Beyond the Fire Season: AI, Climate Extremes and the Future of Emergency Services in Australia

■ Australia no longer has a fire season and a flood season. It has a continuous operating environment of compound, back-to-back events. ■ AI can move emergency services from reactive response to predictive readiness across operations, safety, infrastructure and technology, but only on strong data foundations.

■ The biggest prize is integration: fire, Police, Ambulance and SES working from one picture, because the community sees one emergency. ■ Every AI capability expands the attack surface. Cyber security, data sovereignty and human accountability must be designed in from the start.

■ Operational capability is people, assets, facilities and supplies, ready together. Fleets, aircraft, facilities and inventory need the same intelligence-driven discipline as the response itself. ■ This edition adds the climate context, a plain-English guide to what AI can and can't do, generative AI, AI-specific cyber threats, an illustrative compound-event week, a staged roadmap and questions for boards.

Australia no longer has a "fire season" and a "flood season". It has a continuous operating environment. The Black Summer fires of 2019–20 burned more than 24 million hectares, claimed 33 lives and destroyed more than 3,000 homes.

Within two years, catastrophic flooding hit the Northern Rivers, Lismore and large parts of south-east Queensland. At Lismore, the Wilsons River peaked at 14.4 metres, more than two metres above the previous

record. Some of the same communities, and many of the same volunteers and crews, were in both. Since then, compound and back-to-back events have become the planning assumption, not the exception.

24M+ ha burned in Black Summer, 2019–20 14.4 m Wilsons River peak at Lismore, Feb 2022 2+ m above the previous Lismore flood record That change has big consequences for how our emergency services operate, protect their people, invest in infrastructure and use technology. Artificial intelligence is now part of that conversation. AI won't replace the firefighter on the fireground, the paramedic at the roadside or the SES volunteer on a roof at 2am.

Used well, it can give those people better information, earlier and more reliably, and connect agencies that still too often work in parallel instead of together. Fire moving through bushland in New South Wales as a firefighter watches the edge. Our crews face longer, more intense and more varied seasons.

Source: Photo by Darren Covington Here is where I see the main challenges and opportunities for fire services, and how they link with Police, Ambulance and the SES. The climate context The shift from seasons to a continuous operating environment is grounded in the climate record. The Bureau of Meteorology and CSIRO's State of the Climate 2024 report found that Australia has warmed by about 1.5°C since national records began in 1910.

The same reporting has documented an increase in extreme fire weather and a longer fire season across large parts of the country since the 1950s, and an increase in the intensity of short-duration heavy rainfall events. For emergency services, the implications are practical: ■ Fewer quiet periods. The window for planned burning, training, maintenance and rest between events is shrinking.

■ More compound events. Communities may face fire, flood and heat in the same year, sometimes in the same month. ■ More extreme individual events.

Floods above previous records and fires behaving in ways crews have rarely seen. ■ More pressure on people. The same career staff and volunteers are deployed more often, for longer, with less recovery time.

These trends are why the question is not whether emergency services should change how they operate, but how. What AI is, and what it isn't "Artificial intelligence" covers a range of technologies, and it helps to be specific. ■ Machine learning finds patterns in data and uses them to make predictions, such as how a fire might spread, where ambulance demand will rise, or which appliances are likely to fail.

■ Computer vision interprets images and video, such as spotting smoke in camera feeds, mapping flood extent from satellite imagery, or assessing building damage from aerial photos. ■ Natural language processing works with text and speech, such as transcribing and translating emergency calls or summarising field reports. ■ Generative AI creates new content, such as drafting messages, summaries or reports, based on patterns in large amounts of training data.

■ Optimisation tools recommend the best use of limited resources, such as where to stage appliances or how to roster crews. What AI isn't: it isn't a replacement for professional judgement, it doesn't understand context the way experienced people do, and it can be confidently wrong, especially in situations unlike those it learned from. Climate change makes that last point particularly important, because the events of the next decade may not look like those of the last.

Used well, AI is a tool that helps people see sooner, decide better and act faster. Used poorly, it can create false confidence and new risks. The rest of this article looks at where the benefits are greatest and how to manage the risks.

