Tools built on artificial intelligence offer a mind-boggling array of possibilities for improving public safety, from combating wildfires to reviewing building plans and finding operational efficiencies for fire departments. But as advocates tout the promise of AI, others urge caution in the development of AI tools and their applications. How these technologies will impact public safety raises complex questions that researchers and developers are just beginning to answer.
BY JESSE ROMAN
Last July, a small wildfire ignited in the middle of the night deep in a rugged expanse of the Cleveland National Forest, an hour’s drive east of San Diego, California.
Even for a trained spotter perched high above the valley, the small column of smoke would have been difficult to see rising against the vast night sky. But the machines were watching.
Soon after the ignition, the smoke was detected by a camera loaded with an artificial intelligence program trained to constantly scan the landscape, as far as 110 miles out, and identify minute visual changes that might suggest a fire had started. Once alerted by the camera, Cal Fire dispatchers quickly deployed resources, and “were able to hold the fire to a 10-by-10-[foot] spot out in the middle of the forest,” Phillip SeLegue, CalFire’s staff chief for fire intelligence, told CNN in September. If the AI camera hadn’t alerted the agency of the developing blaze, “by the next morning, that fire would have been a fire of significance,” SeLegue predicted.
From July to September last year, the 1,000 or so AI cameras positioned across California detected 40 percent of all wildfires in the state before a 911 call was received. SeLegue called the technology “a game changer” that allows responders to snuff out fires more quickly before they grow out of control.
The AI camera system being piloted in California is just one of myriad examples of how artificial intelligence seems poised to transform public safety across the entire spectrum of the NFPA Fire & Life Safety Ecosystem®.
As computing power increases, and as breakthroughs continue to lift the ceiling on the ability of machines to learn and make sense of the vast quantities of data we produce, the possibilities for what these tools can do in the life safety realm are almost limitless.
AI CASE STUDY
Intelligent Performance Based Design
When designing complex novel buildings such as a modern airport or a mall with a large atrium, fire protection engineers have long relied on computational fluid dynamics (CFD) simulations to predict how fire and smoke will spread through the structure. Such tests are necessary for designing fire protection strategies for projects relying on performance-based design, where the parameters of the building fall outside the scope and guidance of traditional codes and standards. The problem, experts say, is that these large CFD simulations are time-consuming, costly, and take a lot of computational time, making it difficult for a reviewer to check several different designs for optimal performance. To solve that problem, artificial intelligence researchers are developing new systems run by a trained AI program that can do the same calculations and modelling in a fraction of the time. What’s more, instead of relying on a static model in the manner of traditional simulations, the AI system is able to learn from the experiences and data generated from each new test, thereby improving its accuracy over time. Once the models are trained, an engineer would only need to input parameters such as building dimensions, fire size, and fuel type, and will get results almost instantly.
A quick scan of projects in development suggests that no profession will remain untouched. According to the researchers and companies developing these tools, AI software may soon enable building officials and authorities having jurisdiction (AHJs) to instantaneously review building plans and sprinkler designs for code compliance with the click of a button. Architects and fire protection engineers may have access to AI fire modelling tools that will help them design safer buildings that perform better in a fire. Public safety educators could have tools that can identify the demographics to target, as well as the optimal initiatives to achieve the greatest impact. Building inspectors, system maintenance professionals, event safety managers, and other professionals will all likely be able to turn to AI for insights to address challenges and complete tasks with greater efficiency.
So far, however, perhaps the most fantastical AI ideas are geared toward the fire service. One AI tool now being developed at the National Institute for Standards and Technology (NIST) can use visual cues to predict a fire’s transition to flashover and warn firefighters in time to get out of a building—in tests, the model correctly predicted flashovers one minute beforehand in 86 percent of the simulated fires. AI-trained computers can now evaluate a photograph of smoke and, with reasonable accuracy, determine what type of fuel is burning and the heat release rate, and may one day predict how the fire might grow.
Wearable AI devices, trained by listening to thousands of hours of human heartbeats, may soon warn firefighters when their hearts are stressed by heat or activity and tell them to take a break or seek medical care. Some observers believe that, given enough training and inputs,
AI could one day aid incident commanders at the scene of a fire by reading live building sensor data and—based on the lessons from thousands of pages of post-incident reports, meetings, trainings, and fire science manuals—provide instant guidance on the optimal next moves.
