools.
Notably, the research team included members of the ERR network to ensure cultural and contextual accuracy. They also used GPT-6 (as referenced in the paper's 2026-dated metadata) to assist with translating Arabic interviews and generating candidate codes for thematic analysis, though final analytical decisions remained with the human researchers.
Datasets
This is a qualitative and exploratory study; it does not introduce a new machine learning benchmark or training dataset. The primary "data" consists of:
- Survey Responses (N=89): Capturing demographics, primary roles (coordinators, field volunteers), and Likert-scale agreement on AI utility and risks.
- Interview Transcripts (N=15): Detailed accounts from responders across various Sudanese states (Khartoum, Darfur, Kordofan, etc.).
Key Insights and Ideas
- The Fallback Utility: Responders turned to GenAI specifically when human expertise was "difficult or dangerous to reach." For example, a volunteer might use a chatbot to manage a medical emergency (like a seizure) when taking the patient to a doctor would involve crossing a frontline.
- The "Ground Truth" Gap: Participants emphasized that GenAI cannot independently establish wartime reality. Prices, exchange rates, and safe routes change daily. As one participant noted, AI gives "indicators," but the final decision must come from a human "on the ground."
- Safety Before Use: In a conflict zone, the risk starts before the prompt. Carrying a phone can be interpreted by armed actors as support for the opposition. For some, "having a phone was itself a crime."
- The Credibility Trap: There is a tension in using GenAI to write reports for English-speaking donors. While it helps bridge language gaps, volunteers fear that "AI-polished" reports might look fake or detached from the grim reality on the ground, potentially jeopardizing the very funding that sustains their food and medical aid.
- Collective Responsibility: Verification isn't always done by the person who prompted the AI. One responder might have the internet access to use the tool, while another has the local expertise to verify if the output makes sense in the current context.
Key Results
- Usage Rates: 55.1% of surveyed responders reported using GenAI in their ERR work.
- Top Tasks: The most common uses were translation (44.9%), data analysis (42.9%), and social media content creation (40.8%).
- Verification Consensus: 70.8% of respondents agreed that they "usually need to verify GenAI outputs," and 62.9% felt human review was essential even in urgent situations.
- Barriers to Entry: For non-users, the primary hurdles were a lack of knowledge on how to use the tools (26.7%) and distrust of the outputs (23.3%).
The study also provides a detailed breakdown of the participants' roles and locations, highlighting the diversity of the ERR network:
Limitations and Future Work
The authors acknowledge several limitations:
- Selection Bias: Participation required some level of connectivity and safety, likely underrepresenting the most marginalized responders in the most intense combat zones.
- Self-Reporting: The study relies on what participants said they did, rather than direct observation of their AI interactions (which was declined by participants for security reasons).
- Geographic Skew: Responders from Darfur and Kordofan were less represented in the interview phase than in the survey.
Future Work: The authors call for longitudinal studies to see how AI use evolves as conflicts persist and for technology providers to develop "conflict-sensitive" features, such as zero-data-retention modes and low-bandwidth interfaces tailored for unreliable wartime connectivity.
Summary updated