1. Operations: from reactive response to predictive readiness The challenge. Fire agencies are being asked to do more, more often, across a wider range of hazards.

Urban and rural fire services now routinely support flood rescue, storm response, hazardous materials incidents and extreme heat events. Fatigue builds up. Surge capacity, including interstate and international deployments, is stretched when several jurisdictions are under pressure at once.

Many of our volunteer-based models are under strain from changing demographics and declining availability. The opportunity. AI can move operations up the timeline, from response towards anticipation: ■ Predictive fire behaviour and spread modelling.

This builds on the Australian Fire Danger Rating System, introduced nationally in September 2022, and on simulators such as Spark and Phoenix. Machine learning can combine live weather, fuel moisture, terrain and satellite data to refresh spread predictions in minutes rather than hours.

■ Early detection. AI-enabled camera networks, satellite constellations and drone patrols can detect ignitions within minutes, especially lightning strikes in remote country, when a fire is still small enough to put out. ■ Flood and storm forecasting.

Hydrological models combined with AI-driven nowcasting can help the SES and fire services pre-position swiftwater rescue teams, sandbagging resources and evacuation support before rivers peak, not as they peak. ■ Resource optimisation. Decision support that recommends where to stage appliances, aircraft, ambulances and rescue crews based on the likely shape of an event over the next 24–72 hours.

The important principle is that AI informs the incident controller; it does not replace their judgement. The value is in reducing the time between seeing a risk and acting on it. How predictive readiness could work day to day Predictive readiness is less about a single model and more about a routine: ■ Each morning, forecasts, fuel moisture, river levels and recent activity are combined into a risk outlook for the next 24–72 hours.

■ Resources are matched to risk, with recommendations on where to stage crews, appliances, aircraft, swiftwater teams and ambulances. ■ Duty officers review and decide, adjusting recommendations based on local knowledge and priorities. ■ The outlook is shared with Police, Ambulance, SES, local government and utilities, so everyone prepares for the same picture.

■ After each event, predictions are compared with what actually happened, and the models improve. Getting started Agencies don't need to build everything at once. A practical first step is to bring existing forecasts, fire behaviour predictions and resource data into one daily product for duty officers, then add AI-based detection and optimisation as data foundations mature.

What changes for the incident controller For an incident controller, predictive readiness changes the shape of the day. Instead of spending the first hours of an event assembling a picture from phone calls, they start with an assessment of where the fire or flood is likely to go, what is at risk, and what resources are available and rested. Instead of reacting to each new report, they can test options against the forecast: what happens if we hold this line, or move these crews, or evacuate this town now rather than in three hours?

That doesn't make the job easier. The decisions are still hard, and the responsibility remains theirs. But it gives them more time to think, better information to think with, and earlier warning when the situation is about to change.

2. Safety: protecting the people who protect us The challenge. Our responders face more exposure for longer periods: extreme heat, smoke, contaminants, burnover risk, swiftwater, structural collapse and the cumulative psychological load of repeated disasters.

Firefighter cancer presumptive legislation across Australian jurisdictions recognises the long-term health toll. Mental health and moral injury after events like Black Summer are now widely acknowledged.

A NSW Rural Fire Service firefighter blacks out burnt ground after a fire. Exposure to heat, smoke and contaminants builds up over a career. Source: Photo by Darren Covington The opportunity.

■ Real-time crew tracking and situational awareness. Knowing where every crew and appliance is relative to the fire front or floodwater, with automated alerts when conditions change, directly addresses entrapment and burnover risk. ■ Wearables and physiological monitoring.

These can track heat stress, heart rate and exposure levels, so rotations are based on data rather than on how a crew member says they feel. ■ Hazard identification. Computer vision on drone and appliance cameras can flag structural instability, downed powerlines, hazardous materials and water hazards before crews commit.

■ Fatigue and wellbeing analytics. These look at deployment patterns across a season to identify people and teams at risk of burnout early. They must be handled ethically, with strong privacy protections and the trust of the workforce.

Safety technology only works if responders trust it. Co-design with operational crews, unions and volunteer associations isn't optional. It is what makes adoption possible.

