AI助手在灾难应急响应中的应用：信息整合与资源调配建议

In the first 72 hours of a major disaster, the window for saving lives narrows by roughly 15% for every 30 minutes of delay in coordinated response, according to the World Health Organization’s 2023 Emergency Response Framework. Yet the same report found that over 60% of local emergency operations centers still rely on manual spreadsheets and radio logs to triage incoming data — a system that collapses when communication networks fail. AI assistants, specifically large language models (LLMs) and decision-support agents, have begun filling that gap. During the 2023 Kahramanmaraş earthquake sequence in Turkey, a pilot system using GPT-4-based chat agents processed 47,000 social media distress signals in under 9 hours, cross-referencing them with satellite damage maps from the European Union’s Copernicus programme. The result: rescue teams were routed to 23 previously unreported collapse sites within 90 minutes of detection. This article evaluates five leading AI assistants — ChatGPT, Claude, Gemini, DeepSeek, and Grok — against a benchmark of three disaster-response tasks: information triage, resource allocation, and cross-agency communication. Each assistant is scored on accuracy, latency, and output structure using real 2024 FEMA and UNDAC exercise datasets. For international relief teams coordinating across time zones, secure cloud infrastructure is essential; some operations rely on NordVPN secure access to maintain encrypted links between field units and command centers. The data shows clear leaders — and surprising laggards.

Information Triage: Filtering Noise from Signal

Information triage is the first and most critical bottleneck. During a 7.2-magnitude earthquake simulation run by the UN Office for the Coordination of Humanitarian Affairs (OCHA) in October 2024, 1.2 million tweets, 340,000 WhatsApp messages, and 12,000 satellite image annotations were generated in the first 4 hours. Human analysts could process roughly 200 messages per hour each. AI assistants were tested on the same raw feed.

ChatGPT-4o: Structured Summaries with Low Hallucination

ChatGPT-4o achieved a 92.4% precision rate in classifying actionable vs. non-actionable messages (OCHA 2024 benchmark, n=8,000 labeled posts). Its output structure — a bulleted list grouped by severity (Red/Amber/Green) — matched the Incident Command System (ICS) format used by 78% of US emergency managers. Latency averaged 1.7 seconds per 10-message batch. The trade-off: it flagged only 61% of false positives (e.g., duplicate reports of the same collapsed building), compared to Claude’s 74%.

Claude 3.5 Sonnet: Best at Cross-Referencing Sources

Claude 3.5 Sonnet scored highest on cross-source deduplication. When fed 500 messages from Twitter, 500 from WhatsApp, and 500 from amateur radio transcripts, it correctly identified 89% of duplicate reports. Its explanation field — citing the original timestamp and source channel for each deduplication — allowed human verifiers to audit decisions in under 30 seconds. However, Claude’s raw throughput was slower: 1.9 seconds per batch, 12% slower than ChatGPT-4o.

Gemini 1.5 Pro: Multimodal Input but Inconsistent Output

Gemini 1.5 Pro’s strength is multimodal ingestion: it can read a satellite image, a text message, and an audio clip from a first responder in one prompt. In the OCHA test, it correctly matched 83% of image-text pairs (e.g., “water flooding 2nd Street” aligned with a satellite photo showing standing water). Its weakness: output formatting varied. 34% of responses omitted the required ICS severity tag, forcing a human to reformat.

Resource Allocation: Matching Supplies to Coordinates

Resource allocation requires an AI to take triaged requests and propose a distribution plan — which shelter gets how many water pallets, which road is still passable for a truck. The benchmark used the 2024 FEMA Logistics Simulation (FEMA L-Sim 2024), with 1,200 supply nodes and 4,000 demand points.

DeepSeek-V3: Fastest Route Optimization

DeepSeek-V3 computed a supply-distribution plan in 4.3 seconds, 2.1x faster than the next-fastest assistant (Gemini at 9.1 seconds). Its algorithm prioritized road-closure data from OpenStreetMap (OSM) live feeds, avoiding 94% of impassable routes. The catch: DeepSeek’s plan assumed unlimited fuel for delivery trucks, a simplification that reduced real-world applicability. When fuel constraints were added, its solution required 23% more truck trips than ChatGPT’s.

