When is not using AI safer in a war? Sudan responders described risks before a prompt was sent
An exploratory survey of 89 Emergency Response Room responders and 15 interviews found GenAI could help when expertise was dangerous to reach, while phones, connectivity and sensitive inputs could themselves expose people. The unreviewed study reports experiences—not proof that a tool improved aid.
The short version
- What happened
- The two-stage study used an exploratory survey of 89 responders recruited through network and participant referrals, followed by 15 interviews: nine GenAI users and six non-users. Fourteen interviews were conducted in Arabic and one in English.
- Who it affects
- Responders may gain provisional drafting or technical help when experts are dangerous to reach, but the same conditions can make verification impossible.
- What remains unproven
- Unreviewed exploratory study using network and participant-referral recruitment; the 89-person survey is not representative of Sudan's ERR network.
Keep in mind
Unreviewed exploratory study using network and participant-referral recruitment; the 89-person survey is not representative of Sudan's ERR network.
Fifteen interviews underrepresented Darfur and Kordofan and largely required connectivity, devices and relative safety.
Nine interviewees were GenAI users and six non-users; geographic differences limit attribution of themes to use rather than local conditions.
Usage was self-reported and all but one participant declined screen sharing, so effectiveness and actual error rates were not verified.
Arabic interviews were translated with GPT-6 and manually reviewed; GPT-6 also suggested candidate codes, which researchers checked rather than treating as independent coding.
The study does not measure aid delivery, medical outcomes, donor decisions, safety incidents or a comparator group.
Multi-source analysis · 2 linked sources. Primary analysed source date: 5 November 2026. Latest linked supporting source: 6 November 2026.
Read the full report · 9 min The evidence, methods, implications and limitations
The answer: in active conflict, responsible use can mean limiting or refusing the tool
An unreviewed study of Sudan's Emergency Response Rooms finds that war can make generative AI more useful at the same moment it undermines the conditions needed to use it safely. Responders described turning to chatbots when fighting made a doctor, engineer or colleague dangerous or impossible to reach. They also described circumstances in which carrying a phone, connecting to the internet or digitising beneficiary information could expose volunteers and communities. The paper's central finding is therefore not an adoption rate or a product recommendation. It is that the legitimacy of non-use must sit beside access and training in any responsible-AI policy for conflict settings.
The study does not show that a chatbot improved a medical, logistical or humanitarian outcome. Its survey was exploratory and non-representative, and interview accounts were almost entirely self-reported. One participant demonstrated a workflow; others declined for security and privacy reasons. The value lies in bringing the judgement of community responders into a field often dominated by model benchmarks and organisational policy. The ICRC's current guidance separately stresses that people remain legally responsible for military uses of AI and that adversarial conditions heighten failure risks; this study concerns civilian humanitarian work, not weapon or targeting systems.[1][2]
Who participated and how the research protected them
Researchers first distributed an Arabic survey to Emergency Response Room volunteers aged 18 or over through existing network contacts and participant referrals. In some cases, people with connectivity gave others in-person access to the form. After identical duplicates were removed, 89 unique responses remained. The survey covered roles, GenAI use and non-use, devices, languages, verification, organisational support, training, privacy, security and 23 Likert-scale questions. It was used descriptively to identify issues for interviews and was not designed to estimate GenAI use across the diverse ERR network or Sudan as a whole.
Fifteen follow-up interviews were conducted online in August and September 2026, lasting about 45–60 minutes. Nine participants reported using GenAI in ERR work and six did not. The purposive sample sought different roles, regions and areas controlled by different armed actors; 14 interviews were in Arabic and one in English. The study had institutional review-board approval. Participants were asked not to reveal precise locations or sensitive operations, identifying detail was generalised, data access was restricted and internet costs could be covered. Mental-health support was available if requested, although no participant requested it.[1]
The distinction that matters
What shaped responsible use in the recruited sample
These are descriptive survey results from 89 network- and referral-recruited responders, not national estimates.