Although tantalizing, most of these concepts remain aspirational, though several AI tools spanning a range of life safety industries could launch as early as this year. In the fire service, some larger and more progressive fire departments are already on the verge of incorporating AI into their daily tasks.
“Like many departments, we have a finite amount of resources but an ever-expanding amount of work we have to accomplish—I think by using these tools we can gain a lot of efficiencies,” said David Povlitz, the fire chief at the Arlington County (Virginia) Fire Department, which for at least the past year has been carefully assembling and testing the infrastructure and data sources needed to run AI applications.
At first, AI programs in Arlington and elsewhere will likely be used primarily to help fire departments manage time-consuming administrative work, like writing a new paid-time-off policy or completing the first draft of a new standard operating procedure, Povlitz said. Soon, programs could help departments with other administrative functions including budgeting, staffing, and scheduling. The hope is the tools will help free up staff time to do more hands-on work such as training, community outreach, and community risk reduction.
Eventually, departments may incorporate AI programs into more complex tasks, such as fireground operations and improving safety for firefighters by tracking where they are inside a burning building. Those types of applications are “still a little bit of a dream at this point, but it feels like they could be within our grasp in three to five years,” Povlitz said.
As bullish as some are on AI’s potential to positively impact the fire rotection industry, others urge caution. Even innovative AI programs, like the chatbot. ChatGPT 4, are known to sometimes “hallucinate" facts and will occasionally present wildly inaccurate information with conviction. When the stakes are high, what assurances do we have that AI will be sophisticated enough to accurately interpret subtle nuances in the codes, or careful enough to provide reliable input during a complex fire incident?
“Like many fire departments, we have a finite amount of resources but an everex panding amount of work. AI tools can gain us a lot of efficiencies”
“We need to pay attention to AI and identify the places where we can utilize it to help achieve better outcomes, but we should also be careful not to get too enamored too fast,” said Jim Pauley, president and CEO of NFPA. “AI can help with many tasks, but in the spaces that we serve as NFPA, the human element is still very important. I’m not sure that will ever change.”
How quickly fire departments, code enforcers, building designers, and other stakeholders actually adopt these tools into their operations remains to be seen. As with everything AI, however, progress tends to move faster than anyone expects.
MACHINE EFFICIENCY, HUMAN ELEMENT
The AI industry’s ascent has been meteoric since the viral launch of ChatGPT in November 2022—the month it seems that the entire world collectively woke up to the game-changing potential of artificial intelligence.
Within two months of launch, ChatGPT grew from zero to more than 100 million users and spawned an arms race of sorts, with giants like Meta, Google, Apple, and others suddenly funneling billions of dollars into AI research projects.
All of that activity led Sam Altman, CEO of the artificial intelligence company OpenAI, which developed ChatGPT, to predict at a recent industry event that AI in 2024 will take “a leap forward that no one expected.” A Google executive sitting next to him chimed in, “plus one to that.”
In the simplest terms, artificial intelligence is achieved through what’s called machine learning: that is, by feeding massive amounts of data—visual, audio, hard numbers, text, even codes and standards—into powerful computer algorithms that continually analyze the information to identify meaningful correlations, patterns, and insights. The more data that’s processed, the more the program can learn, with an increased likelihood that the results will be more accurate and applicable.
The California wildfire AI cameras, for instance, are trained to detect signs of wildfire on the horizon by continuously comparing one image of the landscape to the next, taken two minutes apart. When an anomaly is detected, a human checks to see if it is a fire; if it’s something harmless, like a deer wandering up a mountain or a patch of fog rolling in, the system is told to disregard it.
Over time, this training improves the machine’s abilities to spot actual smoke and fire so that its discerning eye far outperforms trained human spotters. With a sufficient amount of high-quality data, a similar process can be achieved in almost any task that a data scientist can dream up.
One reason public safety appears to be such a ripe environment for AI is the industry’s access to copious amounts of data. Public safety agencies, along with adjacent industries such as insurers, fire research organizations, and fire protection device manufacturers, have been collecting robust data for decades and using it to inform their operations.
Now, instead of humans and simple spreadsheets analyzing the data for insights, intelligent machines can ingest entire libraries of information in seconds and search for metaphoric columns of smoke that no human is likely to spot.
What kinds of novel insights andefficiencies this will produce remains to be seen, but many of those working on AI in the fire and life safety world expect AI’s impact on their industries to be profound.
