Senate Standing Committee on Governmental Organization

There we go. Can you hear me now? Yes. All right. Well, thanks, everybody, for joining us here. Today is a joint informational hearing on how California is leveraging AI for effective emergency preparedness and response. Once again, good afternoon, everyone.

With that said, last week I visited the Park Fire and saw the blackened canyons and destroyed homes firsthand. In just a week and a half, this fire became the fourth largest fire in California history. It's a fast moving fire in modern memory. As we all know, climate change is real, and it's making disasters worse in our already disaster-prone state.

Climate change means our current and future hazards are not the same as in the past. That's why we need the newest and best tools available to help us face these challenges. Advanced technologies like artificial intelligence have tremendous potential to enhance our emergency management capabilities.

Fortunately, California is already making great strides in this area. For instance, the ALERTCalifornia Program uses a network of high resolution cameras and AI to detect wildfires early, providing critical real-time data to firefighters. The State Threat Assessment Center uses AI to analyze threats and coordinate responses across the agencies.

And the Cybersecurity Integration Center are conducting a joint risk analysis for threats of California's energy-critical infrastructure using generative AI, with plans to develop a similar strategy for all other critical infrastructure.

Beyond threat detention, AI can also help predict disasters by analyzing data to forecast weather patterns, predict flood or fire areas, and anticipate resources needed before disaster even happens. This allows for better resource allocation and evacuating planning, potentially saving lives. As we elevate and adopt these technologies, it's our duty to ensure they result in practical, public safety focused solutions.

Our partnerships with agencies, tech companies, and academic institutions are vital to driving innovation in this field. We must also coordinate with federal efforts such as Department of Homeland Security, AI roadmap to establish a unified disaster management strategy. By working towards a comprehensive AI strategy, we can address the challenges of today while preparing for the uncertainties of tomorrow. While we have made significant progress, our work is far from done.

It is important that we continue to commit innovation, oversight, and collaboration, ensuring that AI and other advanced technologies effectively protect all Californians and enhance our state's resilience. Thank you all for being here today, and I look forward to the discussions and insights that our experts and representatives will provide today. Now, I'll turn over to my colleagues, if they have some of their own remarks. Senator Dodd?

Thank you, Mr. Chair. Good afternoon. I'd like to thank the chairs of the Joint Legislative Committee on Emergency Management and the Assembly Committee on Emergency Management, and the staff for all your work organizing today's hearing on this very important topic, and a thank you to each of the participants with our panels today.

As we know, California's unique geography and worsening climate change puts us at a higher risk for natural disasters and emergencies, from floods, earthquakes, wildfires, just to name a few. Additionally, California is a global leader in technology, artificial intelligence, boasting 35 of our world's top 50 AI companies and a quarter of all the patents and conference papers.

During today's hearing, we'll learn how we are currently and can improve leveraging our technological advances to better prepare for and respond to these emergencies. For example, Time Magazine recently recognized CAL FIRE, UC San Diego, and industry partner DigitalPath AI fire detection tool, ALERTCalifornia, as one of the best interventions of 2023. I look forward to hearing more about that initiative today, as well as other similar efforts across public, private, and academic sectors.

Thank you again to all the participants today. Let's work together to ensure that California remains on the forefront of AI emergency response and preparations, keeping in mind that when it comes to emergencies, it's one team, one fight. Thank you.

Okay. Thank you very much, Senator. Any other colleagues? Senator Seyarto.

Thank you very much, and thank you all for being here today and bringing us up to date on the new things that are out there to help us on the prevention, detection side of this.

I think it's really important for all of us to realize up here as a Legislature that identifying when we have a problem, that is very, very valuable, but there's also the other piece of that and being able to respond to that problem and having the appropriate resources available to us.

And if we're not recognizing that through our budget process, where in this past budget we cut out camp crews again, or the ability to formulate camp crews, those are people that actually, once the AI has been used to identify an area where we can go in and potentially mitigate a future problem.

If they don't exist, then all we will know is that there is a problem. And we all know that there's a problem already. This is an exciting venture in helping us become even better at identifying those things, but it's not going to force us to provide the resources that we need once we've identified it. And so I'm anxious to hear what the latest and greatest is.

I think it's something that is absolutely a necessity for us to keep our public safe, and especially those areas where we have a lot of WUI areas, where you have the urban wildlife interface areas, because that's where houses are being built now and that's where we need this most, and that's where we need our crews out there, and when there aren't fires to be able to mitigate and respond to the AI information.

And also when there is an incident, catching it early. In Riverside County, it's been, well, hugely successful that these small fires are kind of staying small. We had one larger one this last couple of weeks, but for the most part, a lot of the smaller fires have been bombarded with air, air attack equipment. But there again, when we have multiple fires in multiple regions, we start running pretty thin on that, and we have to keep that.

And if we really want to keep our public safe and respond to the new issues around fire safety and wildland fires in California, it's going to be a combined effort of addressing all of it. So thank you very much, and I appreciate and look forward to hearing what you have to report to us.

Thank you, Senator. Any other questions, Committee Members? No? So with that, we'll go ahead and start with our first presentation. We have Kelly Hubbard, the Director of the Santa Barbara County Office of Emergency. Ms. Hubbard, when you're ready, I believe you're joining us remotely. So if you are there?

Thank you very much. Yes. Can you hear me now?

Perfect. Yes, we can.

Excellent. Thank you so much for having me today, Committee Members and audience and participants. A couple of important notes about some of my comments here. As mentioned in the opening, there are different types of AI, and different types of AI do generate different risks as well as opportunities.

And I won't get into each of those nuances throughout this, but recognizing that concepts like predictive AI in a trained environment is different than generative AI in an untrained environment. AI is being used everywhere, with or without our permission.

And so it's important to note that we are already using AI as we are identifying the risks and as we are identifying our policies and procedures on how we use it, and so recognizing that it's already being used by our government staff, our officials, and not always in ways that we may want to identify or have concerns with when we talk about security of information.

And so I wanted to reflect on some opportunities that AI has presented us at the local government level. Most definitely, AI has presented a fun and engaging concept of developing public education. How can we develop our public education and outreach on emergency preparedness in new and fun ways, using AI to help us develop those messages, that content, that whether those are videos, graphics, various outreach concepts or platforms?

And so it's a really great tool in preparedness for us, and especially in utilizing that tool with our community members. It also is a tool for our own planning so we can use AI to index our local plans.

My agency has over 26 emergency plans that we develop in house and we can use--and we're just starting to explore how we can use AI to help us look for inconsistencies or to ensure that our plans are in alignment with each other or in alignment with our partner agencies, whether that's the state, our neighboring counties, our cities.

And so we're just starting to play around with those concepts while being cognizant that this is content that then ends up potentially in a public venue or space, which generally, I don't look at as a problem for most of our emergency plans.

But there are some where we look at concepts like water utilities, where we want to be careful on how we present that information into an AI platform, and how we protect concepts like our critical infrastructure when reviewing those plans. We're also exploring how we use AI to support our vulnerable populations in disasters. Can we support those who may be deaf or hard of hearing through AI? Maybe I can use AI to generate ASL type videos when we are having evacuations.

