AI is Transforming Crisis Management! Preventing "Help" from Getting Lost in a Sea of Posts — Can AI That Organizes Disaster-related Social Media Be Useful in Japan?

The heavy rain has flooded the roads. Power outages continue in residential areas. Families are stranded, calling for help.

During disasters, social media gathers information that only those on the scene can convey. However, it also includes repeated information, past footage, hearsay, and expressions of concern. As posts increase, the task of finding "information that needs to be confirmed now" becomes more burdensome for responders.

The problem is not just the lack of information. An overflow of information that humans cannot fully read also becomes a challenge in disaster response.

To reduce this burden, a research team from Texas A&M University is developing a disaster information classification technology using artificial intelligence. Phys.org, dated November 9, 2026, introduced research that utilizes large language models to make it easier to find necessary information from social media posts.

The expectation is to reduce the time it takes for personnel to sort through a large number of posts, allowing them to spend more time verifying situations and considering responses. However, it is necessary to separately understand what this technology can do and what it has not yet proven.

Classifying disaster posts from multiple perspectives

The research is titled "CrisisSense-LLM." The results were published in the academic journal "Advanced Engineering Informatics."

The university's introduction article features Bo Li, Ali Mostafavi, and Kai Yin. The research uses a benchmark with human-assigned classification labels for social media posts related to multiple disasters.

The focus is on information related to disasters, weather, and humanitarian aid. It also deals with multiple aspects, such as what events the posts are about, whether they contain information useful for situational awareness, and what type of support information they pertain to.

The publicly available preprint explains a method of using instruction data created from disaster-related posts to further train existing large language models. The idea is to adjust the model to the task of classifying disaster information.

For example, imagine a post like the following. This is a fictional example for explanation purposes.

"The river has overflowed, and the road in front of our house is impassable. We have evacuated to the upper floor, but we are running low on drinking water."

This sentence contains different pieces of information: flooding, road inaccessibility, evacuation status, and a shortage of supplies. A single classification might not fully represent the necessary details contained in the post.

If the information can be organized from multiple perspectives, it becomes easier for the person in charge of investigating road damage and the person assessing the need for aid to find relevant posts. This is where the practical value of this research lies.

"Being able to classify" and "being able to confirm as fact" are different

There is a crucial point to be aware of when reading this research.

Even if a post can be classified as a "rescue request," it does not mean that it has been confirmed that rescue is actually needed at that location.

Understanding the content of a sentence and determining its type is a different task from verifying the facts occurring on the ground.

A plausible fake rescue request could be classified as a "rescue request" based solely on the characteristics of the text. Conversely, short posts written in urgent situations or posts with many typos might be difficult to interpret.

From the introduction article, we cannot conclude that CrisisSense-LLM can reliably detect misinformation. It is also not appropriate to describe it as a system that automatically determines rescue priorities.

For practical use, it will be necessary to ensure that one can return to the original text from the classification results, confirm the posting time, and trace the place names and sources of information.

For example, simply displaying "shortage of supplies" on a screen is not sufficient as a basis for judgment. When, where, who needs what, and whether the situation is ongoing must be clarified. The design must allow organized information to be verifiable.

Not yet operational in real disasters—do not confuse expectations with achievements

The research team clearly states that this system is under development and has not yet been deployed in ongoing disasters.

This point affects how the results are perceived.

There is a difference between classifying pre-prepared post data and using it in disaster sites where information may be interrupted, contradictory, or where new place names and expressions continuously appear.

To verify its effectiveness on-site, it will be necessary to examine not only the accuracy of classification but also how the confirmation time for personnel changes, whether important posts are missed, and whether misclassifications can be easily corrected.

Especially, the impact of misses and false detections is not the same.

If unrelated posts are mixed in, it increases the verification work for personnel. On the other hand, if posts that truly need aid are omitted from the list, there is a risk of losing the opportunity to understand the situation. It is difficult to judge the appropriateness of its use in rescue operations based solely on an evaluation of "overall accuracy."

This research indicates the potential for organizing information to support disaster response. It should be distinguished from the stage where the reduction of rescue time and mitigation of damage have been confirmed in practice.

The complex relationship between disasters and AI as shown by social media reactions

So, how is it perceived on social media?

As of November 10, 2026, within the range of publicly available information, there was no consolidated reaction from general users to this research itself. Therefore, it cannot be introduced as having widespread support or opposition.

On the other hand, actual social media posts related to the use of AI during disasters reveal another aspect. There is both the potential for AI to assist in organizing information and a caution against increased confusion due to AI.

The technology news media "Rest of World" reported on an incident involving AI-generated images depicting rescue scenes during the 2024 floods in Brazil.

In a post on X introduced by the same article, user "@mphistoria" pointed out that images shared on WhatsApp were AI-generated and called for regulation. This is not a reaction to the current research but a response to AI-generated images circulating during disasters.

This incident highlights problems that cannot be solved by the speed of information processing alone.

If inaccurate posts are efficiently organized, the list may become easier to view, but it does not necessarily mean that the foundation for judgment is correct. A well-organized screen or a clear summary might give the impression that the content has been verified.

