Correction: The realism of behavioral theory-based vs. non-theory-based AI agents during a simulated infant formula shortage
Linda Desens, Brandon Walling, Rhys O’Neill, Vanessa Howard, Mary Giammarino, Denise Scannell, Anya Kemble, Taylor H. Wilkerson, Nyalok Nhial, Sara Beth Elson, Maureen Leahy, Scott Rosen · Frontiers in Artificial Intelligence · 2026
As autonomous agent/digital twin (DT) systems become integrated into public policy and health system planning, the credibility of agent behavior is critical. This study provides evidence for policymakers and researchers regarding the realism of behavioral theory-informed autonomous AI agents in crisis situations such as food shortages.The ability to simulate the behavior of households related to health and social systems, and to have confidence that those simulations are realistic ones, would provide an important capacity for policymakers. However, in order to trust the simulations, there is a critical need to assess the accuracy of simulated agents during a crisis. Grounding autonomous agents in behavioral theories could improve their capacity to reflect realistic human decision-making. Our research addresses existing gaps in the literature by making the following contributions:1. Assessing the realism of autonomous agents that are grounded in behavioral theory, comparing them with agents that are not. 2. Building an autonomous agent/DT simulation related to food access.We use this research question: "Do participants perceive theory-based AI agents as behaving more realistically than non-theory-based agents?"This study highlights the importance of incorporating behavioral theory in the design of AI agents. AI agent modeling and simulation should advance from a rule-based decision tree to agent designs that are grounded in empirically validated behavioral theory. To achieve greater realism, agents should be designed with socially adaptive mechanisms that enable them to adjust their strategies based on evolving human norms, values and behaviors to reflect realistic human behavior (Rahwan et al., 2019) and internal processes such as cooperation that reflect human behavior (Diau, 2025). Agents should include reflective reasoning, goal representations, memory, and behavioral theoretical grounding to maintain consistency and behavioral coherency over time (Park et al., 2023). Capturing these nuances in a simulated environment is essential to ensure that autonomous AI agents realistically reflect the decision-making patterns of the populations they represent. In this study, incorporating well-established behavioral theories into agent design ensures that simulated actions are grounded in how people react during a crisis.This study models the behavior of households through the use case of an infant formula shortage. Food security is shaped by an interplay of policy, economic, geographic, and social systems (Roggio, 2019;Sawyer et al., 2021), as was demonstrated during the 2022 infant formula crisis. Families with infants are uniquely susceptible to disruptions across these systems, a concern that has drawn the attention of policymakers (U.S. Department of Health and Human Services, 2025; United States, 2022).The use of autonomous agents and DTs to simulate health and social systems has been growing, including in contexts directly related to our research, such as emergency response and urban Deleted: . ¶ ©2025 The MITRE Corporation. ALL RIGHTS RESERVED health (Jiang et al., 2022;Khan et al., 2023;Park et al., 2025). Review articles emphasize that DT use in these areas is in its infancy, and they point to technical, ethical, financial, and other barriers (Katsoulakis et al., 2024;Ringeval et al., 2025). Despite these challenges, promising case studies have validated the accuracy of DT/autonomous agent simulations (Bilal et al., 2025;Jiang et al., 2022;Park et al., 2025;Von Hoene et al., 2025) or provide a conceptual model for their use in health emergencies (Rodríguez-Aguilar, R., & Marmolejo-Saucedo, J., 2020). The Bilal study (2025) is noteworthy in that it demonstrated predictive validity of an epidemic DT/autonomous agent model, comparing its results to observed COVID-19 data: the model closely matched actual trends in cases, hospitalizations, and deaths (Bilal et al., 2025). Assessing the accuracy of simulations is imperative to their future use: as Park et al. state, "we must develop tools and methodologies so [policymakers] know when they can, and can't, trust these simulations" (Park et al., 2025). Some validation methods, such as Immersive Face Validation, are primarily used to judge the believability of autonomous agents that appear in human form on-screen, for example in videos or games (Bogdanovych et al., 2016;Louloudi & Klügl, 2012). Other methods focus on agent behavior rather than visual aspects. For example, the Virtual Overlay Multi-Agent System uses supervisory agents that monitor the simulation's adherence to predefined behavioral constraints set by human experts (Niazi, 2017). The Standardized Test Suite, which involves human success/failure ratings on agent-generated continuations of real human interaction scenarios, reveals alignment with human judgments (Abramson et al., 2022).Other research that assesses the realism of autonomous AI agents using human raters focuses on specific contexts. For example, a growing body of work assesses the accuracy of autonomous agents for clinical decision support by comparing them to the ratings of expert physicians (Hayat et al., 2025;Ringeval et al., 2025). Park et al.'s 2023 study is similar to our research, in that it uses human raters as participants