A closer look at crew safety technology Crew safety technology works best as a connected system rather than a set of separate gadgets: ■ Location from vehicles, radios and personal devices shows where crews are. ■ Hazard prediction shows where the fire, flood or storm is likely to be. ■ Escape route analysis combines the two to warn when a crew's route to safety is at risk.

■ Physiological data shows when a crew member is approaching heat stress or exhaustion. ■ Communications monitoring shows when a crew has lost contact. Each element is useful on its own.

Together, they let supervisors and crews see danger developing and act before it arrives. The challenge is to present this information simply, without overwhelming crews who are already managing a dangerous job.

Mental health and wellbeing The cumulative psychological load of repeated disasters is one of the most serious long-term risks facing emergency services. AI and data can help, carefully. Deployment records can identify people who have had repeated exposure to traumatic events, prompting supportive check-ins rather than waiting for problems to surface.

Rostering tools can protect recovery time after difficult deployments. But this is sensitive territory. Any use of data for wellbeing must be voluntary where possible, transparent, governed by strict privacy rules, and designed with the workforce and its representatives.

Linking the agencies on safety Responder safety doesn't stop at agency boundaries. Paramedics working near a fire front, police managing roadblocks in smoke and SES volunteers in floodwater all face the same hazards. A shared safety picture, showing where every agency's people are relative to the hazard, would let any supervisor see when partner crews are at risk, and would reduce the chance of one agency's people being exposed because another's information didn't reach them.

Building it requires agreement on location data sharing, privacy protections and common alerting rules, but the benefit is a single safety net across the whole response. 3. Infrastructure: building for the event we haven't seen yet The challenge.

Climate extremes test our critical infrastructure as much as they test our people. Communication towers lose power. Fibre is cut by fire and flood.

Roads and bridges wash away. Fire stations, ambulance stations and emergency operations centres have sometimes been built in places that are now high-risk. The September 2025 outage, when Triple Zero calls failed across parts of South Australia, Western Australia and the Northern Territory, is a stark reminder that the whole emergency system depends on networks that can fail when they are needed most.

The opportunity. ■ Resilient, interoperable communications. This means progressing the national Public Safety Mobile Broadband program, along with satellite fallback, mesh networks and hardened radio infrastructure, so agencies can still talk to each other and to the community when commercial networks go down.

■ Digital twins of communities and critical assets. These model how a town, catchment or road network will behave under fire or flood scenarios. They can inform planning, evacuation routes, station placement and investment priorities.

■ Edge computing on the fireground. AI models running on appliances, drones and command vehicles keep working when connectivity is lost. ■ Data infrastructure as critical infrastructure.

Shared, secure, sovereign data platforms are the foundation every AI capability depends on. Without good data, AI is just expensive guesswork. Planning for climate-ready infrastructure Infrastructure decisions made today will last for decades, so they must be based on the climate of the future, not the past: ■ Station and facility locations should be assessed against future fire, flood and heat risk.

■ Communications sites need backup power, physical protection and alternative paths. ■ Critical systems should be hosted in resilient, geographically separated locations, with tested failover. ■ Fleets should be designed for multi-hazard roles, including flood access and extreme heat.

■ Evacuation routes and refuges should be reviewed against modelled future events, using digital twins where available. The September 2025 Triple Zero outage is a reminder that the emergency system depends on networks owned and operated by others. Agencies need clear arrangements with telecommunications providers and tested alternatives when those networks fail.

4. Assets: fleets, aircraft and facilities that are ready when needed The challenge. A response is only as good as the equipment that turns up.

Every prediction, warning and deployment decision assumes that the right appliance, aircraft or facility will be available when it is called on. On the worst days, that assumption is under the most strain. ■ Ground fleets.

Fire and emergency services run large, dispersed fleets of tankers, pumpers, aerial appliances, rescue vehicles and support vehicles. Many are housed in volunteer brigades far from a workshop, maintained by members who also have day jobs. Breakdowns during a response, or appliances off the road on a high-risk day, directly reduce the capability available to protect communities.

