Earlier this year, Todd O’Hara was speaking on a panel about data at the Aerial Fire Fighting Global Conference when he realized that many of the questions from the audience shared a common subtext: artificial intelligence.
“A lot of the stuff that was coming up from questions from the crowd was kind of hinting at the AI question without mentioning it,” recalled the CEO of TracPlus, a data management company specializing in aerial firefighting, in an interview with Vertical. “I thought, there’s an interesting topic there around how do we actually responsibly introduce AI into a world which is full of data already, but do it in a way that makes sense.”
One month later, TracPlus released a white paper exploring that question titled “AI Makes the Call. Who Answers for It?” Authored by O’Hara, the paper argues that AI systems are increasingly shaping how decisions are made in firefighting operations, yet most fire agencies haven’t formally decided whether that is acceptable or established a framework for who is responsible if something goes wrong.
“We’ve probably arrived at a point where we’re further into AI involvement in the operational ecosystem of aerial firefighting and firefighting in general, but haven’t really taken the time to think about what the impact of that [is],” he said. According to O’Hara, many agencies have incrementally adopted AI-enabled tools such as automated dashboard software without fully understanding how they work or the quality of data they’re based on, potentially setting them up to make faulty decisions when lives and property are on the line.
O’Hara’s white paper hits on one of the central questions facing the use of AI in any industry: to what extent these tools should supplement or substitute for human experts. To O’Hara, they should clearly do the former, not the latter, whether the task is writing a marketing plan or predicting how a wildfire will spread over the next 24 hours.
“That was a big parallel that we were trying to get through in the white paper,” he said. “Embrace [AI], great, but make sure that you understand the thing that you’re asking it to do manually, and run it in parallel and validate that using actual, real, trusted data.”
Yet, for some of the companies developing these tools, substituting for human experts is exactly the point, when the alternative is no expertise at all.
“There are a lot of experts out there who know their job very, very well, who basically do a fire forecast spread analysis in their brain based on their experience,” said Thomas Grübler, the chief strategy officer and co-founder of OroraTech, which offers a wildfire detection and monitoring platform that relies heavily on AI. Grübler told Vertical that as wildfires become more pervasive, organizations that do not have deep expertise in wildland firefighting will increasingly be tasked with making decisions about how to manage them — and he thinks AI can help fill the gap.
“We want to give the experts tools in a way that also non-experts can use them,” he said, explaining that not every agency has the luxury of consulting trained fire analysts with years of specialized experience. For personnel who are “not daily nine-to-five on wildland fire, how can we give the best tools to these people so that they can do a better job?”
A technology leap, with limitations
“Artificial intelligence” has meant different things to different people since the term was invented over 70 years ago. Today, it generally refers to various types of machine learning, including but not limited to large language models like ChatGPT.
Machine learning relies on computational algorithms called artificial neural networks. Unlike traditional software programs that follow hard-coded rules, neural networks are calculators of statistics that identify patterns in training data and apply what they’ve “learned” to new data. Their uses are varied — they can be trained to identify new wildfire starts in satellite images or photos of cats on the internet — and they are much better at many of these tasks than traditional programming methods.
OroraTech, which has its own satellite network dedicated to wildfire management, uses AI to detect new fires from space. Grübler said each satellite is equipped with the AI hardware and software needed to locate new fires, and is able to beam this information back to Earth before an actual image of the fire. “You can use traditional algorithms, but AI improves the quality of detections extremely,” he said. The company also uses AI for its Fire Spread tool, which predicts how a fire will spread based on wind, vegetation, elevation, and other customizable inputs.
O’Hara, whose company announced a two-way data integration partnership with OroraTech in October of last year, sees potential for AI to have a positive impact on firefighting in multiple ways, from enabling better weather forecasts to streamlining paperwork. “The thing that AI is really good at, and much better than humans at, is processing data,” O’Hara said. When it comes to weather modeling, for example, “we’re seeing the advantages of huge volumes of data being able to give people better information and insights into what’s going on, and potentially more accurate suggestions on how to better combat the situation that’s coming up,” he said.
However, because neural networks are inherently statistical in nature, they will produce a likely answer, not always a correct one. The calculations they use to arrive at their output are so complex that humans can’t verify those calculations in the same way that traditional code can be checked line by line. And they’re only as good as thedata they’re fed for training and analysis. If a neural network is only trained on photos of black cats, it won’t be very good at recognizing orange ones, and if a fire spread model is given inaccurate fuels and vegetation data, it won’t output a reliable forecast.
According to Grübler, sourcing ground truth data is one of the biggest challenges in developing and refining machine learning models for wildland firefighting. When it comes to fire detection, for example, fire starts detected by OroraTech’s satellites in remote areas may not be attended by firefighters, so the company can’t always confirm that the detection was accurate.
“You don’t get all the fire parameters always, so it’s a challenge to have this reality to compare against to actually really prove that your algorithms work,” he said. He contrasted this with developing machine learning models for the stock market, where precise and accurate data is readily available.
Nevertheless, Grübler said the company’s models have been consistently improving, and the industry as a whole is investing in gathering better data for model training and analysis. “Fuel maps [are] a big thing, of course, in improving the quality of the forecasts, and more and more customers are flying airplanes to map the fuel better,” he said.
Trust, but verify
Because TracPlus manages so much data for aerial firefighting customers around the world, it has unique visibility into the range and quality of data feeding some of these models. In his white paper, O’Hara notes that data from aircraft sensors can degrade over the course of a season due to the harsh physical environment — and he worries that users of AI tools may not realize when the underlying data is compromised.
“A system consuming degraded inputs does not display a warning. It produces outputs. Those outputs look the same whether the data behind them is sound or not,” he writes. O’Hara draws on decades of automation research to suggest that as AI systems are handed progressively more responsibility, the ability of their human users to critically evaluate them degrades as well: They become more likely to blindly accept an AI-generated recommendation without exercising their own independent judgment.
O’Hara argues that fire agencies should introduce AI tools using the “parallel validation principle,” meaning they should run the AI recommendation alongside an independent human decision, record both and compare them against outcomes over time (a strategy that assumes the presence of fully qualified human experts).
“That comparison is how you know whether the tool is actually helping, whether it performs consistently across your specific geography and fuel conditions, and whether there are categories of scenario where it should not be trusted,” he writes. This approach can also help identify when a system stops performing well as a result ofdegraded inputs, he suggests.
Yet, users of AI tools aren’t the only ones with the incentive and responsibility to ensure they work — their developers are also highly motivated to get them right. Grüblersaid that behind the scenes, OroraTech has developed an AI agent for firefighting that can describe the current status of a fire and undertake complex tasks with minimal human involvement. But the company is not rushing to push it out to customers.
“We have a prototype now, but we also know that we cannot release something that you cannot trust,” he said. “We know that firefighters are really picky on technology, and if something doesn’t work, they will never touch it for the next five or 10 years.”