to assess agent decisions and actions (rather than visual appearance) in simulated non-clinical environments. Park et al. (2023) created 25 generative agents that stored and synthesized memories and were embedded in a sandbox city environment. The researchers "interviewed" agents, asking questions across five categories (self-knowledge, memory, plans, reactions, and reflections). Participants then compared agent interview responses created under different agent types (full, no reflection, no reflection/plan, no reflection/plan/observation, and human baseline) and rated believability across the same five question categories. Results that the generative agent ability to and the behavior (Park et al., 2023). Our research on this the of agents to in with behavioral the use of autonomous agents in simulation into such as crisis public and social policy, that agent behavior realistic human critical. studies have grounded agent in behavioral validation focus on internal consistency or rather than and behaviors (Bogdanovych et al., This study from the Health and the of into autonomous agent theories were they provide The MITRE Corporation. ALL RIGHTS RESERVED for how people decisions under perceive and to social and which are during such as food studies have integrated behavioral theory into agent or of agent et al.'s of existing research (2025) used to that agent behavior over and behavioral et al., 2025). this is a for the of decision-making and in autonomous AI agents over it a model that behavioral theory into agent study that so on using modeling to simulate behavioral and in the of The model simulated of how in agents they are in that a of agents over their environment to their a of Our study into the use of behavioral theory for agent incorporating behavioral theories and using human participants as raters for The behavioral theories used in this study are in are in our for and The is used by participants to assess the behavioral realism of the autonomous agents. of of which are based on theoretical from and as in 2. were they directly to and and are and by participants The in with from the Health and the of as in the The were the for the behavioral ratings used by the participants The is based on an that real decision contexts from the 2022 infant formula & et al., et al., For example, of the agents to the same to formula crisis decision-making or an Other agents support from social when they were to formula from As a example, agents their of infant formula was of this study was to empirically theory-based autonomous AI agents behavioral realism as by in to autonomous AI agents theoretical the following research question and participants perceive theory-based AI agents as behaving more realistically than non-theory-based AI agents using behavioral theories realism study a design the realism of autonomous agent behaviors under agents that are grounded in behavioral theory and non-theory-based agents The was the agent theory-based The were the realism ratings for of the of and of from the 2022 formula including actions and the constraints on formula our The 2022 formula was by the of a for of the formula and was by related to the COVID-19 2025). constraints a in the formula shortage. for for and for were it for to the infant formula et al., and such as in food or to to to the for et al., on the other as to formula through and et al., et al., of and to the 2022 infant formula the types of actions into our actions were into behavioral by responses and patterns from the 2022 and them the behavioral theory For example, responses from the 2022 to to for it et al., et al., and it through and including food et al., et al., et al., Other 2022 actions the of formula that was an of our The MITRE Corporation. ALL RIGHTS RESERVED they were to used including formula or it with such as or infant formula & et al., and by in 2022 were used to the crisis For example, strategies based on households were more to formula than in a for households compared to for other households et al., households were more to have the to to over a of these households or more and more than to infant formula et al., In of of time and to formula et al., households constraints that other households they were to using the Food and of to of or of infant formula during the (U.S. 2025). were more than to use & et al., of that their infant to in they as et al., behaviors how and shaped responses during the which has a food than the as a provides a for our Food is as or of and or or ability to in socially (U.S. Department of food and and have and health for et al., et al., food in was in 2023 compared to the of and food was compared to the of has of the in and the was compared to the of in 2023 (U.S. in the are as and to food (U.S. Department of and were food (U.S. Department of the interplay of that were for our of raters was from a This was based on a a and a of raters was ©2025 The MITRE Corporation. ALL RIGHTS RESERVED for a et al., participants a behavioral health or public health no in the behavioral and social or was The were designed to be by behaviors were simulated in a twin of during a infant formula twin to as closely as the to provide the agents with for infant formula during the crisis To provide the twin and AI accuracy the of the real were into their state, and were and into and of the twin of barriers for the AI twin five in specific were drawn as for the agents would on and to formula in the 2023). The to formula in were formula and for was integrated into the of the simulation modeling to more the and of infant formula from to The and and that the agent of infant and were based