■ Aviation, fixed and rotary wing. Aerial firefighting relies on a mix of fixed-wing aircraft, from large air tankers to single-engine bombers and air attack platforms, and rotary-wing aircraft for water bombing, winching, mapping and crew transport. Much of this capability is contracted for the season through national arrangements, and longer, overlapping fire seasons put pressure on availability when several states, or several countries, need aircraft at once.

Maintenance windows, crew duty limits and weather all shape what is actually available on the day. ■ Facilities. Stations, depots, airbases, warehouses and emergency operations centres must keep working through heat, smoke, flood and power loss.

Some sit in places that are now high-risk, and many depend on networks and power supplies that fail during the very events they support. ■ Purchasing and inventory. New appliances and specialist equipment can take a long time to specify, build and commission.

Spare parts, breathing apparatus, PPE, hose, fuel, retardant, sandbags and consumables are often tracked in separate systems, or on paper, which makes shortages hard to see until they bite. The opportunity. AI and better data can bring the same predictive discipline to assets that this article proposes for operations: ■ Predictive maintenance for ground fleets.

Telematics on appliances and vehicles can flag pump, engine, brake and electrical problems before a breakdown. Maintenance can then be scheduled around the risk season, rather than in the middle of it, with support targeted at the volunteer brigades that need it most. ■ Aviation availability forecasting.

Combining aircraft maintenance records, crew duty hours, contract terms and weather gives air operations staff a forward view of which fixed-wing and rotary-wing aircraft will be serviceable, and where, over the coming days. That supports better pre-positioning and earlier requests for interstate or national support. ■ Facility condition and resilience.

Condition monitoring of critical systems such as power, water, fuel storage and communications, combined with climate-risk modelling, shows which stations, airbases and operations centres are most at risk, and where backup power, water and hardening are needed first. ■ Smarter purchasing. Whole-of-life cost modelling, demand forecasting and supplier risk analysis help agencies decide what to buy, when and from whom.

Common specifications across agencies make parts interchangeable and support joint purchasing.

■ Real-time inventory. Stock levels for critical items, shared across agencies and jurisdictions, with demand forecasts tied to the risk outlook, so cylinders are refilled, PPE is replaced and supplies are positioned before the season peaks. People, assets, facilities and supplies: ready together.

Source: Darren Covington Why this is operational, not administrative Asset availability, maintenance, purchasing and inventory are often treated as back-office functions. They are not. Operational capability is people, assets, facilities and supplies, ready together.

An intelligence-driven plan that cannot be resourced is just a forecast. Agencies that report readiness honestly, by capability and region, and that link it to the daily risk outlook, can make better decisions about pre-positioning, interstate requests and when to warn the community that response may be stretched. Securing the asset chain Fleet telematics, aviation maintenance systems, purchasing platforms and inventory systems are part of the operational chain, and they are targets.

A compromised maintenance system could take appliances off the road; a compromised supplier could disrupt critical stock. These systems need the same protection as dispatch: multi-factor authentication, segmentation from corporate IT, secure vendor access and tested manual fallbacks. 5.

Technology: the promise and the pitfalls The challenge. Emergency services technology in Australia is fragmented. There are multiple computer-aided dispatch systems, different data standards across states and agencies, legacy platforms and limited funding for long-term digital transformation.

Add AI to that environment without addressing the foundations and you risk faster confusion, not better decisions. There are also real risks to manage:

■ Model reliability under unprecedented conditions. Climate change means future events may fall outside the historical data AI models were trained on. ■ Over-reliance and automation bias.

Decision-makers may defer to a model when experience is telling them something different. ■ Misinformation. During disasters, AI-generated false content can spread faster than official warnings.

■ Equity. AI-driven warnings and services must reach everyone: remote and First Nations communities, culturally and linguistically diverse communities, older Australians and people with disability. The opportunity.

Approached with discipline, AI can make a real difference: ■ AI-assisted Triple Zero and dispatch. Call triage, automatic location and translation during surge events, and ambulance demand prediction free call-takers to focus on the most critical calls. ■ Public warnings at scale.