Xinyan Huang, an AI researcher at the Research Centre for Fire Safety Engineering at Hong Kong Polytechnic University, thinks that once the AI spigot is turned on, the work performed by professionals in these fields will be fundamentally altered forever. “We are still in the very early stages of this technology, but when the moment comes, I think every company will adopt AI,” he said. “I believe it will be widely used by every fire protection engineer.”
AI CASE STUDY
Design Toolbox for Buildings
Several AI-related projects in the works aim to use AI to make it easier to design better fire protection systems for buildings. For instance, one AI program is being developed to instantly calculate the ceiling heights and slopes in an entire building and estimate the optimal location for each sprinkler and smoke detector to maximize effectiveness in the event of a fire. Another AI program under consideration is akin to a toolbox for sprinkler design; by training it on various codes and standards, the program may soon be capable of telling a sprinkler designer where exactly in a building to place each sprinkler to achieve compliance and optimal performance. Still another project, initiated by the Fire Protection Research Foundation, has trained AI cameras to go through a building and use visual data to identify various objects and calculate the total fuel load of the building and its contents. Such measurements are critical in performance-based designs for ensuring adequate fire protection. In one early test, the AI surveyed three office buildings and found that “the measured total fuel load densities…were considerably larger than values found in older surveys and most code provisions,” according to the project’s final report. Future variations of the tool may soon be used to improve fuel load requirements in NFPA 557, Standard for Determination of Fire Loads for Use in Structural Fire Protection Design.
One major way that AI will be used in fire protection, Xinyan believes, is as a copilot—an assistant with extensive training in codes and standards and design, and one that can also instantly run fire dynamics models or read an entire layout of a building and tell you the optimal locations for sprinklers and smoke detectors.
“When engineers are doing a design, they can ask it questions like, ‘What can I do when I meet this particular problem?’” Xinyan said. “Or maybe the AI can sense what step an engineer is doing and provide some hints about what they should do next and what kinds of common errors they may face.”
These kinds of robot AI assistants may be closer to reality than most of us realize. As soon as this year, tools using AI will be available on the market that claim to be able scan a PDF of a building plan and perform a code compliance review within seconds. If the program finds something that’s either not permitted or that may be questionable—such as a room that is further than the allowed distance to the nearest exit—it flags it for the fire protection engineer to review, similar to the spell check function in word processing software.
Developers say the increased efficiency of these tools could be a boost for city building departments faced with staffing shortages and significant backlogs. For instance, a San Francisco Chronicle analysis from 2022 found that it takes an average of 627 calendar days for an applicant in the city of San Francisco to obtain a full building permit to construct a multifamily housing project. It takes an average of 861 days to gain the same approval for a single-family residence, the analysis found.
Developers claim that AI-based code-compliance tools could significantly reduce that time by saving humans from having to conduct the necessary initial compliance checks, which can take hours to complete.
Concepts like these are intriguing, but only work if a human is there to double check the machine’s accuracy and provide a final assessment.
“As bullish as some are on AI’s potential, others urge caution. Even innovative AI programs are known to sometimes “hallucinate” facts and will on occasion present wildly inaccurate information with conviction.”
How far beyond that a machine might someday be able to go in interpreting and applying the code is difficult to say, but the implications become complicated very quickly.
For instance, if an AI program is used to walk an electrician through a complex installation to comply with the National Electrical Code®, what happens if the program gets it wrong? Was the model trained on the right edition of the code? Did it have all of the context of the installation? Is the program, the developer, or the installer at fault?
“When it comes to applying the code, there are so many situations that are unique and that require you to understand the particulars in order to do it correctly,” said Pauley, who worked extensively with the NEC before coming to NFPA.
“If you ask an AI tool a code question, such as ‘Can I do X, Y, or Z?,’ the fear is that it will wrongly say ‘yes’ because it doesn’t have an awareness of all of the circumstances. So before we get too far down the road, we need to better understand what AI can do and what it can’t. AI may be able to fill in some pieces, but we need to preserve the human element in this.”
HARD QUESTIONS
That sentiment is shared by experts across a range of fields in the fire and life safety world. While it’s easy to see the utility of an ultra-smart AI assistant and imagine the ways these tools might make our lives easier—and potentially make our world safer— there are also risks and potential downsides to these developments.
If AI continues to progress as rapidly as some experts, including Sam Altman, predict, the fire protection community will inevitably run up against some difficult questions.