Reported use
49 of 89
30 non-users; 10 unsure
Verification
70.8%
Said outputs usually need checking
Sensitive data
50.6%
Concerned about exposure
Stable internet
44.9%
Reported sufficient access
The survey shows perceived usefulness alongside weak conditions for verification
Forty-nine of 89 respondents—55.1%—said they used GenAI in ERR work, 30 reported non-use and 10 were unsure. Nearly two-thirds, 65.2%, agreed that it could be useful in fast-paced work, while 70.8% said outputs usually required verification. Only 41.6% agreed that generated information was usually accurate. Half, 50.6%, expressed concern about exposing sensitive information, and 44.9% reported sufficiently stable internet. Only 49.4% agreed that GenAI understood local words, places and context in Sudan. These figures describe the recruited sample and should not be projected to every responder.
The pattern matters because a tool can be attractive precisely where the human checks around it are weakest. Interviewees described using GenAI for drafting, organisation, translation and provisional technical suggestions. One respondent described using Gemini to assess possible shade on solar panels at a dialysis centre when measurement equipment and return access were uncertain, then discussing the result with the responsible engineer. Another described why a chatbot might be consulted during an urgent health incident when moving someone to a doctor could itself be dangerous. Neither account validates the advice; both show why a provisional answer can feel valuable under constraint.[1]
Ground truth, phones and sensitive data can fail before model accuracy is tested
Participants said prices, exchange rates, population movements, supplies and security could change faster than documents or online sources. Some used GenAI to organise information but refused to let it define community needs. Their alternative was direct contact with volunteers and affected people who could see what was happening. That distinction is important: retrieval can ground an answer in available sources, but it cannot retrieve knowledge that has not been safely recorded. In a fast-changing war, a fluent response can be stale or detached from the conditions that determine whether an action is feasible.
Several risks arose before anyone evaluated an output. One non-user described hiding or abandoning a phone because armed actors could interpret messages or searches as support for an opponent, exposing the owner to detention or death. Users also worried about entering beneficiary, financial or operational information into a service whose retention and access they could not control. De-identifying a prompt may reduce risk but cannot guarantee anonymity when role, place and circumstances can identify someone. A humanitarian AI checklist that begins with accuracy therefore starts too late; device safety, connectivity, data sensitivity and risk to other people come first.[1]
Donor reports reveal a second problem: assistance can improve language while weakening trust
Emergency Response Rooms depend on donors to sustain food, medical and other aid. Participants said GenAI could help volunteers draft English-language reports for funders, particularly when English was not their first language. The same assistance could make a document look detached from what happened on the ground or raise suspicions that claims were fabricated. Interviewees emphasised that reporting must preserve actual activities, feedback, challenges, local languages and community understanding. Fluency is not the same as factual or cultural fidelity, and disclosure alone does not establish that a report is accurate.
The authors suggest that donors distinguish translation, editing and organisation from fabrication, and assess the evidence supporting claims rather than penalising all AI-assisted language. Organisations can preserve user-provided facts, mark generated additions, require local review and link claims to documentation without exposing sensitive operational detail. Providers can help by making retention terms clear, offering low-bandwidth interfaces and giving users meaningful control over data storage. Those are proposals derived from participants' accounts; the study did not test whether they protect people, improve donor decisions or increase funding.[1]
What humanitarian organisations can do—and what evidence would change the assessment
A task-specific protocol can ask whether using a device or network creates danger, whether the prompt contains sensitive information, whether current local knowledge exists, what harm follows from an error, and who can verify the result. Low-consequence drafting support is different from medical triage, needs assessment or resource allocation. When qualified review is unavailable, an output may remain explicitly provisional and should not become the sole basis for a consequential decision. Teams also need safe non-AI alternatives: transferring a task to another volunteer can move rather than eliminate the risk.