on 2022 (U.S. Food and 2023). The model was designed to disruptions and such as and emergency To model the for based on and behavior when of were into the and were used as policy to as realistic to infant formula were into this of the twin to provide realistic barriers for with infant formula This integrated 2025) to provide realistic for the agents, including and across the a and model were from (U.S. 2025) and integrated into the This to formula for agents. and that infant formula in the specific on realistic and were when with This during emergency response used the to assess the realism of the behavior across the by a on the for of the of and of was to the consistency realism ratings to participants on how to and the of that the response was with how a real under similar crisis would be to a of behavior that was realistic For example, on the an agent an or a of was rated a In formula was was rated and a were for the of of to ensure across were an that the study and the of their the participants were a to the form that and the a of actions over the simulation and include theoretical from for a of The theory-based agents behaviors of theoretical from and and non-theory-based agents that with no theoretical an of in a order using the To order participants were into with different This ©2025 The MITRE Corporation. ALL RIGHTS RESERVED that in in the an of across theory-based agents in this study were designed to simulate realistic human decision processes by from and The decision was by the agents to reflect on the strategies that their and in response to et al., 2023). and were into the theory-based agents agent that used for were embedded into a of and social the goal of infant and were as strategies social to food or was on the theory-based and the of the agents a rule-based decision tree social or adaptive This a for Agents a decision are from and social then by that To a is during behavioral theoretical is and agent for the of the of the is in example how a theory-based agent into a decision using behavioral theory the the the agent infant formula is its is The agent and from households similar are as rather than the Health the of formula and the of and the of and work constraints behavioral for to other households formula through the of The MITRE Corporation. ALL RIGHTS the internal the The on infant its from through social and The agent then an with this such as a food or to constraints such as and agent behavior to over time in response to grounded in behavioral theory, actions that reflect realistic human decision-making under crisis rather than of to behavioral theory and decision is in the non-theory-based agents decisions based on rule-based The agents a decision tree that a when was and was the agent a and when was the agent would for The decision tree different under different of and However, the non-theory-based agents have that social or adaptive This created a for the to which into the agents realism in simulated crisis For a of theory-based non-theory-based agents, non-theory-based agents were designed as a rather than as a realistic or predictive was to provide a that rule-based used in the study to the of behavioral theory to realism decision tree was for validity by experts with in public health food systems, and crisis to ensure that the actions were and of or was as the study realism agent types rather than behavioral for agents were using from the Food and (U.S. Department of 2025) and the (U.S. 2025) to and the the of need or ability to formula during or the need to a time of and to different the System (U.S. for and 2025) and health were to agent or adaptive with the to to from and (U.S. Department ©2025 The MITRE Corporation. ALL RIGHTS RESERVED of Food and 2025) and to which and are through agents to areas and are by based on realistic and for crisis simulation a set of agents that and support mechanisms that would be during an infant formula shortage. As an of the into the these systems agents food and or agent was designed with and so that their behavior with how these systems realistically including these agents, the model the households and the responses that during a agent types were created to model crisis response The agent food from Health and Human 2025) and 2025) to model to agents emergency formula when households were in the The twin simulated to be to the 2022 infant formula in five in five were they a of food and The model integrated This simulation was using and with and to support and of the agent decision-making validation ensures for for and system the of agents the agent the twin the research question and comparing realism for the theory-based compared to non-theory-based agents, a was the compared realism for theory-based non-theory-based For to the realism by of the a was used to the in ratings or were as for and for participants rated the theory-based and non-theory-based AI agents for realism across the for As in theory-based agents realism compared to non-theory-based agents the confidence for these with the of the theory-based the of the non-theory-based evidence of a in realism theory-based and assess the consistency of ratings using the were for the ratings across and was that the realism across and were and by agent was for ratings than for theory-based The results consistency raters across values for theory-based reflect the greater behavioral by from behavioral theory. Some agent behaviors were in In agents to decision consistency across for are in and the realism and confidence for theory-based and non-theory-based agents across for of the realism by for theory-based non-theory-based agents and