Therefore, disaster-oriented AI needs a mechanism that allows humans to distinguish not only "what is written" but also "what is unverified."

Japan also has a history of analyzing disaster-related social media

This topic is not a distant overseas study for Japan.

The National Institute of Information and Communications Technology (NICT) has been conducting trial releases of the disaster social media information analysis system "DISAANA" since 2015 and the disaster situation summarization system "D-SUMM" since 2016.

D-SUMM was introduced as a system that analyzes disaster-related social media posts and organizes and summarizes disaster reports by region. Japan has been tackling the challenge of making it easier to understand situations from a large number of posts during disasters.

However, the public trial releases of both systems ended on December 28, 2023. NICT explains this in the context of the spread of private services utilizing the technology. It is important not to introduce them as currently available trial services.

When viewing the current U.S. research, it is important to avoid the understanding that "analyzing social media with AI is a first."

There is value in verifying how large language models can handle various expressions and how multiple pieces of information can be organized, in light of the efforts accumulated in Japan.

If used in Japan, place names and "unwritten information" become issues

From here, the discussion is not about the confirmed content of the current achievements but about the points to consider if used in Japan.

First, expressing locations in Japanese disaster posts is challenging.

Instead of formal addresses, expressions like "the usual bridge," "behind the station," or "beyond the old road" might be used. Posts may include district nicknames or facility abbreviations. Even if words can be classified, if the location to respond to is unclear, it is difficult to connect to on-site verification.

Handling Japanese does not simply mean translating sentences. It is necessary to verify how well it can handle posts that include omitted subjects, dialects, short fragments, or conversion errors.

Moreover, damages not posted on social media cannot be grasped through social media analysis alone.

Regions where communication is cut off or people who do not have the capacity to post may not appear on the screen. It is important not to interpret "regions with few posts" as "regions with little damage."

When using analysis results, it will be necessary to cross-reference with phone calls, on-site reports, and information held by local governments. While social media is useful as a supplementary information source for situational awareness, it does not capture the entire region without omissions.

Towards an AI that can be questioned in the future, but with the need for evidence

The research team aims to develop an AI agent that disaster response experts can interact with next.

They want to advance from the current classification-centric system to one where personnel can ask questions to find information.

For example, future use cases might include questions like "Are there any posts requesting aid in this area?" or "Which are the new reports on road damage?" However, these are hypothetical examples and not features reported as implemented in the current article.

If it becomes interactive, users might reduce the effort of opening classification items one by one. On the other hand, designing a system to verify the basis of responses becomes even more important.

A single sentence like "Rescue is needed in this area" does not reveal whether the original information is from one source or multiple, or how many hours ago it was.

Future systems should ideally show the original post, time, and confirmation status along with the response, and communicate any information shortages. There is still a gap to bridge between answering in natural language and being usable as a basis for disaster response decisions.

Can human judgment be placed beyond organized information?

The potential shown by the CrisisSense-LLM research lies in assisting with the overwhelming amount of information during disasters that humans cannot fully read.

To measure its value, it is necessary to look not only at the number of posts it can process but also at whether it could reach the necessary information, reduce oversights, and allow personnel to verify the basis.

Finding a post that says "Help." Verifying its content. And connecting it to appropriate action.

These three are connected but are not the same task.

How to link AI classification to subsequent verification and action is crucial, even when considering its use in Japan. The way it is used on-site is as important as the technology's performance.

Sources and References

1. Phys.org:Published on November 9, 2026. Introduces the research overview, classification targets, non-deployment in real disasters, and future plans for interactive AI.
   https://phys.org/news/2026-11-ai-tool-crisis-communication.html

2. Texas A&M University: University research announcement
   Published on November 6, 2026. Primary source confirming the research team and development status.Texas A&M Stories
   https://stories.tamu.edu/news/2026/11/06/researchers-develop-ai-tool-to-improve-crisis-communication/

3. Advanced Engineering Informatics: CrisisSense-LLM published paper
   The paper is titled "CrisisSense-LLM: instruction fine-tuned large language model for multi-label social media text classification in disaster informatics." The full text from the publisher was not obtained this time, so detailed performance figures are not used in the text.
   https://doi.org/10.1016/j.aei.2026.104836

4. arXiv: Public preprint of the same research
   Used to confirm research methods of instruction-based additional learning and multi-perspective classification of disaster posts. Referenced separately from the final published version in the academic journal.arxiv.org
   https://arxiv.org/abs/2406.15477

5. Rest of World: Report on Brazil floods and AI-generated images
   Published on May 23, 2024. Verified the post by X user "@mphistoria" through media reports. Introduced as a reaction to AI during disasters, not a direct response to the current research.restofworld.org
   https://restofworld.org/2024/exporter-brazil-floods-ai-misinformation/

6. NICT: Announcement of the trial release of the disaster situation summarization system D-SUMM
   Primary source of efforts in Japan to analyze, organize, and summarize disaster information from social media.NICT - National Institute of Information and Communications Technology
   https://www.nict.go.jp/press/2016/10/18-1.html

7. NICT: Regarding the end of the trial release of DISAANA and D-SUMM
   Material confirming the start of the trial release and its end on December 28, 2023.##HTML_TAG_236