The flip side of that is, how do we verify and confirm how accurate or consistent that messaging is with our messaging that we might be developing in English or other languages that my staff may be native speakers of? The other concept for alerting and vulnerable populations that we're looking at is in Santa Barbara County.

We have a significant population of Mixteco or Indigenous populations who have unwritten languages. So again, looking at how we could use AI to communicate with a component of your community who is definitely vulnerable and may not be receiving information in real time.

As noted earlier, some of the other areas that CAL FIRE and other partners will be talking about as part of this committee hearing is some of the concepts such as ALERTCalifornia, how are you using AI to recognize threats and recognize and do risk analysis? It's also being utilized for scenario processing.

So looking at evacuation planning, both prior to an incident, but during an incident as well. Some of the concepts associated with that is, how well is that data already informed to help make sure that whatever analysis is being provided back to us--especially if it's real-time disaster analysis--to ensure that that disaster analysis is accurate, so that if we're making life and death decisions based on that, that we feel confident and know where the data source is coming from, and that that resource through AI is going to be an accurate, safe source of analysis.

And I think that kind of points to the fact that none of this really ever replaces localized knowledge that our first responders, our public works, emergency management, local government officials may have about our jurisdictions.

I've started talking a little bit about the concerns, and I'll talk just a few more points on these. Non-generative AI in its current state absolutely has the ability to make up things. You can ask for it to develop content, and it can do--I've seen examples of it developing content that is not accurate or historically correct.

And so how do we ensure we're looking at that and validating the information we're receiving? There's concerns about purposeful misinformation utilizing AI. So can bad actors utilize AI to develop content to dissuade, change how people act and respond in a disaster?

I don't want someone to create content that looks like my content as a local government agency and maybe potentially give the community information that is not accurate or in alignment with how we'd like to ensure that our community is reacting in a disaster. One of my biggest concerns with AI is cultural incompetence.

It is well-documented that AI is generally--and please, no offense to anyone in the group--generally has an older, White male implicit bias in its content. And so how do we ensure that, as we use AI in emergency management, especially when we start talking about learning and how we support our community, how do we ensure that it has cultural competence in the content that's created?

And this is all areas that are still being explored by emergency managers on the local level, I think on the state and federal level as well, but one of my biggest concerns is how do I review and look at the content being created and understand what its biases may be?

And I think lastly, one of the things that have come up at a local agency level for myself is what are our legal liabilities as an agency to use AI and disaster preparedness, response, and recovery, and concepts like--simple concepts from--I've had a lot of meetings recently where someone has turned on Read AI, which is a note-taking system, and that then opens up me, as a government agency, to PRA considerations.

But also, if someone's using that to take notes, one: can those notes, which might have been a confidential planning or a confidential decision-making process associated with the disaster, those notes could maybe be pushed out publicly prior to them being validated and ready to go out publicly.

But two: if it has inaccuracies and it's note-taking, just think about how when you speak into your phone and you take notes with your phone and you use the voice concept where there's inaccuracies, you're going to have those same concepts within AI generated notes.

And so we look at those legal liabilities and how do we ensure we're evaluating those, recognizing them, and taking the appropriate risks where AI can support us and benefit us, but also not exposing our residents, specifically our residents, to any life safety risks, but also ourselves as local government agencies to any liabilities.

I think if I was to end with a note of where I'd like to see and hope us as government move forward with AI is this really powerful tool, and as I mentioned before, we are already using it as we are developing those policies, and so having collaborative, open space for government to learn from one another, build and grow together, develop those policies together to ensure that all levels of government are identifying both risks, opportunity, and considerations as we move forward, as mentioned, as one community. And so with that, I will leave it to some of the other speakers and be available for some questions as well.

Thank you very much, Mrs. Hubbard. Next, we'll have Michael Crews, who is the Chief Information Officer at the California Governor's Office of Emergency Services. Mr. Crews, when you're ready.

Great, thank you. Good afternoon, Chair Rodriguez, Chair Dodd, and Members of the Committee. My name is Michael Crews, Chief Information Officer at the Governor's Office of Emergency Services, and thank you for the opportunity to speak today about artificial intelligence and its potential benefits for emergency preparedness and response.

I'd like to briefly touch on Cal OES's responsibility and then provide some insights into some AI evaluations we conducted to look at the potential benefits not only for the organization, but for the state to enhance our emergency preparedness and response capabilities. So Cal OES leads the state in courting disaster response and ensuring preparedness for all hazards.

We actively support local governments and efforts towards emergency preparedness, response, recovery, and mitigation. For emergency preparedness, we develop comprehensive emergency plans, conduct training sessions, promote public education, oversee public alert systems, and disseminate real-time information during emergencies through the California State Warning Center.

During response operations, our teams operate from the State Operations Center and coordinate mutual aid and resource allocations throughout the state. Following response operations, we coordinate post-disaster recovery efforts. Our focus is on restoring communities and infrastructure and providing assistance to affected individuals and businesses. Our goal is to restore our communities and improve their resilience to future disasters.

So we all know emergency response is complex. It's challenging due to its unpredictable nature of disasters. During emergency, time is of the essence. We rely on technology to help us sift through data we collect from various agencies and organizations.

The goal is to ensure direct reward, and our state agency partners have access to the same information to make timely and effective decisions. Over the last several years, we've invested in modernizing our information systems. These enhancements have been instrumental in real-time information sharing and enhancing our situational awareness during emergencies.

Since the Governor's AI executive order was released, our Cal OES data and analytics team has focused on training, streamlining processes, and ensuring that our data is prepared to adopt AI technologies. At present, we have not adopted any AI dependent workflows in our response efforts.

However, we have explored and evaluated several use cases involving the declaration process, analyzing catastrophic plans, and streamlining the FEMA appeals and arbitration process. I'm going to touch a little bit more on those use cases. For the use cases we evaluated, we took a safety-first approach.

We established three pillars for the safe and secure use of AI; one: ensuring our research only occurred in closed environments not open to the Internet, two: we excluded using classified personally identifiable information and any other sensitive or private data, and then three: we ensured that we had a human in the middle that ensured that staff retained all decision-making authority.

Our initial approach was to target low-risk but high-impact areas for our initial evaluation and aligning these efforts with broader safety and efficiency goals. We used AI to evaluate disaster assistance appeals pending FEMA review.

We looked at the historical trends of successful appeals in the past and identified attributes that increased the likelihood of success for California applicants, particularly for our communities. We address the manual and time-intensive process for collecting data for the disaster declaration process, which typically takes 36 to 72 hours.

By integrating AI, we've reduced that completion time by 20 percent, enhancing our responsiveness in critical situations. Our initial results are encouraging and and we are considering other use cases in the future. This includes FIRIS, which is our Cal OES Fire Integrated Real-Time Intelligence System.

We conduct incident aerial mapping with FIRIS and plan to explore automating that analysis of those images that come in through our aircraft, deploying real-time language translation services at our disaster recovery centers, utilizing also AI for in-depth analysis of our catastrophic plans, identify gaps and reinforce compliance across planning, preparedness, and prevention, and improving the efficiencies and accuracy of our technical elements of the grants process through AI-driven analysis.

