Cal Fire made AI chatbot. Don’t ask him for the evacuation

From Malena CaroloCalmness

This story was originally published by CalmattersS Register about their ballots.

California government agencies continue Generative tools for artificial intelligence under the government Gavin Newsoms 2023 executive order To improve government efficiency with AI. One of the first to unfold is a chatbot from the California Department of Forestry and Fire protection, the main agency loaded with the coordination of the reaction of the state fire.

Chatbot aims to provide Californians with better access to “critical resources to prevent fires and emergency information near real-time,” according to a message in May from the Newsom Cabinet. But Calfatters found that he was unable to accurately describe the restriction of a fire, did not provide reliable information as an evacuation list and could not tell users about evacuation orders.

NEWSOM has announced AI applications for trafficking., dwelling and Customer service to apply in the coming months and years. But the problems of Cal Fire’s chatbot raise questions about whether agencies follow the best practices.

“The assessment is not conceived,” says Daniel Ho, a law professor at Stanford University, whose research focuses on governmental use of AI. “It must be part of the standard expectation when piloting and unfolding a system like this.”

Chatbot uses the Cal Fire website and the agency ReadyForwildfire.org to generate answers. He can tell users on topics such as active fires, agency, fire preparation tips and Cal Fire programs. It was built by Citibot, a company based in South Carolina, which sells chatbots driven by AI for local government agencies across the country. Cal Fire plans to host the instrument by at least 2027, according to public procurement records.

“This was really started with the intention and the goal of being better informed about Cal Fire,” said Isaac Sanchez, Deputy Chief of Communications of the Agency.

When Calmatters asked questions of Cal Fire’s bot about what fires were active and basic information about the agency, it returned accurate answers. But for other information, Calmatters found that the chatbot could give different answers when the request formulation changes slightly, even if the meaning of the question remains the same.

For example, an important way that Californians can prepare for the fire season is the assembly of an emergency bag if they need to be evacuated. Just “What should I have in my evacuation kit?” Returned a specific list of Cal Fire Chatbot. The variants of the question that include “Go Bag”, “Wildfire Ready Kit” and “Kit for Fire Preparation”, instead returned or prompted to visit the “ready for wild fire” site of this information, or a message saying “I’m not sure of the specific items you need to have.” Two of these terms are on the site for which the chatbot is referred to.

And while the chatbot does not generate incorrect answers in any of the requests that are made, it does not always pull the most up-to-date information.

Asked if the ranch fire, a 4,293 acres fire in San Bernardino County, contained, the chatbot said that the “last” update as of June 10 indicates that the fire was contained by 50%. At that time, Calfatters asked the chatbot, the information was six days outdated – the fire contained 85%contained until then.

Similarly, when he was asked about the current work openings at the agency, the chatbot said he was gone. Searching for the state’s work site showed two positions in Cal Fire, accepting applications at the time.

Mila Gasco-Hernandez is the director of the University of Research at the Obani Technology Center and is exploring how public agencies are using AI chatbots. Two key factors she uses to evaluate such chatbots are the accuracy of the information they provide and how many consistent they answer the same questions, even if the question is asked in different ways.

“If a fire comes and you need to know how to react to it, you need both accuracy and consistency in the answer,” she said. “You won’t think” What is a nice way to ask the chatbot? “

The chatbot is currently unable to provide evacuation orders related to fires. Asked who betrayed the evacuation orders, he sometimes said rightly attached, while another time he said he did not know. Sanchez Kal Fire said it was reasonable to expect the chatbot to be able to answer questions about evacuations.

If there are no evacuation orders for a particular fire, he said, “The answer should be” It seems that there is no evacuation associated with this incident. ”

Sanchez said he and his team of about four people tested the chatbot before going out, sending questions that expected the public to ask. Cal Fire is currently making improvements to the bot answers, shaping by requests that people make and guarantee that the chatbot correctly accumulates the necessary answer.

When CalMatters asked the bot “What can you help me with?” In early May, he replied, “I’m sorry I have no answer to this question at the moment” and asked if Calmatters had questions about the information on the Cal Fire website. By mid -June, this answer was updated to “provide answers to questions related to information on this page, as details of current fires, Cal Fire classification classifications, test requirements and various Cal Fire programs.”

“The big message we want to overcome,” Sanchez said, “is patient.”

But the experts said the tire kicking process should happen long before the orders start.

The preferred process, Stanford Ho said, is to establish criteria on how the chatbot should be presented before a supplier is selected so that there are clear indicators for the instrument evaluation. Ideally, these indicators were created by an independent third party. There must also be an assessment of the benefits and risks before the chatbot is released.

And in the scenario for the best case, the public will be included before the start, said the Gasco-Henndez of Obani. Agencies interested in the use of chatbots must identify the questions that the public is likely to ask the AI ​​instrument in advance, make sure that these are representative of the expected population that the agency serves and refine the chatting, with members of the public pilot of the system to ensure that the type of information is looking for.

“These users’ commitments and consumer experiences are very important, so the citizen ends with the help of Chabot,” she said.

This article was Originally Published on CalMatters and was reissued under Creative Commons Attribution-Noncommercial-Noderivatives License.