It is conceivable, for instance, that AI will someday be relied upon to do most of the heavy lifting, from designing safer versions of complex buildings like malls or airports to inspecting a hospital for code compliance, or even running response operations on a complex fire. But ultimately, how much control over these important—sometimes life-ordeath—tasks and decisions should be entrusted to a machine? What guardrails need to be in place to make sure the technology doesn’t lead us astray?
Preet Bassi, the CEO of the Centers for Public Safety Excellence, has thought a lot about what could go wrong if public safety agencies fail to take proper precautions with AI. In December, she moderated a panel, titled “Safely Harnessing the Potential of AI for Public Safety,” at the Technology Summit International held by the International Association of Fire Chiefs. The panelists included researchers and fire department representatives from the U.S. and the Netherlands, as well as Microsoft, the largest investor in OpenAI.
During the standing-room-only session, Bassi electronically polled the audience of mostly fire service leaders on what they see as the biggest risk of AI in the industry. The most common response was the potential lack of validity and reliability of AI tools.
In an interview the day after the panel discussion, Bassi shared two scenarios to illustrate her point: the Ghost Ship fire in Oakland, California, in 2016, and the West Fertilizer Company fire and explosion in West, Texas, in 2013. The Ghost Ship fire killed 36 people inside a warehouse that had been illegally converted into a residential artists’ collective and performance space. In the Texas example, 15 people were killed—including 12 firefighters at the scene— and more than 160 were injured when a storage facility filled with ammonium nitrate caught fire and exploded near a residential neighborhood.
“In both of those incidents, there was poor community data—the fire department in Oakland knew the Ghost Ship was an abandoned building, but because of a lack of inspections they had little concept that it was being used as an underground artists’ commune. In West, Texas, they knew that it was a commercial property, but they did not understand the level of hazard in that property,” Bassi explained. “An AI tool to predict a go/no go would have been built on the data that existed. But without having inspection activities to drive your community risk assessment, as in Oakland, or more sophisticated classifications of the types of commercial properties that are out there, as in West, Texas, the outcomes frankly would have been the same whether it was an AI tool or a human driving that decision.”
In some scenarios, it’s conceivable that an AI assistant working with faulty data could make an incident more dangerous by drawing the wrong conclusions. If an AI tool has determined, for instance, that a warehouse is abandoned, it could recommend an extreme defensive strategy for firefighters even though people are trapped in the building. Conversely, poor inspection data could lead to an AI tool falsely informing firefighters that there is no threat even when they are in harm’s way.
The point is that without robust data from inspections, an AI system could make potentially deadly tactical errors based on faulty assessments from poor data, situations that might otherwise have been avoided had a human made a different decision based on observation and experience. One fear is that “an automation bias might kick in, where you just begin to accept what the machine is telling you,” regardless if the information is credible, Bassi said.
The dangers of AI and poor data can manifest in other ways as well. What if a fire department is presented with a novel situation where multiple outliers coalesce to create a disaster without precedent?
Could the AI tool, using existing data or even information gleaned from textbooks, provide an incident commander with useful guidance? “AI is planning for the bell curve. It’s planning for the middle 66 percent of the situation, but our emergency services live at the fringes,” Bassi said, paraphrasing a panel member from the December event. “How does a model built to take away the mundane tasks and address the everyday begin to address what emergency services actually does?”
YAPAY ZEKA ÖRNEK ÇALIŞMASI
Yangın Söndürme İçin Dijital İkizler
Bazı uzmanlar, sensörlerin ve yapay zekanın birleşimini kullanarak, ileride yangın departmanlarının dijital ikiz adı verilen bir yapı, tünel veya başka bir yapı üzerinde gerçek zamanlı ve ölçeklendirilmiş dijital bir temsil kullanarak uzaktan yangın müdahalesi yönetebileceğini düşünüyor. Yangın senaryosunda, yangın komutanları bilgisayar ekranında yangının gerçek zamanlı konumunu görebilmek için bir yangının dijital ikizini oluşturabilirler. Bu, yangının binada nasıl büyüdüğünü ve yayıldığını gösteren noktaları içerebilir. Gerçek itfaiyecilerin veya yangın söndürme robotlarının yangını söndürmek için çalıştıkları yerleri temsil eden noktalar da olabilir. Ekran üzerinde, sprinkler veya yangın pompalarının durumu ve konumu, çıkış yerleri, sıcaklık okumaları, hava hızı ve yönü, yayaların hareketi veya dakika dakika koşulların nasıl değişebileceğine dair tahminsel analizler gibi her türlü faydalı bilgi bulunabilir. Yangın komutanları, bir düğmeye tıklayarak yapı içinde bir tavan penceresini açarak duman tahliyesi gibi müdahalelerin bazılarını kontrol edebilirler. Uzmanlara göre, böyle bir teknoloji için çok daha fazla sensör, akıllı bina teknolojileri ve gerçek zamanlı veri transferi gereklidir ve bu teknolojinin yaygın olarak kullanılmaya başlanması yıllar alabilir. Yine de, dijital ikizler, üretim hatları gibi alanlarda uygulama bulmaya başladılar. Burada bir çalışan, bir bilgisayar başında Sim City video oyunu gibi üretim hattını çalıştırabilir, makinelerin ne yapması gerektiğini işaret ederek ve tıklayarak söyleyebilir ve hatta ekran üzerinde gerçek dünya değişkenlerini manipüle ederek basitçe değiştirebilir.