So, in conclusion, as AI continues to advance, its integration within the emergency management domain is inevitable and it's essential. It is. At OES, we are committed to deepening our understanding of AI technologies and exploring innovative ways to safely incorporate them into our operations, ensuring they bring substantial benefits while upholding the highest safety and security standards.

I appreciate your attention, and this time this afternoon, I'm available to answer any questions regarding our initiatives or any other future direction in using AI to enhance our emergency management. Thank you.

Thank you, Mr. Crews. Finally, we have Phillip SeLegue, the Chief Intelligence at the California Department of Forestry and Fire Protection. Mr. SeLegue, when you're ready.

All right. Well, good afternoon Chair Rodriguez, Senator Dodd, and Members of the Committees. My name is Phillip SeLegue, and I'm the Staff Chief of the Fire Intelligence Program at CAL FIRE. So thank you for the opportunity to testify today and the topic of how CAL FIRE is leveraging AI for effective emergency preparedness and response.

We are leveraging the power of AI in several key areas, including wildfire detection and monitoring, prediction and wildfire behavior, wildfire spread modeling. I'd like to begin by talking about wildfire detection and monitoring. In July of 2023, CAL FIRE and UCSD's ALERTCalifornia program announced an ongoing cutting edge partnership using AI to expedite fire detection and provide real-time fire monitoring to assist in decision-making.

This partnership utilizes the network of ALERTCalifornia's high-definition, high-resolution imagery, pan-tilt-zoom cameras that provide 24-hour persistent coverage and can sweep 360 degrees every two minutes. Of the almost 1,100 cameras, CAL FIRE sponsors 208 of them.

This new AI tool was rolled out to us in CAL FIRE in late summer of 2023 and analyzes camera feeds across California for anomalies. If an anomaly is detected, the CAL FIRE Emergency Command Center and first responders are notified of a potential fire, sometimes even minutes before the first 91 phone call has been received.

For example, between September 1st and December 31st, there were 278 dispatches for wildland fires. Of those, the ALERTCalifornia AI detection alerted 190 of those fires at the same time, if not before a 911 phone call was received. In at least six of those fires, no 91 call was ever received.

During the early detection of all of these fires, they were kept under one acre. The utilization of the AI and fire detection exemplifies by systems like ALERTCalifornia marks a transformative step in our early wildfire identification and response.

The ALERTCalifornia employs the AI algorithms to analyze data from a network of strategically placed cameras that monitor fire-prone areas continuously. These AI-enabled systems are capable of detecting smoke and flame with high accuracy, significantly reducing the time between the ignition and the detection by processing real-time video feeds.

The AI algorithms can identify fire indicators more quickly and reliably than the traditional human monitoring, even in challenging conditions such as low visibility and/or at nighttime. The early detection capability is crucial for initiating rapid response efforts, therefore minimizing the potential spread and impact of wildfires.

Integrating such advanced detection system with predictive models enhances the overall wildfire management, enabling a swift transition from detection to strategic response and resource allocation. CAL FIRE also uses AI to help predict wildfire behavior.

Traditional methods of wildfire prediction have relied heavily on historical data and basic models, which often falls short of accounting for complex and dynamic nature of fire behavior. In 2019, CAL FIRE requests for innovative ideas, RFI-squared, made possible through Governor Newsom's Executive Order N-4-19, looked to find a comprehensive solution that could evaluate fire behavior at the risk statewide to support preparedness and response.

Technosylva, which is a GIS-based company specializing in fire predictive analysis solutions and fire emergency management, was selected amongst 131 proposals. An intrinsic part of the solution is the Wildfire Analyst.

Wildfire Analyst includes extensive use of AI and machine learning to support this challenging modeling which needs a vast amount of data and in real-time to enable more accurate and timely predictions. By leveraging the Wildfire Analyst solution, we can enhance our ability to forecast fire spread, intensity, and impact, ultimately improving response, strategies, and minimize damage.

The AI and machine learning algorithms can process and integrate diverse data sets, including weather conditions, topography, fuel characteristics, and satellite imagery. These algorithms can identify patterns and correlations that may not be apparent through other conventional analysis.

For instance, the machine-learned models can be trained to recognize the influence of specific variables of fire behavior, such as the effect of wind on fire spread or the impact of humidity levels on fuel moistures. A list of AI power models inside the Wildfire Analyst includes improvements on three key areas of the fire behavior triangle: fuels, weather, and terrain. The three components are basic inputs to computer fire models used to calculate wildfire risk up to five days in advance through the state.

The integration of CAL FIRE data into the Wildfire Analyst exemplifies the benefits of this approach by incorporating extensive observational data on conditions leading to different fire behaviors. The models can be finely tuned to reflect real-world scenarios and more accurately, resulting in more reliable predictions.

Moreover, the predictive power of AI and machine learning extends beyond the immediate fire behavior to assessing long-term fire risk. By analyzing historical fire data alongside current environmental concerns, these technologies identify areas of higher risk for future fires.

Consequently, AI and machine learning not only enhance the ability to respond to active fires, but also contribute to reducing the likelihood and severity of future wildfires, fostering a more resilient and prepared landscape.

Integrating CAL FIRE data into a predictive model like the Wildfire Analyst ensures that predictions are grounded in a comprehensive understanding of local fire dynamics, further bolstering the effectiveness of fire management strategies. This innovative approach, in collaboration with academia industry partners, has positioned the department as a world leader in adopting, validating, and operationalizing these transformative tools.

CAL FIRE is actively collaborating in the adoption of these advancements, not only in partnering agencies locally, but throughout the state, nationally, and even internationally. I would also like to note that CAL FIRE's Office of the Wildland Fire Technology and Development, created through SB 109, authored by Senator Dodd in 2021, serves as the state's central organizing hub to identify emergency emerging wildfire technologies.

We are committed to reviewing new wildfire technology ideas from outside the department, including the use of AI within the parameters outlined within the Governor's Executive Order N-12-23 on generative AI use. In closing, the AI techniques and methodologies leveraged by the department are used in different scenarios to support the mission to keep Californians and our landscape safe. Thank you for your time. I look forward to answering any questions that you may have.

Thank you, Mr. SeLegue and the rest of the panel for your testimony. Now we'll open it up for questions. Any question from Committee Members? No? Assembly Member, you had a question. Go ahead.

Yes, Mr. Chair. Thank you very much. I really appreciate the presentations. I think that each of the presentations really pointed out a very encouraging trend in the use of AI to improve California's emergency response. Ms. Hubbard described the use of AI really elegantly about finding a creative new public access and engagement tools.

Mr. Crews spoke specifically about the new cost savings that are generated by doing repetitive tasks and being able to use, you know, humans more efficiently, and Mr. SeLegue spoke very eloquently about how the potential for mitigating disasters and fires that could get out of control without being able to process all of that information.

So what do you think is the next increment of this tool? Each of you come at this from a very different perspective. How do you see the most opportunity for using AI in the next year or two in your individual spheres of interest?

I could take that, not being able to.

Okay, go ahead.

Go ahead, ma'am.

Thank you. Assemblymember Hart. I think where we really see potential is how we can support our access and functional needs, community and our most vulnerable community members. So how can we leverage AI to develop new tools and resources to help those community members who are most vulnerable?

