Responders in Sudan’s community-led Emergency Response Rooms described using generative AI for suggestions and provisional technical advice, while relying on people to judge changing local needs. Others chose not to use it because their work depended on knowledge built with communities.
Those accounts appear in a new arXiv preprint, Responsible GenAI Use During Active Armed Conflict: Evidence from Sudan’s Emergency Response Rooms, by Dalia Ali, Yearin Lee and colleagues. Its accessible findings show how volunteers weighed possible assistance against gaps in current local information, dangerous connectivity and uncertainty about sharing sensitive data. They do not establish that AI improved aid outcomes.
The first version was submitted on 5 November 2026. As checked on 7 November, the arXiv record showed no journal reference or evidence of peer review. This report covers the accessible methods and early findings; we could not inspect the later discussion or appendices.
The 15 interviewees came from the 89 survey respondents
Emergency Response Rooms are a community-led humanitarian network in Sudan. For their exploratory survey, the researchers recruited adults through ERR contacts and participant referrals. They obtained 89 unique responses to an Arabic questionnaire. Some connected responders helped peers without connectivity participate in person.
The researchers then selected 15 interviewees from survey respondents who had agreed to follow-up and supplied contact details. These were overlapping samples, not 104 separate people. The online interviews took place in August and September 2026 and included nine users of generative AI and six non-users.
The authors explicitly say the survey was not intended to produce representative estimates across ERRs. Recruitment through personal networks, followed by interviews with reachable, consenting respondents, also limits whose experiences the findings can capture. These self-reported accounts cannot establish how common a practice is across Sudan or measure a chatbot’s accuracy.
A chatbot’s technical suggestion still went to an engineer
In the paper’s early findings, responders reported using chatbots for suggestions for children’s activities and distress-related support. The accounts do not establish clinical benefit or evaluated safety.
One responder described asking Gemini for a provisional assessment of a solar power system, then taking the issue to the responsible engineer. According to the account, the engineer had already considered the issue and said it would be addressed.
The account describes a provisional AI assessment followed by expert consultation, with no measured technical improvement.
Local knowledge shaped both use and non-use
Participants described limits that arise when information changes with conditions on the ground. In the section on local context, prices and needs were matters to check with people working in the affected communities.
A selective user described using AI to organize existing knowledge while obtaining project needs from volunteers and communities. A non-user emphasized the iterative work of understanding a problem locally and adapting a response.
A responder may decide that the relevant knowledge comes from ongoing relationships and observation.
Phone danger and feared AI disclosure were different risks
The privacy and safety accounts describe risks at different stages of access. One participant recounted arrest and danger associated with carrying a phone during wartime work. That account concerns an existing device risk; it does not describe an arrest caused by generative AI.
Participants also feared what could happen if sensitive information were disclosed through an AI service. Some described withholding such information or asking for generic drafts and adding sensitive details themselves afterward. These were reported precautions against possible disclosure, not evidence of a demonstrated leak.
The accounts did not establish consistent safeguards. The authors describe three other participants whose reported workflows appeared to involve original records or identifying information, without making clear whether those materials had been de-identified. Stated boundaries therefore leave an important uncertainty about actual handling of data.
The researchers also used AI to interpret the interviews
The authors report using AI assistance in translation and qualitative analysis. A native Arabic-speaking author checked translations against the originals. After initial coding, GPT-6 suggested candidate codes and patterns that researchers checked against transcripts; it was not treated as an independent coder.