Tailored, multilingual and location-specific messaging aligned to the Australian Warning System, delivered through every available channel, now including AusAlert, the national cell broadcast system available from October 2026. ■ Intelligence fusion. Combining social media, sensor, satellite and field reports into one common operating picture, with clear provenance so commanders know what has been verified.

■ Post-event recovery. Automated damage assessment from aerial imagery can speed up insurance, recovery funding and rebuilding decisions. Empowering the frontline: integrated apps over satellite Most of the fastest, highest-stakes decisions in an emergency are made by crew leaders, strike team leaders and volunteers on the ground, often where there is no mobile coverage.

AI-supported insights are of little use to them if they stay in the operations centre. Frontline volunteers and staff need integrated, fit-for-purpose applications that put the same real-time picture in their hands: ■ One aligned toolset, not a dozen apps. Tasking and commander's intent, the shared operating picture filtered to the user's role, crew location, escape routes and safety alerts, fire spread and flood predictions, warnings, photo and voice reporting, availability and fatigue hours, and appliance and stock status, all in one integrated experience.

■ Fit for purpose on the fireground. Designed for gloves, glare, smoke and stress, with simple screens, large controls and voice input, and co-designed with the volunteers and staff who use them. ■ Connected anywhere.

Offline-first apps that sync over mobile networks when available, and over low-earth-orbit satellite terminals on appliances and command vehicles, or direct-to-device satellite messaging on phones, when they are not. ■ Radio stays primary for safety-critical voice. Apps add the shared picture and data; they don't replace the radio.

■ Secure by design. Managed devices, multi-factor sign-in, encryption, remote wipe for lost devices, and data held under Australian jurisdiction. A crew leader should be able to see the predicted fire run, their escape route, the intent from the incident controller and the status of the nearest tanker on one screen, and send back a photo and location in seconds, even in a valley with no mobile signal.

Generative AI in emergency services Generative AI tools are developing rapidly and are already being explored by emergency services around the world. Potential uses include: ■ Drafting warnings and public information from incident data, for review and approval by authorised staff. ■ Translating messages into community languages quickly, with checks by qualified translators where time allows.

■ Summarising large volumes of reports, such as field updates, social media and situation reports, so decision-makers can see the key points. ■ Supporting after-action reviews by organising records and highlighting patterns. ■ Answering routine questions from the public through well-governed digital assistants, freeing people for more complex calls.

The risks are significant. Generative AI can produce plausible but wrong information, which is dangerous in a warning. It can leak sensitive information if used carelessly.

And it can be manipulated through crafted inputs. Any use must keep a qualified person responsible for what is published, use approved and secure tools, and never put sensitive community or operational data into public AI services. “Add AI to fragmented systems without fixing the foundations and you get faster confusion, not better decisions.” Avoiding automation bias in practice Automation bias is easy to describe and hard to prevent.

Practical measures include: ■ Show uncertainty, not just a single answer, so users see when a model is unsure. ■ Ask for reasons, by requiring decision-makers to record why they accepted or rejected a recommendation. ■ Train with failures, including exercises where the model is deliberately wrong, so people learn to spot it.

■ Keep independent checks, such as field observations and local knowledge, that don't depend on the model.

6. Cyber security at the core Cyber security at the core: secure by design, not bolted on. Source: Darren Covington As emergency services become more digital, they become more attractive and more consequential targets.

An attack on dispatch, warnings or crew tracking during a major event could cost lives. A manipulated data feed could send crews to the wrong place. A fake warning could cause panic or false reassurance.

And AI introduces new risks of its own, including poisoned training data, manipulated inputs and leaks of sensitive information. Cyber security therefore has to be at the core of every AI and data initiative: ■ Secure by design. Security requirements are defined before systems are bought or built, and tested before they go live.

■ Zero trust across agencies. Every user, device and data feed is verified, with least-privilege access when fire, Police, Ambulance and SES share information. ■ Protect the data that matters most.

Information about vulnerable residents, workforce health and operational plans is encrypted, held under Australian jurisdiction, access-controlled and audited. ■ Assure the AI. Models are tested against manipulation and extreme scenarios, their outputs are explainable, and there is always a human accountable for the decision.