That we can create processes that helps them go through the very complex FEMA process? Maybe we can develop concepts that support them through those application processes that can be very daunting to a local community member.

How can we communicate with our community members who maybe not have, may not have traditional language or accessibility concepts, especially when we're talking about a learning and immediate life safety decision-making that has our biggest opportunity risk as well, but as long as we do it in a coordinated manner, I think it has a lot of opportunity.

Thank you. With that, Miss Hubbard, I probably have a follow up question, maybe you already answered some of it. So how do we ensure that AI-driven decisions are transparent and accountable, especially with the impact those of vulnerable populations during the disaster? Because we see sometimes some of those areas get left out, so to speak. Right.

So how can make sure that AI is addressing those vulnerable communities?

I think that's one of my big questions right now that I've been exploring personally with my team and with other counties members that are my peers. You know, can we use tools or resources that we can pre-vet? Right.

So is there a language AI concept that we can pre-vet and have local native speakers verify that as we do these translations? Are the translations accurate and culturally appropriate? Because accuracy and cultural appropriateness are two different concepts. And so looking at, you know, can we pre-vet these concepts?

Can we really feel confident in them before we use them? And then when we use them, one of the things, our county has an AI Committee, and we're exploring what is our legal but also ethical requirements or concepts we want to consider with how we identify what's created or developed using AI?

Do we need to create an asterisk at the bottom that says develop through AI? And we're really exploring that, not just legally, but ethically, to ensure that our community understands what we're using and when.

Thank you. If there's no other questions. zero, Senator, do you have a question?

Thank you, Chair Rodriguez. I appreciate the presentation. Very informative, and I really appreciated the comment made by Cal OES's leader about making sure there's a human in the middle and ensuring that a human in the middle is always there when there is high risk and trying to look for Low-risk areas.

I think that is something that the state, I'm glad to hear the state is doing. I think that's something that a lot of industries are looking to and certainly workers who are very concerned about the impact of AI on careers and on the workforce.

So as we're getting this technology to strengthen our ability to tackle what seems to be these intractable issues that are coming about due to climate change, you know, it is very important to think about, you know, where are the workers in this process and how this technology will affect even our first responders.

So I had a couple of questions. One, my question was to be put to the Chief Intelligence Officer. Tell me your last name. I want to make sure I say it correctly.

SeLegue.

SeLegue, thank you. I just had a question about the wildfire analysts because that sounds like a position, too.

Just before AI did this work, were there staff members who actually study kind of how these wildfires grow? Kind of what is the cultivation of them, how fierce they are?

I think we've seen technology sort of step in, but we've been pretty good at starting to predict and to respond to fires in the state, unfortunately, because we've had to.

Just curious, how does this integrate with the current sort of work that is being done within Cal Oes and particularly within our first response and firefighters and all of the experts and analysts and science behind that? How do these things integrate?

Will AI remove what might be a cadre of experts who have been in this field over time, or is there a sense that these things will be integrated?

That was my one question and then really appreciated Miss Hubbard's point about cultural competence and whether it's surveilling potential wildfire areas and the people who might live in those when there isn't a wildfire right and or to making sure that we're translating in a way that communities trust the messenger and respond.

I'm just curious, what are the ways that the Cal OES is really addressing this inherent issue of bias, both racial and gender bias, that has been pretty front and center in this technology from the very beginning.

And then my final question is just how do we ensure we have the accurate staff at the state level to really oversee and regulate and ensure we're putting the best tools forward? How will we bring those experts on?

And also will they have specialization in this whole issue of cultural competency and the way that we have to engage what is the most diverse state in the union? So those are my three questions, and you all feel free to respond.

Sure, I could take the first one regarding the wildfire analyst. So it's a suite of software that we procured through the N 419, through the request for innovative ideas. And it's really transformed the way that we have predictive analysis and fire spread predictions. The original fire models were created over 50 years ago from a Rothenberg.

So it was very basic binary models. And now we're looking at a very dynamic environment, taking hundreds of inputs in and being able to process. So the speed of which AI and the machine learning has the ability to speed that decision making up within the 20 seconds of response to taking our Fire Behavior

Analysts that are traded at a higher level to understand and interpret that data, to make actionable outcomes for operational use of where the fire is going, what is it going to be affecting? What critical infrastructure is in the path of this?

That's really the power of where we have seen embracing that AI and that machine learning abilities within that software. So there are personnel that are utilizing it, not only on a minute by minute, daily, weekly, hourly basis, but those that are in the field.

So if you went to the park incident now, there would be a staff of four members that are specific just to the park fire that are working off of the suite of products from the wildfire analysts through technosylva. I'll pause there.

Sure. So to translate that same analysts higher level of information coming in, and so their job changes into being able to take the new information that they wouldn't ordinarily have access to and applying it to what they're doing out in the field.

Correct.

Thank you.

In terms of the other question. oh, sorry. Go ahead.

Go ahead, sir.

Okay. In terms of the other question about cultural competencies, as was mentioned earlier, my testimony is we really wanted to look at really low-risk, high-impact. As a technologist, I'm really excited about AI. I think we're all excited.

But sometimes we need to pause and think about potentially the impacts, particularly how it's going to affect what biases, making sure it's fair, showing empathy. I think for us as emergency managers, sometimes we get really into the response, but I think the most impactful thing is showing empathy to our communities.

They've just been devastated by a wildfire, whatever it may be. Our goal is to try and get those folks back to where they were before disaster. So I think empathy, transparency, really looking at what technologies are really going to benefit us.

Again, I think low-risk, high-impact, and we're really taking that position, making sure that policy's in place, the guardrails are there for us and for our communities. So that's really critical.

But in terms of the integrations, I know with CAL FIRE and our other partners, it's really important as well in terms we're responsive, in terms of the data we collect.

So I think even before the technology, I think those partnerships that we have not only with our local counties, with our state partners, our federal partners, that we all operate using the same data, you know, it's important that we all respond in a certain way, and that data is critical, that it's accurate.

And I think that's some of the biggest challenges right now is a lot of us have different, you know, I think elements of data that that's important for us. But how does that collate together to have a statewide response that's effective and it's quick and responsive as well. So.

And if I could, and I believe it was, Senator, I think the other consideration on cultural is that there are cases where some of the AI companies have been working to correct some of their cultural biases and are still struggling with that.

And so I think important to your point is that, and some of the points that are being made here is that we cannot, I don't think, fully replace some of the concepts that we do with AI. Right.

We're always going to have to have that localized, culturally competent, community based review and identification of how we're using those resources and cross verification of what they're creating, what content AI might be creating to ensure that we're making sure that we're meeting those needs of our entire community and not just how the content might be developed through AI.

Thank you. So with that, I want to thank the first panel for the very informative information we're given. AI is really coming a long way and I'm sure the future is bright in everything we do. So with that now we'll transition onto our second panel.

Panel two is a current effort to integrate AI in emergency response disaster mitigation. On this panel, we'll first be hearing from industry and academic experts. We'll start with Dr. Bistra Dilkina, the co Director for the center of AI in Society based at the University of Southern California. Doctor, when you're ready.