These data quality issues are not exclusive to the fire service. According to experts, one of the main obstacles that may prevent AI tools from maximizing their usefulness for life safety applications center around data access. As in many industries, huge quantities of data in the life safety world—from insurers, device manufacturers, installers and maintainers, private researchers, and other entities—is proprietary and closely held. That dynamic could severely limit the amount of data available to train AI tools in a number of fire and life safety applications, and ultimately make its reliability much weaker.
Another existential question that the fire safety community is grappling with is the potential that these AI tools could eventually put firefighters, engineers, installers, and designers out of work. Whether those concerns are warranted, nobody knows for sure, but opinions are abundant and run the gamut.
Some AI advocates claim these tools won’t cost any fire and life safety jobs, in large part because the number of skilled tradespeople, including code officers, is rapidly shrinking as workers age into retirement and not enough younger people are rising to take their place. On top of that, AHJs including city building departments are already struggling to find enough staff to keep up with permitting, observers say.
The argument is that AI tools will allow the staff that AHJs do have to approve permits much more efficiently, freeing them to conduct more inspections of buildings and construction sites as well as other tasks that can benefit safety.
Other experts, including Xinyan, argue that AI will eventually lead to a shrinking workforce, at least in the private sector.
AI CASE STUDY
Evacuation and Crowd Control
Existing novel methods for smart evacuation and crowd management could soon get a boost from faster AI programs. Dynamic exit signs (as featured in the Summer 2023 issue of NFPA Journal) direct occupants to the optimal and safest route of egress in the event of a fire or other emergency, depending on the specific conditions inside the building. For such a system to work, a huge amount of data from sensors must be processed in real time to understand where in a building smoke and heat exist—and critically, where they are moving— and then coordinate potentially hundreds of different exit signs throughout the building to display the correct egress information. As conditions change, so might the optimal path of exit. The realization of such complex exit strategies will almost certainly depend on AI and machine learning systems to process the information and then predict how the hazards will change over time. Crowd management tools are also expected to get a big boost from AI. Basic tools already exist that can process information from security cameras, sensors, cell phone data, and even social media to evaluate crowd count in a given area and create density maps. Layering powerful AI algorithms over these programs could help officials at major events analyze crowd movement patterns faster to proactively manage the flow of people and avoid potentially hazardous situations. This system, used in conjunction with tools including dynamic signage, may one day automatically detect bottlenecks and other hazards at large events and use signage to steer people to prevent overcrowding.
“I think for the consulting companies, as long as these tools can save them manpower, they will have a big motivation to use them,” he said. “If someone tells a company, ‘Hey, you can reduce 30 percent of your engineers if you use this AI tool,’ I think every company will adopt it because, honestly, software is much cheaper than humans.”
Early adopters such as Chief David Povlitzin Arlington are adamant that AI tools can never replace the knowledge and decision making of actual firefighters. During an interView, nearly every positive or theoretical application he assigned to AI included a caveat with caution and the need for human oversight. If the data isn’t sound, if the public isn’t aware and on board, if the excitement to use the technology gets ahead of the safeguards to ensure its safe application, things could go south fast, he said.
“We can’t let the technology drive us—we have to know where we want to go with it and how we want to apply it,” he cautioned. “We just can’t turn the tool on and release it in our public safety ecosystems, which affect critical infrastructure and critical systems, without truly knowing some of its effects and impacts. We can’t just take our hands off the digital wheel and say, ‘the machine has all the answers.’”
JESSE ROMAN is senior editor at NFPA Journal and host of The NFPA Podcast.
Comments