Responsible GenAI Use During Active Armed Conflict: Evidence from Sudan's Emergency Response Rooms
B · Useful, limited A Technical University of Munich-led team, with members of Sudan's Emergency Response Rooms, gives rare first-hand evidence from responders in an active war; the survey is exploratory and non-representative.
Summary based on the full text·Original: about 45 minutes to read
In short
An arXiv paper (5 November 2026) by Dalia Ali, Yearin Lee and colleagues (TU Munich, University of Pennsylvania, Saarland University, and Sudan's Emergency Response Rooms). It studies how volunteer humanitarian responders in Sudan use or avoid GenAI during the war. The core finding is a trap: war makes GenAI most useful where its outputs are hardest to verify, because human expertise is unreachable. Responders also fear that AI-written reports will cost them credibility with donors.
Key numbers
- 89survey respondents from Sudan's Emergency Response Rooms, plus 15 semi-structured interviews
- 55.1%respondents using GenAI in their ERR work (33.7% not, 11.2% unsure)
- 41.6%respondents agreeing GenAI-generated information is usually accurate
- 49.4%respondents agreeing GenAI understands local words, places and context in Sudan
- 50.6%respondents concerned about exposing sensitive information
As reported by the source.
English·Web page·opens arxiv.org
What it reports
- Of 89 respondents, 49 (55.1%) used GenAI in ERR work; top tasks were translation (44.9%), data analysis (42.9%) and social media content (40.8%).
- Top reasons for non-use: not knowing how (26.7%), distrust of outputs (23.3%), unreliable internet (16.7%).
- 65.2% agreed GenAI could be useful in fast-paced ERR work; 70.8% agreed outputs usually need verification; only 41.6% agreed outputs are usually accurate.
- Risks start before use: carrying a phone or going online can expose volunteers to danger, and sharing beneficiary data with a model can endanger people.
- Verification is social: users rely on colleagues with local knowledge to check outputs.
- Donor credibility: participants worried AI-assisted reports would undermine trust from the funders who pay for food and medical aid.
How they got it
A two-stage design: an Arabic-language exploratory survey distributed through the ERR network by referral, then 15 interviews (14 in Arabic) with nine users and six non-users, August-September 2026. IRB approval, de-identification before LLM-assisted translation and coding (GPT-6), and an offer to cover internet costs.
Limits
- Referral sample; the survey does not estimate use across the more than 700 ERRs.
- Likert answers measure stated agreement, not behaviour.
- Two authors belong to the network studied, which aids access but may shape interpretation.
Why it matters for companies
The findings apply to any field team that works with poor connectivity, sensitive data and no expert on call: verification capacity, not model quality, sets the limit on safe use. Organisations funding or supervising such teams should publish clear rules on AI-assisted reporting so that disclosure does not cost credibility, and should treat a decision not to use AI as legitimate. Data-protection guidance needs to cover what staff type into public tools, not only what systems store.
Summary based on the full text. Read the original for the full argument and data.Read the original on arxiv.org
- Tags
- Future of work