Hello, dear Chair and members of the Committee, thank you for giving me an opportunity.

When you're ready

to speak today. Yes, can you hear me?

Yes, we can hear you.

Okay, excellent. Thank you for giving me an opportunity to speak today on leveraging AI for effective emergency preparedness and response. My name is Bistra Dilkina. I'm an Associate Professor of Computer Science at the University of Southern California and the Co-Director for the USC Center for AI and Society.

The Center for AI and Society was founded in 2016 and is a unique collaboration between the UIC School of Engineering and the UIC School of Social Work, and it advances the scholarship and application of AI for the well being of vulnerable communities through partnerships with community stakeholders, organizations, and policymakers.

Today, I'm here to highlight one of our successful partnership projects related to disaster mitigation planning.

The project was sponsored and funded by the Department of Homeland Security Science and Technology Directorate and its Office of University programs through the Critical Infrastructure Resilience Institute at the University of Illinois, Urbana Champaign, which is a DHS science and Technology Directorate Center of Excellence.

So disaster damage to critical infrastructures can have extremely detrimental effects, highlighting the need for strategic pre-disaster mitigation planning that makes these infrastructure networks more resilient and minimizes negative impacts when failures occur. This project leveraged artificial intelligence methods, including planning and optimization algorithms, to develop state-of-the-art techniques to find targeted water infrastructure disaster mitigation plans.

AI is a broad technology area that includes predictive and forecasting models that can be used, for example, to predict hurricane trajectories or detect wildfires in satellite images.

It includes generative AI with its recent rising successes and impacts in text, image, and video generation and finally, AI methods that focus on solving hard resource allocation problems with complex constraints and objectives to maximize. This third type of AI is what our project utilizes.

Water infrastructure is critical as it provides access to drinking water, functioning of fire departments, healthcare, and other important services that are essential to post-disaster periods. Hence, proactively fortifying water networks where most need is essential to develop our solution. The UC Center for Air and Society worked closely with the Los Angeles Department of Water and Power.

The LA Department of Water and Power is the nation's largest municipal water agency. It serves about 4 million people over 465 square mile service area. Los Angeles has an ambitious and comprehensive resilience plan, which includes provisions to build a seismically resilient pipe network.

It basically calls for upgrading targeted water pipes to earthquake resilience ones to create a sub-network of pipes that will resist most earthquake risks and ensure water delivery to critical customers, such as evacuation centers, hospitals, and police stations, as well as ensure Fire Department coverage of most buildings, basically from interviews with several utilities and other end users.

Currently, when long term water infrastructure mitigation planning is attempted, it is largely done through using GIS information and possibly hazard risk information to manually identify candidate pipes for bioengineers and subject matter experts working at the utilities.

But given the complexity of water networks, the spatially varied seismic risk and locations of critical customers, and the limited resources that are available for infrastructure upgrades, this planning problem is actually extremely challenging and it's currently addressed in an ad hoc manner.

Basically, what is lacking is the ability to generate service zone level master plans or of infrastructure upgrades that will provide the needed system-wide resilience connectivity. So this mitigation planning problem is actually not unique to LA, but is faced by many cities across the US exposed to earth risks.

So USC designed this AI based tool based on scalable resource allocation methods, which was capable of generating optimized service zone or even district scale master plans and takes into account community hazard impacts, resilience needs, economic costs, and spatial and network dependencies.

I want to emphasize that the need metrics and approaches were closely defined in collaboration with our main partner, the LA Department of Water and Power, and was piloted with their infrastructure data and specifications. We delivered optimized mitigation plans for the whole LADWP service area.

But more importantly, what we did is we gave them a tool and we trained in-house engineers to use it so they can update infrastructure information cost needs and rerun the planning tool iteratively and as needed.

It is really meant to serve as a data-driven decision support tool that can provide suggested mitigation upgrades that can then be further refined with local knowledge, additional engineering and locational constraints, and be morphed into actionable mitigation. I hope that this work will serve as a starting point for further adoption by other municipalities.

In the course of this project, we consulted with FEMA Region Nine, the East Bay Municipal Utilities District, the Metropolitan Water District of Southern California, the Seattle Public Utilities KUBOTA Membrane USA, which is a manufacturer of water pipes and water engineering consulting firms.

In closing, Los Angeles use of AI to inform water infrastructure disaster mitigation planning illustrates the potential that is unlocked by academic public sector partnership in harnessing the most up to date capabilities of AI and optimization in the context of disaster mitigation planning, and so can help achieve better effectiveness and spare human and economic costs when future disasters inevitably occur.

Thank you for the attention and I'm happy to answer any questions.

Thank you Doctor. Next we'll hear from we have Dr. Brian D'Agostino, the Vice President of Wildfire and Climate Science at San Diego Gas and Electricity. When you're ready sir.

Committee Chair, Committee Members, thank you very much for the opportunity to be here. As you heard, I'm the Vice President of Wildfire and Climate Science for SDG&E, but also a meteorologist for the organization where I've been kind of helping to advance the weather program for the last 15 years.

And AI has had a, a big part of that. I'm going to focus on three areas today. One is public safety power shutoff and how we look at AI and how it helps us anticipate, prepare, and recover from public safety power shut off.

I'm also going to look at the alert cameras and how we use those slightly differently. We were very proud to bring that network to California back in 2017 as the alert, SDG and e network.

So we've had a lot of experience with the camera network and critical infrastructure, and then also going to look at drones and how we take the images from drones and then do intelligent image processing to give us a lot of intel on our critical infrastructure.

But starting off first with public safety power shut off, of course, huge impacts to our region. AI first comes in four to seven days before we implement a public safety power shutoff through the assessments of fuel moistures, trying to identify how severe the fire potential actually is going to be.

So that's where it first gets introduced to that, then by the time we get within 96 hours. So in that three to four day period, AI is first telling us what is the peak wind gust that we are going to see in every neighborhood in San Diego.

It doesn't replace what our meteorologists do, I think, to some of the questions we had earlier, but it gives the meteorologists a new element of tools, allows them to be more granular, more specific, hyper-localized on what neighborhoods could potentially see public safety power shut off.

Now, this gives us the ability three days in advance to notify fire agencies, community leaders, folks in kind of who have a very vested interest in supporting the community through this event.

And then by the time we get to 48 hours, it gives us enough to call every customer who could potentially experience a public safety power shutoff by the time we get into the event itself. And this is where a lot of our AI is growing right now.

It's doing the real-time analysis between wildfire risk and public safety power shut-off risk. So AI is now starting to tell us what is the likelihood of a failure on the electric system, what's the probability of vegetation and trees blowing into the system? And AI is helping us with those calculations.

And then that can inform into making those decisions of, do we implement a public safety power shutoff or nothing? But very importantly, we've heard a lot about that community vulnerability right now.

It also is helping us understand what is the impact of public safety power shutoff on a community so that we can look at those risks side by side and make very informed decisions. Before moving on from the wildfire risk modeling, this is an area that we've focused a lot on.

We have just opened a wildfire and climate resilience center within SDG&E, which houses our emergency operations center, but it actually houses a new lab where we specifically look at wildfire risk and AI implications there. So that's an area of direct focus for us at SDG&E.