■ Authenticate public warnings. Only verified, authorised channels can issue warnings, and agencies actively counter fake ones. ■ Plan for the outage.

Every critical capability has a tested fallback when networks, systems or power fail, whatever the cause. The Australian Signals Directorate's Essential Eight, the Security of Critical Infrastructure framework and the Cyber Security Act 2024 give agencies a solid baseline to build on.

AI-specific threats explained AI systems introduce some threats that traditional security doesn't fully cover: ■ Data poisoning: an attacker corrupts the data a model learns from, so it makes systematically wrong predictions. ■ Adversarial inputs: carefully crafted inputs, such as altered images or sensor readings, that cause a model to misclassify what it sees. ■ Prompt injection: hidden instructions in text or documents that cause a generative AI system to behave in unintended ways.

■ Model theft and leakage: extracting a model, or the sensitive data it was trained on, through repeated queries. ■ Over-reliance: not an attack, but a vulnerability: people trusting a model's output without the checks they would apply to a human's. Defences include protecting training data and pipelines, validating inputs, testing models against manipulation, limiting what AI systems can do on their own, monitoring their behaviour in operation, and keeping people accountable for decisions.

A lightning storm over a city. Compound events demand systems that keep working when everything else is under pressure. Source: Photo by Tasos Mansour, Unsplash 7.

One emergency, many agencies: why integration is the real prize The most powerful opportunity isn't any single technology. It is connection. In a major flood or fire, the community doesn't see separate agencies.

It sees one emergency: ■ Fire services are fighting fire, conducting rescues and managing hazmat. ■ The SES is leading flood and storm response, with many of the same volunteers often serving in multiple roles. ■ Police are managing evacuations, road closures, missing persons and public order, and in many jurisdictions holding overall emergency coordination.

■ Ambulance services are triaging and treating, moving vulnerable patients and keeping health services running through the surge. The community sees one emergency. Our systems should too.

Source: Darren Covington Every royal commission and inquiry of the past two decades, from the 2009 Victorian Bushfires Royal Commission to the 2020 Royal Commission into National Natural Disaster Arrangements, has returned to the same themes: interoperability, shared situational awareness, consistent warnings and national coordination. AI and modern data platforms let us act on those findings in practical ways: a shared, secure common operating picture across fire, Police, Ambulance and SES; joint predictive modelling; vulnerable-persons data protected by strong privacy safeguards; and cross-jurisdictional resource sharing informed by national visibility of capability and demand, building on the work of AFAC and the National Emergency Management Agency.

Data sharing done right Integration depends on agencies sharing data, and data sharing depends on trust. Good practice includes: ■ Clear purpose: each data-sharing arrangement states what is shared and why. ■ Minimum necessary: only the data needed for the purpose is shared.

■ Legal authority: sharing is grounded in legislation, agreements or consent. ■ Security: shared data is protected to the standard of the most sensitive party. ■ Accountability: access is logged and reviewed, and misuse has consequences.

■ Review: arrangements are reviewed regularly and after major events. Illustrative scenario: a compound week The following scenario is illustrative and set a few years in the future.

Day 1. A heatwave is forecast, followed by severe storms. The integrated risk outlook flags three concerns: high fire danger in the ranges, heat stress in older residents, and flash flooding on recent burn scars.

Fire, Police, Ambulance, SES, health and local councils receive the same outlook. Day 2. AI camera networks detect smoke from two ignitions within minutes.

Crews are dispatched while the fires are small. Ambulance demand modelling predicts a rise in heat-related calls, and additional crews are rostered. Day 3.

A power outage affects thousands of homes. Ambulance and health services identify residents dependent on powered medical equipment, using protected data, and check on them. A spoofed message claiming an evacuation order circulates on social media; it is identified and corrected within the hour through official channels.

Day 4. Storms bring intense rainfall to the burn scars. Pre-positioned SES teams respond to flash flooding, and warnings tailored to the affected streets reach residents through AusAlert.

Day 7. Automated damage assessment from aerial imagery speeds up recovery support. The after-action review uses data from all agencies to improve the next outlook.