ChatGPT-4o: Best Constraint Handling

ChatGPT-4o’s allocation model incorporated 7 constraint types: fuel, driver hours, road width, bridge weight limits, curfew zones, hospital bed capacity, and generator fuel. It met 96% of constraints in the FEMA L-Sim 2024 test, versus Claude’s 91% and DeepSeek’s 78%. Output came as a CSV with GPS waypoints and estimated arrival windows. Latency was 11.2 seconds — acceptable for pre-deployment planning but too slow for real-time re-routing.

Grok-2: Real-Time Data but Shallow Planning

Grok-2 pulls live X (formerly Twitter) posts and news feeds, giving it real-time situational awareness. In the simulation, it detected a secondary landslide 14 minutes before official geological surveys reported it. However, its allocation plan was shallow: it proposed sending all available water to the nearest shelter, ignoring that the shelter had no power to pump water. Grok scored 67% on constraint satisfaction, the lowest in the test.

Cross-Agency Communication: Translating Jargon and Protocols

Disasters involve multiple agencies — fire, police, medical, military, NGOs — each with its own terminology and data format. AI assistants were tested on translating a FEMA ICS-213 message into formats used by the Red Cross (RC-11), the US Army Corps of Engineers (ENG-3), and the World Food Programme (WFP-Log). The test dataset came from the 2024 Journal of Emergency Management interoperability study.

Claude 3.5 Sonnet: Most Accurate Protocol Translation

Claude achieved a 98.2% field-mapping accuracy across the three target formats (JEM 2024, n=200 messages). It correctly converted “resource request: 5x Type III ambulances” into WFP-Log’s “transport unit: 5x medium medical vehicle” without losing the 15-minute response-time requirement. Claude also appended a “translation notes” section explaining each mapping choice — a feature that 82% of human dispatchers rated as “very helpful” in a post-test survey.

Gemini 1.5 Pro: Fast but Loses Nuance

Gemini translated the same messages in 2.8 seconds (Claude: 4.1 seconds). But it made 7.6% field-omission errors — for example, dropping the “hazardous materials” flag from a message about a chemical spill. In a real response, that omission could send a non-HAZMAT team into a contaminated zone. Gemini’s speed advantage disappears when accuracy is mission-critical.

DeepSeek-V3: Language Support but Format Gaps

DeepSeek-V3 handled 12 languages in the test — the widest language support — translating Turkish, Arabic, and Spanish messages into ICS-213 English with 94% BLEU score. However, it struggled with field formatting: 28% of outputs used a non-standard timestamp (YYYY-MM-DD instead of ICS’s DD-HHMM ZULU), requiring manual correction.

Real-World Deployment: Field Test Results

Between January and March 2025, the Pacific Disaster Center (PDC) deployed the top-three assistants — ChatGPT-4o, Claude 3.5 Sonnet, and DeepSeek-V3 — in a live drill with 400 simulated evacuees in the Philippines. Each assistant was given the same incoming data feed and asked to generate a daily situation report (SitRep).

ChatGPT-4o: Highest SitRep Completeness

ChatGPT-4o’s SitReps included all 12 required ICS sections (situation overview, operational period, logistics, etc.) in 94% of cases. It automatically populated the “current resource status” table from the live inventory feed. The PDC evaluators gave it a 4.6/5 for completeness. The downside: its SitReps averaged 1,400 words, longer than the 800-word target, slowing briefings.

Claude 3.5 Sonnet: Best for Executive Summaries

Claude produced concise SitReps averaging 760 words while still covering 11 of 12 ICS sections. It omitted the “weather forecast” section in 31% of reports — a gap that the PDC flagged as critical for typhoon-prone regions. When prompted to add weather, it complied in 1.2 seconds, suggesting the omission was a default-prompt issue, not a capability limit.

DeepSeek-V3: Fastest Generation but Missing Metadata

DeepSeek-V3 generated a SitRep in 22 seconds (ChatGPT: 41 seconds; Claude: 37 seconds). However, 18% of its reports lacked the “reporting unit ID” and “classification” fields — both required for inter-agency sharing. The PDC concluded that DeepSeek is best suited for internal, low-formality updates rather than official external communications.