And then it's also worth noting that as we're coming out of a public safety power shut off, minimizing that impact is understanding when those winds are going to drop to a point that helicopters can get in the air quickly.

And if you're ready to start that inspection as soon as those winds go down, you can really start to minimize the impact on our community. AI is helping our meteorology team say exactly when those winds are going to come down. So that's a big impact there.

We've heard a lot about the camera network, so I'll take it a slightly different point that understanding impacts to critical infrastructure pathways is crucial. So identifying not only where the fire is, but we've seen the extreme heat waves impacting our region.

If we end up losing major transmission corridors due to fire during these high-load scenarios, that can increase the stress on the electric system. So understanding very early and identifying where fires are with proximity to critical infrastructure is key. And that's a major way that we've been using the camera network as well.

The last point that I'll bring up is how we use it with drones. So we go in and we do risk modeling. AI helps us with our risk modeling, back to the probability of the system failing in different areas, and it hones us into the highest-risk areas within San Diego County.

And then we fly drones on those highest risk areas. And then we run all of those photos through intelligent image processing. It tells us exactly what equipment is, where, It can tell us if there were loose bolts, it can tell us even things that you can't see from the ground, like woodpecker damage on the top of the poles and other things.

So all of this is over time. We had qualified electric workers that were going through all the images, flagging them, and training the models as we went along.

Now we run everything through the models, and the images that get flagged go back to those qualified electric workers. So they're still in the mix. But that's part of the advancement. So I'll leave it there. Look forward to answering any questions that potentially come up, and thank you for the opportunity.

Thank you. Third, now we'll have Ahmad Wani, the CEO of one concern. I believe he is joining us via phone, remotely. Are you there?

Hi. Yes. Can you guys hear me?

Oh, there you go. Yes, we can hear you. Yes.

Thank you so much for having me and, yeah, really informative session. So, my name is Ahmad. I'm the Co-Founder and CEO of a company called One Concern. One Concern does three things. Map, analyze, and monitor every piece of the world's built environment and understand its connection to impact and really the global economy.

So, you know, I'll start with some statistics to give you an understanding of the problem we are trying to solve. And it, you know, I feel like some of the, some of the use cases which we actually, which we actually operate under, have been touched by some of these speakers and assumed guests which we heard from.

So if you look at the listed companies in the United States and the New York Stock Exchange today, they have been doing climate risk and physical risk disclosures from a number of years.

And just this last calendar year, 2023, you know, the US corporates lost north of $100 billion in uninsured business interruption losses. Yet zero companies said that physical risk was a material problem or a material risk for them and at the bottom of those disclosures, people like Deloitte and KPMG and all the auditors sign and say, no materiality.

So when we double click on the problem as to why these things actually, everybody gets that green report card in terms of their disclosure, we find that status quo technology is GIS, as actually Doctor Dilkina was mentioning.

Folks, what they do is they look at their locations, whether this is in a particular flood zone or in a wildfire zone or in an earthquake zone. And guess what? 99.6% of listed company locations are not in any high hazard zone. That number is 99.8% for government critical infrastructure in the United States.

I don't know the exact number for California, but I'm pretty sure it's right up there.

However, while those numbers are super high, that they don't really have any direct exposure to wildfire risk or earthquake risk or floods, et cetera. 95% of them get actually affected during disasters by virtue of externalities, whether it's power outages, whether it's supply chain disruptions, water and wastewater infrastructure, folks not being able to get to that critical infrastructure, ultimately affecting the function inside those facilities, not necessarily the facility itself, but you're not able to actually operate that facility.

So what we started to do was recognize this problem back in 2017-2018 timeframe.

And the realization was that the risk, bulk of the risk, actually more than 90% of the risk, is in the infrastructure, it's not in the critical facilities, and how we have to somehow get the data for this infrastructure to be able to then analyze it.

Now, unfortunately, there's no download button for all the world's power lines and all the world's water pipes and transportation networks, etcetera. So we had to do it city by city. We started to do local cities in the US and several states.

We collected information for north of 70 cities in the United States, including most of the large metropolitan areas, including LA, Seattle, etc. And ultimately, we created, think of it like a spider's web, a graph. So, as compared to a two-dimensional GIS network, think of us as like a spider's web of nodes and edges.

Every facility is a node, every other facility is another node, and they're connected through those edges, which we call as dependencies. And those can be our critical infrastructure lifelines. Whether it's the water networks, whether it's the power networks, whether it's the Internet cables and the silver lines, the transportation networks.

Now, once you hit this graph with any disturbance, whether it's an earthquake or whether it's a flood, whether it's an incoming hurricane within the next 12 hours, or whether it's sea level rise over the next 30 years, you are now either shaking a node or breaking an edge.

And now you're simulating ripple effects, 2nd, 3rd, 4th, or 5th effects. Our product is called Domino, because we only focus on these domino effects or ripple effects. After doing a lot of POCs for a number of years, this was my PhD project back in the day at Stanford.

We were able to actually do our first country level implementation in Japan. Luckily or unluckily, I found myself presenting on the floor of the Japanese Senate, you know, back in 2020, and it really resonated with them. And the reason was Fukushima had basically the exact same characteristics, just like the US critical infrastructure.

In fact, the reactor was not inundated by the tsunami, so all their models from the University of Tohoku, to simulate this explosion were failing. And even after 10 years, they were unable to recreate it. What had actually happened was it had suffered a power outage from several miles out, and then it started to get overheated.

All they needed to do was get in and hook it up to a generator on the bottom floor, but the roads were blocked due to the tsunami waters. So essentially the two dependencies there which failed were power and transportation, ultimately leading to that explosion.

So ultimately we basically collected a lot of this information for the Japanese municipalities by collaborating with the MLIT, the Ministry of Economy and several other agencies in Japan. And then ultimately that led to a deployment ultimately helping the Japanese both in terms of prioritizing their emergency infrastructure budgets, and then ultimately also in response.

Their emergency infrastructure budget is a ¥7 trillion budget, which is really a planning budget, helping them prioritize actually to, several of you guys pointed out, which water pipe or which transportation line to fix, which power line to underground. You might prioritize upgrading on highway, which ultimately leads to nowhere and might not have much utility.

So ultimately doing a system-wide analysis would be really helpful to, to really understand those kinds of systems. So the system is pretty straightforward, it simulates these ripple effects and it is helpful in understanding what could happen 2,3,4 orders of magnitude down.

Now, because it is primarily based on foundation models and machine learning systems, in order to both do data synthesis as well as running catastrophe models on top of it. We are primarily our first use as we developed the US graph last year has been in insurance.

So our main customers include folks like Swiss, Rezuray, and those kinds of people who helping them do business interruption insurance better, folks like Goldman doing valuing equities better. But I feel like there are obviously applications in safety and in planning from a critical infrastructure standpoint and as well as also from a community standpoint.

So yeah, happy to answer any and all questions.

Thank you Mister Wani. Last but not least, we have Reymund Dumlao, the Director of State and Local Government and Education Cloud Sales for Google Cloud west region. When you're ready, sir.

