On 31 March 2026, the Scientific Advice Mechanism (SAM) and Public Safety Communication Europe (PSCE) co-hosted a webinar bringing together scientific experts and emergency management practitioners to explore how artificial intelligence is reshaping crisis response across Europe. The session drew on our recently published rapid evidence review and featured real-world case studies from Germany and Portugal.
With: Simon Franke - ICT Services of the German Red Cross; Jorge Gomes - Operations Coordinator at VOST Europe; Thomas Kox - SAPEA Working Group Member on 'AI for Crisis Management'; Marta López Saavedra - Postdoctoral researcher at IDAEA-CSIC
SAPEA report: principles over tools
Thomas Kox introduced the SAPEA evidence review report on AI for crisis management, developed to provide the European Commission with an independent, synthesis-level view of the field. Rather than evaluating specific tools, a task rendered impractical by the pace of technological change, the report focuses on underlying principles that should govern AI deployment throughout the full crisis management cycle, from risk monitoring and early warning through to response and recovery.
The technologies in scope are deliberately broad: crowdsourcing platforms, conversational systems, training simulations, and automatic analysis tools all fall within the review's remit. The report addresses performance requirements, legislative frameworks including the EU AI Act and GDPR, ethical principles, data challenges, and the question of trust, illustrated through four detailed case studies.
From the dispatch room: AI in German emergency services
Simon Franke of the German Red Cross ICT Services provided a window into the operational pressures facing emergency dispatch centres in Germany. Dispatch operators already work under severe time pressure, and the challenge is intensifying: rising call volumes, staff shortages, tighter budgets, and an expanding array of communication channels (video calls, wearables, e-calls, social media) are all converging simultaneously.
Traditional rule-based systems were not designed to handle this complexity. The German Red Cross has already deployed a real-time translation tool enabling dispatchers to communicate across language barriers during emergencies, with question suggestions provided alongside transcription. Research projects are advancing further, exploring image-based hazard detection, smoke propagation modelling, medical triage support, and optimised ambulance routing.
Franke's long-term vision points toward human-AI collaboration built around multi-agent architectures, where AI assistants operate autonomously in the background while a human operator retains the role of decision-maker. This framing raises important open questions: how should AI autonomy be bounded, what does effective human-AI team design look like, and how is accountability maintained when decisions are shared?
Digital volunteers and open-source intelligence: the VOST Portugal model
Jorge Gomes, Operations Coordinator at VOST Europe, presented a different but complementary approach built around what he described as a "human in the middle" architecture. In this model, AI ranks and suggests; humans validate and act.
VOST Portugal has deployed several live AI systems:
- A satellite monitoring tool using Meteosat third-generation data detects fire radiative power at ten-minute intervals, cross-referencing satellite readings with ground data to confirm active fires and identifyre-ignitions that would be invisible to human observers.
- A social media intelligence system, developed in collaboration with the Joint Research Centre, extracts multilingual help requests from open platforms. During the 2023 Turkey earthquake, it mapped 42 such requests, directly contributing to 21 rescues.
- A predictive fire tool, Pyrocast, integrates fuel accumulation and meteorological data to project fire threat levels and spread rates across a seven-day horizon.
Yet Gomes was candid about the fragility of this infrastructure. Platform access has become an acute problem: the cost of the Twitter/X API rose from zero to €5,000 per month, and Meta's discontinuation of CrowdTangle rendered previously trained social media intelligence models unusable. A digital divide means communities without reliable internet access or digital literacy cannot benefit from SMS- or web-based warning systems.
Gomes also highlighted that data divide across Europe remains stark. As such, Portugal publishes real-time structured data on fire incidents, while equivalent data from Belgium and Germany is unavailable or arrives in non-machine-readable formats. Legislation such as the Digital Services Act and the AI Act exists, Gomes observed, but enforcement has not kept pace.
Most immediately, eight million trees were felled in recent Portuguese storms, creating fire conditions with no historical precedent in the training data, a reminder that AI systems built on past patterns can be overtaken by unprecedented climate-driven change.
Panel discussion: accountability, bias, and the limits of automation
The panel discussion brought together all three speakers — Franke, Gomes, Kox, and postdoctoral researcher Marta López Saavedra of IDAEA-CSIC — and surfaced several themes cutting across allpresentations.
On responsibility, there was clear consensus: AI will become pervasive in crisis management, but accountability cannot be delegated to an algorithm. Kox described the need for four distinct layers: those who build models, organisations that deploy them, operators who use them, and systems that record outcomes and feed back into future decisions. Transparency, the panel agreed, must be operational rather than rhetorical. Logging every decision so that it can be reviewed is more meaningful than philosophical commitments to openness.
On data protection, the panel discussed how GDPR permits data use for life-protecting purposes provided core principles are respected. The harder challenge is sharing data for model training while preserving anonymisation. The EU AI Act shifts a significant share of responsibility onto AI service providers, requiring transparent documentation of training data and formal risk assessments. Call-taking is explicitly classified as a high-risk application under the Act.
On bias, the panel raised the problem of models trained predominantly on anglo-centric data, exhibiting cultural blind spots. Addressing this requires not only acknowledging bias but sharing training data and outcomes openly.
On automation, the panel warned against what is sometimes called automation bias: the tendency for operators to approve AI suggestions routinely, without critical review. The remedy, participants agreed, is explainability: operators need to understand why a system makes a recommendation, not simply receive it. Black-box outputs in life-affecting decisions are not acceptable.
The discussion also yielded concrete guidance for practitioners considering deployment. Systems should be validated in controlled environments before going live and built on infrastructure robust enough to remain available during power outages and network disruptions. Equally important is involving operators from the very start of the development process to ensure that tools are shaped by genuine operational needs rather than by what happens to be technically feasible.
The webinar offered a productive dialogue between the evidence base SAPEA has assembled and the lived experience of practitioners working at the intersection of technology and emergency response. Both perspectives point in the same direction: AI in crisis management holds real potential, but realising it safely requires human oversight, open data infrastructure, enforceable regulatory frameworks, and tools that are designed with and for the people who will use them.
Read the Rapid Evidence Review Report.