Equity: making sure AI protects everyone AI systems learn from data, and data is not evenly spread. Remote communities may have fewer cameras and sensors. Culturally and linguistically diverse communities may be under-represented in social media monitoring.

People without smartphones may miss app-based warnings. Older residents and people with disability may need different kinds of support. If these gaps aren't addressed, AI could improve protection for well-connected communities while leaving others behind.

Practical steps: ■ Check coverage: map where sensing, data and warnings are weakest, and prioritise investment there. ■ Test for bias: evaluate models on how they perform for different communities and places, not just on average. ■ Use many channels: combine apps and cell broadcast with radio, door-knocking, community networks and accessible formats.

■ Work with communities: involve First Nations communities, multicultural organisations, disability advocates and remote communities in design and testing. Volunteers and AI For volunteers, AI should make involvement easier, not harder. Promising uses include automatic availability matching, simpler incident reporting by voice, better information sent to phones on the way to an incident, and tools that help brigade leaders plan training and maintenance.

Volunteers should also be part of the conversation about AI's limits and risks, since they often know their local area better than any model. Building the skills Using AI well requires new skills across the organisation: ■ Executives and boards need enough understanding to govern AI risk and ask the right questions.

■ Incident managers need to know how to interpret model outputs, including confidence levels, and when to challenge them. ■ Analysts and data specialists need skills in data engineering, modelling and evaluation. ■ Frontline crews need to know what AI-supported tools can and can't do.

■ Security teams need to understand AI-specific threats. Partnerships with universities, research organisations, the Bureau of Meteorology and industry can help build these skills faster than any agency could alone. Buying AI well Much of the AI emergency services use will be bought rather than built.

Procurement should ask: ■ What data was the model trained on, and how does it perform in Australian conditions and extreme events? ■ Can we see how the model reaches its outputs, and how confident it is? ■ Where is our data stored and processed, and who can access it?

■ How is the model protected against manipulation, and how will we be told about vulnerabilities? ■ Can we get our data out, and switch providers, if we need to? ■ What happens when the model is wrong, and who is accountable?

The case for national coordination Many of the AI capabilities described here would be expensive and wasteful to build eight times over. Detection networks, satellite analysis, national resource visibility, shared models of fire and flood behaviour, and secure data platforms all benefit from scale. A model trained on fires across the whole country will generally be more robust than one trained on a single state's history.

National coordination could include: ■ Shared detection and satellite services feeding every jurisdiction. ■ Common data standards for incidents, resources and warnings. ■ Shared model development and testing, with results available to all agencies.

■ National assurance of AI systems used in emergency management, so each agency doesn't have to test everything itself. ■ Shared threat intelligence on cyber and AI-specific attacks against emergency services. This builds on the national coordination already provided by NEMA, AFAC, NAFC and the Bureau of Meteorology, and on the national approach taken with the fire danger rating system and warnings.

Ethics and public trust Emergency services enjoy high levels of public trust. That trust is an operational asset: it is why people follow warnings and evacuation advice. AI can strengthen it, through faster, more accurate and more personalised warnings, or damage it, through errors, privacy breaches or decisions people don't understand.

To protect trust, agencies should be open about how they use AI, explain decisions that affect people, give people ways to raise concerns, protect personal information rigorously, and acknowledge mistakes quickly. A published statement of AI principles, developed with the community and the workforce, is a simple and powerful first step. Recovery: where AI already shows its value Recovery is often where AI's value is easiest to see.

After a fire or flood, AI analysis of aerial and satellite imagery can identify damaged buildings across large areas quickly, helping governments and insurers direct support sooner. Natural language tools can help process large volumes of applications for assistance. And data from the response can guide decisions about rebuilding in safer ways.

These uses carry lower operational risk than real-time decisions, making them a good place to build experience and confidence. Where to from here? My view is that the next decade will separate emergency services that treat AI as a series of pilots from those that treat it as part of how they operate.

A few principles I'd put forward: 1 Start with the problem, not the technology. Pick the operational decisions that matter most and work backwards. 2 Fix the data foundations.