Confidence would rise with more diverse recruitment, direct observation that does not endanger participants, longitudinal follow-up and comparison with other conflict-affected humanitarian networks. Future work should separate tasks, tools and access conditions, document actual errors and trace whether advice changes decisions or outcomes. The present evidence supports listening to responders who choose to use, limit or avoid GenAI and treating local knowledge as part of the safety system. It does not support a claim that chatbots are effective emergency clinicians, that 55.1% of all ERR volunteers use them or that any single policy can be transferred unchanged to another conflict.[1][2]
Keep following the evidence
Join The Impact Brief founding list for selected AI reports, their effect on people and the evidence limits. Read a sample first.
Continue the story
Read next
Continue exploring generative ai: compare the finding and its evidence limits.
Australia now requires AI use to be disclosed in workplace cases—what changes?
From 20 October, anyone using generative AI to prepare a document for Australia's Fair Work Commission must say so, explain how it was used and verify the result. Witness statements face an extra safeguard: they must remain the person's own knowledge, words and truthful account.
Read the next reportExplore the full source and evidence record
What this means for people
- Responders may gain provisional drafting or technical help when experts are dangerous to reach, but the same conditions can make verification impossible.
- Carrying a phone, connecting to the internet or submitting beneficiary information can create risks before model quality is considered.
- Communities and donors need reports grounded in events and local language; fluent generated text can support communication without becoming evidence of what happened.
Global context
The study centres Sudanese community responders, a population commonly missing from responsible-AI research, but conflict conditions and humanitarian structures differ across countries. The generalisable question is not whether every organisation should adopt or ban GenAI. It is whether people can safely access it, protect information, verify outputs against current local knowledge and refuse use without losing support. Those conditions must be assessed locally with affected communities.
What the evidence does not yet show
- Unreviewed exploratory study using network and participant-referral recruitment; the 89-person survey is not representative of Sudan's ERR network.
- Fifteen interviews underrepresented Darfur and Kordofan and largely required connectivity, devices and relative safety.
- Nine interviewees were GenAI users and six non-users; geographic differences limit attribution of themes to use rather than local conditions.
- Usage was self-reported and all but one participant declined screen sharing, so effectiveness and actual error rates were not verified.
- Arabic interviews were translated with GPT-6 and manually reviewed; GPT-6 also suggested candidate codes, which researchers checked rather than treating as independent coding.
- The study does not measure aid delivery, medical outcomes, donor decisions, safety incidents or a comparator group.
What to watch next
- Peer review and publication of the full survey instrument, coding materials and ethics detail that can be shared safely.
- Longitudinal evidence on how access, rules and verification change as conflict conditions change.
- Task-level observation of errors and benefits that does not expose responders or beneficiaries.
- Provider support for low bandwidth, clear retention controls and genuinely usable zero-retention options.
- Donor policies that distinguish language assistance from unsupported factual claims.
Living evidence record
Impact record IAI-0DINR92
Evidence stage
Studied
Confidence
Supported
Reporting basis
Multi-source analysis
Independent or research support
Present
Record status
Monitoring
Last checked
6 November 2026
Source trail
2 direct sources across 2 source types.
People impact
Documented in this record.
Uncertainty
Limits and next checks are explicit.
Stages describe the evidence available—not whether a technology is good or bad. See the public method.
Evidence trail
Sources used for this report
Links checked 6 November 2026
Research question, topics and regions
How community-based humanitarian responders in Sudan use, limit or avoid generative AI during active conflict, and which conditions make responsible use harder
Topics: Humanitarian response · Armed conflict · Generative AI · Digital safety · Local knowledge · Human verification
Regions: Sudan · Africa · Global
This report is labelled multi-source analysis. We summarise and analyse source material in our own words; company statements remain attributed claims until independently supported. Translated summaries preserve the meaning of the original source and link back to it. Read our editorial standards.
Reader commentary
What does this mean where you are?
Tell us what you have seen at work, in your community or in everyday life. A question is welcome too. No account is needed; we review notes before publishing them.
Explore commentary across the portal →Published reader notes
0No published reader notes yet. You can start the evidence-led discussion above.
Prefer a private correction or response? Contact the newsroom.