�鶹��Ѱ�����

Researchers at �鶹��Ѱ�����Boulder’s Institute of Behavioral Science and the Natural Hazards Center are finding new ways to personalize alerts that could help people take more timely action in disaster scenarios

Wildfires. Tornadoes. Hurricanes. Floods. The increased frequency of natural hazards and disasters cropping up all over the world is staggering—as is the potential loss of human life.��

To limit casualties amid these disasters, emergency management organizations use a variety of methods to communicate warnings and evacuation alerts. Whether or not people comply with them is another story.��

“We know that many individuals don’t really pay attention to disaster alerts until it's too late to take any kind of action,” says Amir Behzadan, professor in the Department of Civil, Environmental and Architectural Engineering and fellow in the Institute of Behavioral Science’s . “We're trying to bridge this gap by ensuring that disaster warnings translate into timely action, particularly among communities, individuals and population groups that may have greater barriers to accessing, processing and responding to this type of risk information.”

��

�鶹��Ѱ�����Boulder scholars Amir Behzadan (left) and Mary Angelica Painter (right) collaborated on research employing generative AI to explore a new approach to getting people to respond more quickly to disaster alerts.

In a new study, funded by the Federal Emergency Management Agency (FEMA), Behzadan, (a research associate at the Natural Hazards Center) and a team of �鶹��Ѱ�����Boulder graduate students employed generative AI to explore a new approach to getting people to take action more quickly.��

Leveraging Google’s large language model (LLM), Gemini, the team generated and distributed more personalized alerts—which included reminders about pets, dependents and medications—to sample groups to determine how well they would be received and how they might motivate recipients to take action. were published in September in the International Journal of Disaster Risk Reduction.��

Getting personal��

Study participants were first asked a series of questions about their sociodemographic factors (i.e., gender, age, education level, language preference and prior disaster experience). Knowing these factors was key to understanding how people from different walks of life might perceive alerts.��

Participants were then asked about personal traits that might impact their needs in a disaster scenario (e.g., whether or not they lived close to family, if they had dependent family members and/or pets, if they used medication and what type of home they lived in). The participants’ ZIP codes were also requested so the messages could be tailored to the types of disasters they would be most likely to experience, given their geographical location.��

Their responses were then fed into the LLM, along with examples of the tone the messages should mimic. “The role of AI here was to focus on those personal traits and then try to cater the content and style of the message to the participants that possess those traits,” says Behzadan.�� ��

One portion of participants was given the AI-created messages, while another portion was given standard, human-written messages with no personalization.��

After reviewing the message, participants were asked to rate it on six different metrics—clarity, trust, relevance, influence, confidence and certainty—that are common parameters known to influence a person’s behavior and decision-making intentions in disaster scenarios.��

How AI could level the playing field��

The results showed the clearest discrepancy between sociodemographic groups in relation to the standard alerts: women, people with lower education levels and older adults all rated the human-crafted alerts with higher trust and confidence than their counterparts (men, people with higher education levels and younger adults).��

Infographic: Amir Behzadan

But that did not hold true among the AI alerts. “Across gender, age and education, what you noticed was that as soon as we introduced AI personalization, those differences tended to diminish,” Behzadan says.��

The study also explored how the type of disaster might influence people (e.g., whether a tornado warning might be taken more seriously than a flood warning). “We found that the type of hazard is not that relevant to how people perceive the quality of the alert,” Behzadan says.��

While it does seem as though the personalized alerts could engender more trust and inspire more action among all sociodemographic groups, Behzadan’s team does not recommend abandoning the use of conventional messaging formats.��

“We're thinking that these two should work together to augment each other,” Behzadan says, adding that future personalized alerts could target specific subgroups within the population that are more likely to trust the customized messages.

Along those lines, Behzadan has already secured additional funding from the Natural Hazards Center and �鶹��Ѱ�����Boulder’s Research and Innovation Office (RIO) to pursue work on the potential to use AI to translate disaster alerts into American Sign Language (ASL) for Deaf and Hard-of-Hearing individuals.��

Given that ASL is a completely different language than English, many people in the Deaf and Hard-of-Hearing population would have difficulty reading and understanding disaster alerts transmitted in plain written English (and some may not be able to read it at all), Behzadan says, adding, “And if they can't understand it, they can't act upon it!”��

The trust issue��

Any broad use of AI is bound to face complications, especially when human lives are at stake.��

For this study, the use of AI was not disclosed to participants so the team could gather more realistic reactions to the messaging. “We knew from past research that some people may think that AI-generated content is less legitimate and may have bypassed human agency,” says Behzadan, adding that disclosing the AI use most likely would have skewed people’s perceptions.

But in practice, FEMA and/or state and local disaster agencies would most likely be required to disclose their AI use. So, while personalized alerts could inspire more trust in the short term, the use of AI may inadvertently lead to longer-term distrust of generated alerts, Behzadan says.

Another hurdle in deploying more personalized disaster messaging? The personal information required from recipients.��

“There's always a trade-off between privacy and accuracy,” Behzadan says. “If you want a highly accurate AI or technology output that genuinely reflects your specific situation, needs or expectations, you generally have to provide it with some additional context and personal information. Then, the real question is: How much personal information are you willing to give up in exchange for greater accuracy and personalization?”

Did you enjoy this article?����Passionate about natural hazards research?��

��