Interoperable, trusted, secure data is the precondition for everything else. 3 Put cyber security at the core. Design it in, fund it properly and govern it at executive level.

4 Keep humans accountable. AI should advise, and people should decide. 5 Co-design with the frontline, including career staff, volunteers, unions and communities.

6 Invest nationally and integrate locally. Climate extremes don't respect state borders, and neither should our capability. Australia's emergency services are among the most respected in the world.

Their professionalism, and the commitment of tens of thousands of volunteers, is a national asset. The question isn't whether climate extremes will keep testing them; they will. The question is whether we give them the tools, infrastructure and secure, integrated systems to meet that test safely and effectively.

A staged roadmap First 90 days ■ Map current data sources, systems and AI pilots across the agency and its partners. ■ Agree an AI policy consistent with national guidance, including where AI must not be used. ■ Review cyber security for the systems that matter most during emergencies.

■ Baseline the real availability of appliances, aircraft and critical facilities on high-risk days. First year ■ Produce a daily integrated risk outlook shared with partner agencies. ■ Pilot AI-assisted detection and incident fusion in one region.

■ Introduce fatigue-aware rostering and crew safety monitoring trials, co-designed with crews. ■ Close the highest-risk gaps against the Essential Eight.

■ Fit telematics to priority appliances and build a shared view of critical inventory. ■ Pilot an integrated frontline app for crews, with satellite terminals on priority appliances. Years two to three ■ Scale proven pilots across the state.

■ Integrate fire, Police, Ambulance and SES data into a secure common operating picture. ■ Introduce generative AI for drafting and translation, with strict human review. ■ Test AI systems against manipulation and extreme scenarios before operational use.

■ Introduce predictive maintenance, aviation availability forecasting and whole-of-life purchasing. Measuring success ■ Earlier detection: time from ignition to confirmed detection. ■ Earlier action: time from risk identification to resource deployment and warnings.

■ Crew safety: injuries, near misses and fatigue exceedances. ■ Warning reach: share of people at risk receiving timely, understandable warnings. ■ Integration: agencies sharing data in the common operating picture.

■ Operational capability: availability of appliances, aircraft and critical facilities on high-risk days, and critical stock held against target. ■ Security: Essential Eight maturity and AI assurance testing completed. Questions for boards and executives ■ Where could AI make the biggest difference to our community's safety, and where are we not ready?

■ What is our policy on generative AI, and who is accountable for AI-assisted decisions? ■ How would we detect if a model or data feed we rely on had been manipulated? ■ How are we protecting vulnerable residents' data while using it to keep them safe?

■ How are crews, volunteers and unions shaping how AI is used? ■ On our worst day, how many of our appliances, aircraft and facilities would actually be available, and how do we know? Glossary ■ Computer vision: AI that interprets images and video.

■ Data poisoning: corrupting the data a model learns from. ■ Generative AI: AI that creates new content, such as text or images. ■ Machine learning: AI that learns patterns from data to make predictions.

■ Operational capability: the people, assets, facilities and supplies available to respond, ready together.

■ Nowcasting: very short-term forecasting, typically over the next few hours. ■ Prompt injection: hidden instructions that cause a generative AI system to behave unexpectedly. Further reading ■ Bureau of Meteorology and CSIRO, State of the Climate 2024.

■ Royal Commission into National Natural Disaster Arrangements, Report (2020). ■ Australian Government, Voluntary AI Safety Standard (2024). ■ Digital Transformation Agency, Policy for the responsible use of AI in government (2024).

■ Australian Signals Directorate, Essential Eight Maturity Model. Where are you seeing AI make a real difference, and where are the barriers still too high? #EmergencyManagement #AI #CyberSecurity #ClimateResilience #Bushfire #Floods #FireServices #SES #PublicSafety #DigitalTransformation #Australia #ResponsibleAI Darren Covington writes on intelligence-driven emergency services, climate resilience, technology and cyber security in Australia.

Photo credits: Darren Covington; Tasos Mansour / Unsplash. Unsplash photos are used under the Unsplash Licence. Other photos and graphics © Darren Covington.