Thank you members. My name is Reymund Dumlao and I am the Director of State and Local Government for Google Cloud for the Western Region. Thank you for the opportunity to testify today on Google and Google Cloud's ongoing commitment to leveraging artificial intelligence to improve government services and specifically for enhancing emergency management capabilities.

As a global leader in AI research and development, Google has a unique responsibility to apply our technological expertise to address critical societal changes, including disaster preparedness, response, and recovery. I will highlight key initiatives, partnerships, and research efforts that demonstrate our dedication to harnessing AI's potential for improving emergency management outcomes.

This includes providing early warning systems, optimizing resource allocation, facilitating information dissemination, and supporting long-term recovery efforts. We recognize that AI is not a panacea, but when deployed responsibly and ethically, it can be a powerful tool for saving lives, protecting communities, and building resilience.

Google's approach to AI is guided by a set of core principles that prioritize social benefit, safety, fairness, and accountability. Google Cloud offers a suite of AI-powered tools and services specifically designed to support emergency management agencies and organizations.

These solutions leverage years of Google research, advanced machine learning algorithms, and vast computational resources to analyze data, predict risks, and optimize response strategies. Broadly, these AI tools can offer early warning systems. The models can analyze real-time data from various sources, such as weather sensors, social media feeds, satellite imagery to detect early signs of impeding disasters.

This enables emergency managers to issue timely warnings and alerts, giving communities more time to prepare and evaluate, if necessary. Risk assessment and mitigation. By combining historical data with AI-driven predictive modeling, we can help identify risk areas high-risk areas for specific types of disasters.

This information is crucial for developing targeted mitigation strategies such as reinforcing infrastructure, establishing evacuation routes, and pre-positioning resources. For example, we work with the Hawaii Department of Transportation to evaluate the sea level rise and how it impacts infrastructure and emergency response.

Thirdly, damage assessment. In the aftermath of a disaster, Google's AI tools can quickly analyze satellite and aerial imagery to assess the extent of the damage. This real-time information helps emergency responders prioritize their efforts and allocate resources where they are needed the most. Crisis communication. During emergencies, clear and efficient communication is essential.

Google Cloud's AI-powered chatbots and virtual agents can handle a high volume of inquiries from the public, bringing up human responders to focus on critical tasks. Additionally, Google Translate can facilitate communication with non-English speakers, ensuring that everyone has access to vital information.

Resource allocation. AI algorithms can optimize the allocation of resources, such as personnel, equipment, and supplies to ensure the most efficient and effective response. This can particularly be valuable in situations where resources are limited and time is of the essence.

Specifically, our research teams are focusing on three areas that produce valuable insights that can be used by government at all levels to save lives. Forecasting floods. Floods are the most common natural disasters, causing thousands of fatalities and disrupting lives of millions every year.

Our flood hub platform leverages AI and geospatial analysis to deliver real time flood forecasts, and recently we have expanded coverage to 80 countries, providing forecasts up to seven days in advance for 460 million people across Africa, the Asia Pacific Region, Europe, South and Central America.

Additionally, late last year we announced the expansion of this to the US and Canada, covering more than 800 locations by rivers where more than 12 million people live.

Tracking wildfires. As wildfires unfortunately become more frequent, we're working to provide information about where fires are, and we are exploring how AI can predict where a fire will spread. In 2023, our SOS alerts have provided timely safety information to more than 30 million people across 120 wildfire events around the world.

To help map fire boundaries, our wildfire boundary tracker uses AI and satellite imagery to map large fires in close to real-time and update every 15 minutes. This is available on Google Search and Maps in fire-prone parts of the US, Canada, Mexico, and Australia, and we're working to expand our coverage.

In addition to knowing where wildfires currently exist, firefighters need to anticipate where they will go and how to mitigate them well before they start. We're collaborating with us Forest Service to make biggest updates to our current fire spread model in 50 years.

This model leverages machine learning to model more fire dynamics to help fire authorities train firefighters, plan effective fuel treatments, and battle large-scale fires more safely and effectively while in the field. Responding to extreme heat. Extreme heat also impacts public health, and health-related deaths are on the rise.

We launched extreme heat alerts last year, so when people search for information on extreme heat, they see details about a heat wave that's predicted to start and end, tips on staying cool, and related health concerns to be aware from the Global Health Information Network. Cities are also looking for ways to prevent heat islands.

Our tree canopy tool combines AI and aerial imagery to show where shades are in the city, helping cities better understand where to plant more trees to reduce heat. Another way we're helping to reduce heat islands is by providing insights about reflective roofs called cool roofs.

Our cool roofs tool uses AI and aerial imagery to map out the solar reflectivity of cities so urban planners and governments can identify which areas would benefit most for deploying a cool roof solution, such as a white roof. The pilot is now live in 15 cities such as New York, Nashville, and Melbourne.

Early warning systems and predictive analytics. We are excited about the future of AI and one of the most promising applications of AI in emergency management is the development of early warning systems and predictive analytics. Google Cloud AI platform provides tools and infrastructure for emergency management agencies to develop custom predictive models.

These models can analyze vast data sets, identify part patterns, and generate actionable insights for anticipating potential emergencies and optimizing response strategies. By integrating AI into existing infrastructure management workflows, we can improve situational awareness, decision making, and ultimately save lives. During a crisis, rapid response and efficient resource allocation are crucial for minimizing damage and saving lives.

Google is actively collaborating with emergency management agencies and nonprofit organizations to develop AI-powered tools that streamline these processes. Our SOS alerts, integrated with Google Search and Maps, provide critical information during and after a crisis, including evacuation routes, shelter locations, and relevant news updates.

Additionally, machine learning algorithms can analyze real-time data on traffic, infrastructure damage, and resource availability to identify optimal routes for emergency vehicles, allocate personnel effectively, and prioritize the delivery of essential supplies. By automating and enhancing these critical decision-making processes, we can significantly improve the speed and effectiveness of emergency response efforts.

Effective communication is essential during a crisis, both for informing the public and coordinating response efforts. Our public alert systems deliver timely and accurate alerts to individuals affected areas through Google Search and Maps.

Additionally, Google Cloud's natural language processing capabilities are being used to analyze social media data, news articles, and other sources of information to identify emerging trends, rumors, and misinformation. This information can be used to inform public messaging, address concerns, and combat the spread of false information during a crisis.

By leveraging AI to analyze and interpret vast amounts of data, we can provide accurate and relevant information to the public when it matters most. The aftermath of a disaster often presents significant challenges for communities as they strive to rebuild and recover. Google is committed to supporting long-term recovery efforts through AI-powered tools and resources.

For example, our crisis response platform provides a centralized hub for information and resources related to specific disasters, including donation opportunities, volunteer opportunities, and recovery guidance. Our AI capabilities are being used to analyze satellite imagery and other geospatial data to assess damage, prioritize recovery efforts, monitor progress over time.

This information can be used by governments, NGOs, and other stakeholders to allocate resources, effectively identify vulnerable populations, and build more resilient communities in the face of future disasters. This also helps with claims damage. We have worked on assessing whether the damage was for wind or flooding for hurricanes in Florida.

And we've actually worked with Cal OES to analyze Street View camera data and assess wildfire damage for FEMA. In conclusion, we are committed to using AI for good, and emergency management is key, is a key area where we believe technology can make a significant impact.