Cost and Scalability: Per-Response Economics

Deploying AI assistants at scale requires understanding cost per query and throughput under load. The benchmark used 10,000 concurrent queries — simulating a city-wide evacuation — on identical cloud instances (AWS g5.12xlarge, 48 vCPU, 192 GB RAM).

ChatGPT-4o: Reliable but Expensive

ChatGPT-4o processed 10,000 queries in 14.3 minutes at a cost of $0.042 per query (API pricing, March 2025). Cost predictability was high: 95% of queries fell within ±8% of the average latency. This makes it suitable for well-funded government agencies but less so for cash-constrained NGOs.

DeepSeek-V3: Cheapest per Query

DeepSeek-V3 cost $0.009 per query — 4.7x cheaper than ChatGPT-4o — and completed the batch in 11.8 minutes. The trade-off: 8% of queries timed out (>30 seconds) under peak load, versus 2% for ChatGPT. For budget-limited field operations, DeepSeek offers the best cost-efficiency if timeout risk is acceptable.

Claude 3.5 Sonnet: Middle Ground

Claude cost $0.031 per query with a 13.1-minute completion time. Its timeout rate was 3.1%. The PDC noted that Claude’s higher accuracy on cross-source deduplication could offset its per-query cost premium by reducing the number of redundant queries by up to 18%.

Ethical and Privacy Constraints

Disaster data often includes personally identifiable information (PII) — names, phone numbers, medical conditions. The 2024 IEEE AI in Emergency Management Ethics Guidelines requires that any AI processing such data must strip PII before storage and allow full audit trails.

ChatGPT-4o: Strongest PII Redaction

ChatGPT-4o automatically redacted 97.3% of PII in a test set of 5,000 messages (IEEE 2024 benchmark), including partial matches like phone numbers embedded in free text. Its audit log recorded every redaction with a timestamp and the original token position. This is the highest compliance level among the tested assistants.

Gemini 1.5 Pro: Weakest Audit Trail

Gemini redacted 91.1% of PII but its audit log was opaque: it recorded “PII removed” without specifying which tokens were removed. In a legal review scenario — common after a disaster for liability assessment — this lack of granularity could fail evidentiary standards. The IEEE report flagged this as a “significant compliance risk.”

Grok-2: No Local PII Option

Grok-2 processes all queries through cloud servers that may store data for model training. In the IEEE test, 0% of PII was redacted because the model was not configured to do so. The vendor’s terms of service explicitly state that data may be used for “service improvement.” This makes Grok-2 unsuitable for any disaster response involving protected health information (PHI) or victim names.

FAQ

Q1: Can AI assistants replace human emergency managers during a disaster?

No. In the OCHA 2024 simulation, AI assistants processed data 40x faster than humans but made errors in 6-12% of cases. Human managers are still required to verify outputs, make ethical decisions (e.g., triage priority), and communicate with the public. The best current use is as a decision-support layer that handles volume while humans handle judgment.

Q2: Which AI assistant is best for a small NGO with a limited budget?

DeepSeek-V3 offers the lowest cost at $0.009 per query and the widest language support (12 languages). However, its 8% timeout rate and missing metadata fields mean it should be used for internal situational awareness, not external official reports. For a budget under $500 per month, DeepSeek is the most practical choice.

Q3: How do these models handle offline or low-connectivity environments?

None of the tested assistants run natively on local hardware. ChatGPT and Claude require a persistent internet connection. DeepSeek-V3 offers a lighter API with a 50% smaller model (67B vs. 175B parameters) that can run on a single A100 GPU if pre-loaded, but still needs periodic sync. The PDC 2025 drill showed that all models became unusable after 90 minutes without connectivity.

References

• World Health Organization. 2023. Emergency Response Framework, 3rd Edition.
• UN Office for the Coordination of Humanitarian Affairs (OCHA). 2024. AI in Humanitarian Information Management: Benchmark Report.
• Federal Emergency Management Agency (FEMA). 2024. Logistics Simulation (L-Sim 2024) Technical Summary.
• Journal of Emergency Management. 2024. Interoperability of AI-Translated ICS Messages: A Controlled Study, Vol. 22, Issue 3.
• Institute of Electrical and Electronics Engineers (IEEE). 2024. AI in Emergency Management Ethics Guidelines: PII Compliance Test.