Authors: Michael Goldsbury, CEM, Associate, High Street Consulting L.L.C, and Joshua Harper, Technical Analyst, High Street Consulting L.L.C
Within the past year, Artificial Intelligence (AI) and Large Language Models (LLM) have grown from esoteric technologies to household names. Models like OpenAI’s GPT (the force behind ChatGPT), Google’s LaMBDA (soon to be integrated into google under the name “Bard”) and thousands of other offshoots have become commonplace not just in the tech-sector, but in many fields and workspaces throughout the world. The growing popularity of these LLM’s backed by major corporate players has fueled a surge of new AI-based startups, models, and competitors.
In an earlier article, we discussed some of the ways in which these technologies may be useful within the field of Emergency Management (EM) and provided a short example of how a planner might use AI to brainstorm and workshop an exercise scenario. We also reviewed several drawbacks and reasons for caution. We believe it is important to continue the conversation surrounding these technologies because AI and LLMs present opportunities for greater efficiency, while also providing outlets for those wishing to do harm. To that end, our goal in this article is to explore beyond theory to practical applications and look at some real-world examples of how AI technology is being used, as well as ways in which AI technology can be built out to meet the needs of a given organization or individual.
How is AI continuing to affect Emergency Management?
What implications does the changing landscape of AI and LLM have for EM professionals? Even prior to the advent of the GPT-3.5 model and the current wave of AI-based software, experts had identified numerous ways that AI could contribute to preparedness and risk reduction throughout the world. Examples include the processing and modelling of large quantities of data (such as in meteorological forecasting1), or sociological analysis of conflict and mass migration to predict and potentially avert famines2. The application of AI and machine learning has seen significant successes in recent disaster response operations as well.
The Department of Defense’s xView program has been employed to analyze satellite imagery of damage from the recent devastating earthquake that impacted Türkiye and Syria3. Data from this program has been used to feed critical information to rescuers and other first responders about the extent of damage and estimate high-priority areas in which to search for survivors. Unlike ChatGPT or other private AI technologies, xView is entirely open-source and free, allowing any organization or individual to make use of the underlying code.
Virtual SME with AI
As advancements in AI and LLM software continue, they offer opportunities for customization tailored to the unique needs of both organizations and individual users. Leveraging software frameworks such as Langchain, we can quickly create applications powered by LLM. Consider the process of transforming vast knowledge bases and datasets and storing them in a database that an LLM can read from. This provides our LLM application with a memory bank of specific details we intend to retain and that can be used to inform the model’s output.
Imagine a situation where an AI-powered chatbot is acquainted with your organization’s data, capable of answering questions with no or little human intervention. A tool like this could prove invaluable for both client-facing applications and internal operations. The addition of data to your database could even be automated, ensuring a seamless integration process.
These kinds of systems could free Subject Matter Experts from time-consuming routine inquiries, enabling them to focus their expertise on more critical tasks. We have succeeded in training chatbots with a deep understanding of Section 508 Compliance and Public Assistance.
The 508 Compliance training allows our AI to navigate the complex waters of accessibility laws and regulations, which is especially beneficial for clients who need to ensure digital content is accessible to individuals with disabilities. Meanwhile, with training in Public Assistance, the AI can provide quick and accurate responses to questions regarding public aid programs. A user can easily cross-reference any answer with known sources and a human SME can verify their accuracy as necessary, and in turn help improve the model for successive use. Such capabilities make these AI solutions not just time-saving tools, but also valuable assets that enrich the effectiveness and productivity of our organization. Of course there are many valid concerns related to data security, the creative process, and knowledge sharing that any organization needs to reckon with before using this kind of approach.
Autonomous Agents: AutoGPT
AI agents (i.e. an AI acting autonomously) like AutoGPT are experimental AI tools that highlight the power of LLMs. AutoGPT is an AI agent that attempts to achieve a given goal by breaking it into sub-tasks and using the internet and other tools in a loop to complete the goal autonomously.
We hypothesize that AI agents like these have enormous potential to bolster natural disaster management. Users could use AI agents to autonomously collect, analyze, and summarize social media data during disasters, which could provide valuable real-time insights and help authorities in making informed decisions. For example, during a flood or hurricane, an AI agent could sift through thousands of posts, tweets, and messages, identifying those signaling immediate danger or providing real-time updates on the evolution of the incident. This information can then be compiled into social media listening reports, offering emergency management teams a data-driven foundation for strategic decision-making and targeted response. With the ability to operate continuously and process information at a scale far beyond human capacity, these AI agents can significantly enhance the effectiveness and timeliness of disaster response efforts. However, more research is needed to determine the effectiveness, feasibility, and in many respects the safety of this use case.
Closing and Conclusion
Microsoft’s recent investment of $11 billion in OpenAI (the company behind ChatGPT) portends a near-future with AI technology integrated into the MS Office Suite, used by hundreds of millions of professionals every day. Google’s work on its own LLM technology seems poised to integrate AI into the most used search tool in the world. Concerns about data security and privacy have limited many organizations from implementing LLMs. However (at least in the case of the GPT Model), OpenAI has recently addressed this issue by making it easier for organizations and individuals to opt-out of having their data used to further train the model. Furthermore, data used when accessing OpenAI’s servers externally (such as in an API request) remain private and are not used to train the model, or otherwise stored in a way that the model can use or potentially provide to another user.
Data collection methods and AI development are central to natural disaster management, and their successes and limitations shape the effectiveness of AI in this field. Sensor networks and innovative uses of satellite-derived imagery are among the recent successes in data collection that have proven beneficial for monitoring events like flash floods and avalanches. These technologies, in combination with AI, can enhance the timeliness and accuracy of detections and forecasts, and improve emergency communications. However, several technical challenges must be addressed when curating data for AI-based algorithms, including issues related to data quantity and quality. The availability of data, especially for no-notice or short-notice events like avalanches or earthquakes, can be a limiting factor, and solutions like producing synthetic data based on a physical understanding of these hazards, or using machine learning algorithms that require as few as one training event, have been proposed. In short, the use of AI and LLM will depend heavily on the quality of data that can be obtained and fed into a model.
Useful adoption of AI in disaster management will require interdisciplinary, multistakeholder, and international collaboration to develop standards that facilitate implementation. Any EM professional would be well served to stay abreast of developments in AI and LLM, both for its potential applications in the field, as well as the challenges it may present future practitioners.