We also believe it is extremely important to build trust with our customers and users. By investing in research, developing innovative tools, and collaborating with partners, we are working to improve early warning systems, optimize response efforts, facilitate communications, and support long-term recovery.

We are proud to be working alongside dedicated professionals across the country to enhance public safety and resilience in the face of natural disasters and other emergencies. Thank you for your time today. I'm happy to answer any questions the Committee may have.

Thank you, Mr. Dumlao, and thank you for the panelists. So bringing it back to Committee Members. Any questions? Senator, go ahead.

Thank you. And thank you for the presentation. And representing a South Central Los Angeles, we don't too often in the core of my district, deal with wildfires, but I appreciated the nodes and the analogy of the ripple effect.

And I guess this is my question for one concern, presenter, you mentioned that Japan is spending $7 trillion, yen, ¥7 trillion, to sort of look at this ripple effect.

And it just made me think about to what extent given, you know, our state, given, you know, the beautiful topography we have, but also the challenges that we have. You know, when you think about the scale of this kind of analysis, does AI create an opportunity to really reduce the cost? And it's really assessing this?

And how could the State of California wrap their arms around this kind of deep, deep study and collaboration across technology to give us the kind of roadmap that would help us not just, you know, deal with the crisis in real-time, but to prevent these kinds of crises.

Thank you, Senator. And I should clarify, they're spending ¥7 trillion in actually upgrading their infrastructure every single year. So that's, that's around $70 billion a year.

Thank you for that.

Infrastructure upgrades. Yes, they call it, they call it the emergency infrastructure fund for some reason. I don't know why, but it's an actual budget which they deploy almost every single year in upgrading levees and upgrading bridges, etcetera, from a point of view of disaster mitigation. So I'll give you an analogy.

And the first city which we actually did this for was the City of Seattle. And I went to the port and the airport and I was collecting information for like, hey, when was this crane upgraded? What about that one? And what about this airport runway?

And it took us three years to do one city for the entire Japan it almost took us six years. We've already launched the United States. So we have data for every electric pole in California today. We know the health of every substation. And so those components we've already built. And the reason is because of AI.

So now I know what a type 32 substation or an airport runway actually looks like, when was it built? And then therefore apply a sort of what we call as a fragility function, understanding the engineering strength of that. So AI has definitely sped us up in a great way.

We are still not global, so we have the coverage for the United States right now. Having said that, there are certain limitations in terms of using a combination of real data and synthetic data. For a lot of municipalities, we have real data who have been able to provide us this data.

I won't name the cities, but there are several cities, even in the Bay Area, who don't know where their own water pipes are. They have 10 streets worth of water pipes, and 11 to 15 is missing. Essentially, we basically are then using some combination of urban planning and machine learning and AI to sort of guess.

Okay, you know, if it's a 10-story building with 1000 people, it might have these many combinations and this much pressure in three-inch and six-inch water pipes, for instance. Right? Or silver lines, etcetera. So it presents an opportunity to scale to launch these projects really fast.

But I believe, you know, the Cal OES Executive, he mentioned that. And then also the Santa Barbara leader mentioned the idea that, hey, we have to be extremely careful in terms of understanding the biases which come in.

So we realized that when we are synthesizing this data and then the decisions which we are actually recommending, there's obviously a human in the loop in every decision. That's obvious. But while our precision is in the high nineties, our recall is in the mid-seventies.

What that means is if you were to give me 50 critical infrastructure, and you were to tell me which of these is the most vulnerable stack, rank them, or pick the ones which are most vulnerable, I would pick the most vulnerable with a high 90 precision, but I would leave some, and those would be 30% of them, which I would be leaving out.

And there are ways to actually optimize that and get more humans to then train it better. But essentially, we have to really focus on specific use cases where we can be mindful. But these are the opportunities. And so if you recall, you know, obviously many, many upgrades have gone into the Hetch Hetchy water system.

But, you know, most of the hospitals in the Bay Area, you know, and in fact, even in LA, are serviced by specific aquifers and specific water systems. And those have high. Oh, sorry.

Thank you for that answer. I'm very, very complete.

Senator, sir, I think you had a question.

I do. And this is for Mr. D'Agostino. You know, PG&E up here has been doing a lot of undergrounding. SDG&E probably has to look into a lot of that. I don't know how much of that is going on.

All of our public utilities, that's a big push to try to minimize what you were talking about, the PSPS shutoffs. Yes, because the worst time to shut off somebody's electricity is when it's 105 degrees and they've got fire coming through and smoke and all of that.

So to what extent is that AI being used to determine where it is practical and can be done? Because it is impractical to underground everything. But I think there's a lot of opportunities out there. So is AI being used to determine that so we can do that in our planning?

It absolutely is. And that's in that risk modeling and on both sides of the risk modeling. So I mentioned there's that likelihood. So we have to find those areas on the electric system that are prone to the most extreme winds and are prone to seeing a failure that potential ignition.

But then also we've kind of heard about how AI is being used on where would the biggest fires happen? What would be the biggest consequence? We even bring in some of Google's structure analysis to look at the impact. But the short answer is yes. AI is being used to identify the highest-risk locations to underground.

So with that said, you know, I had a quick question and maybe it's a dumb question just hearing about the technology and AI, where it's come from, where it's going. I'm looking at seismic activity, right? All these earthquakes, you know, that happens year after year.

You think there will be a way to put all that data into AI where they would kind of predict future earthquakes come in? Obviously we're overdue for the big one, but I don't know. Would that be something you guys think? I don't know.

Maybe it's not in your privy to answer, but I'm just kind of thinking outside the box a little bit as we talk about responding and use of AI for everything. You think something like that could come in the future? As a question, if anybody can answer.

As an earthquake engineer, I feel. Thank you for the question. Assemblymember. As an earthquake engineer, I feel like it's almost an impossibility unless we have a lot of data to simulate. The problem is we cannot dig beyond 18 to 20 miles, and some of the seismic activity actually starts way deeper than that.

And because we don't have enough elements in the periodic table which will enable us to dig deeper than a certain depth, we are unable to put sensors at that depth and therefore unable to understand. There are indirect ways by which folks have tried to sort of say, okay, there is an earthquake in Peru.

Maybe it reached Washington, you know, shook the Washington state monument after some period. So. So there could be some ripple effects there and that kind of stuff. Obviously, there's early warning, which you're very familiar with, but I don't really see data is the problem, is my point. Right. So thank you.

Any other questions? So with that, I believe that will come to an end of our hearing. Almost to the end of the hearing. I will just move. There's no other questions. We'll move on to public comment. If there's any Member in the public for a comment which doesn't look like there's none.

So with that said, I want to thank everybody for participating in this very informative hearing. AI has come a long way, and it's. The future is bright in what they're doing. Brian, just want to talk. I've been to PG&E's weather station down there in San Diego. It's fabulous.

I would encourage all my Committee Members or folks that if they ever get opportunity, what you're doing out there is just. You guys got the, a good thing going on there, what you're talking about, predicting the weather and everything you do over there. So, with that said, this hearing will come to a close.

Thank you all for participating and thanks for my colleagues joining me here today. Thanks. Meeting is adjourned.