compared to non-theory-based agents, greater realism and consistency in agent the research question and and participants the theory-based agents to be more realistic when compared to non-theory-based agents, a was used to the realism across The results that the in realism across was was with and as of were to the that the of was for and for the interaction were in the to the same as the a for a for and a interaction results are in the of the of using were over the from the of other the were Participants rated the theory-based agents as more realistic than non-theory-based agents for the and the of in the and of were for a of the of agent for agents, the of from the and from the of For non-theory-based agents, the from the of other were that the interaction was primarily by realism ratings for agents in the and of and by specific the agent or assess the of our the a using confidence for the theory-based and non-theory-based across The confidence that the observed are with confidence for the of and of results provide evidence that the of theory-based agent design is across with the constraints of our results of the study provide support for the that the behavior of autonomous AI agents grounded in behavioral theory is as more realistic by participants compared to the behavior of agents that the same theoretical This addresses a in the literature the validation of twin agents, accuracy to behavior as by the of realism decision of and use of social theory-based agents realism The in realism ratings theory-based and non-theory-based agents was with a that the of behavioral theory the realism of ©2025 The MITRE Corporation. ALL RIGHTS RESERVED agent behavior in food The of this was through confidence which demonstrated that theory-based agents were rated across when for on these for agent and realism as as a interaction agent and This that theory-based agents non-theory-based agents in realism, the of this across the different of the realism was for the use of social the importance of modeling social and adaptive behaviors during crisis and including that the and behavior of social with research that agents are of social and processes and that human raters their behavior more and in simulated (Park et al., of the study the of as a design and from and into agent and with these this study a for and behavioral realism in autonomous agent This addresses a critical in the rule-based theory-informed agent design that the of human decision-making and behavior under crisis results have for the and of autonomous AI agents in twin related to food and crisis the that theory-based agents are rated as more realistic by participants the importance of behavioral theory into agent This or accuracy and that realistic behavior is critical for simulations that are and be by This that behavioral theories into the AI design the of decision support tools in of these to the of simulation models in public emergency and social As twin to the realism of agent behavior is essential for simulation agent design to ensure that simulation results have for for crisis a question of trust in Our that grounding agents in behavioral theory behavioral agents decisions how real people and their actions become to the experts must on simulation when are or public health need agents that decisions based on social they and theory-based provides that policy contexts agent behavior with validated theories and for twin This could of decision support in public health and emergency for realism and simulation used in this study be as this However, these be to researchers to and are to this participants agent realism through that an the infant formula shortage. The agent are and to for during the constraints that use which would the participants to assess the agents a agents decisions how real people and their actions become to the experts must on simulation compared to the agent the are that of how the agents were on the realism of agents, assess theory-based agents would to behavioral behavioral theory was used to agent the goal was to assess realism, that more or realistic agent models be to behavioral this on the and embedded in the rather than agent studies should how be designed to support or behaviors in response to crisis our use case was and on infant formula shortage. Other public health crisis or emergency different is in the theory-based compared to non-theory-based for the which that more work to be to the for realism theory-based However, the realism ratings across and of realism was used to theory-based to non-theory-based for was which the ratings across raters provides a for This is in with in social and behavioral ratings are used to and the of (Bogdanovych et al., et al., study which for or the design by through and the confidence the of our the to of or the of a was based on it for the across or the interaction studies should to improve of and enable more modeling of in realism our have the specific This research on participants to an infant formula which is a crisis populations The to which these would in populations or contexts infant social support trust in and across contexts and actual behavior and of agent work should behavioral theory-based agents in and across different types of and under of to the validity of these behavioral theories and the of agent AI model with the grounding of the autonomous agents with behavioral theories and using human for the this the of which from such as theory and in the simulation environment that across populations or The study behavioral and to of these However, future research should include as of twin et al., et al., study that when autonomous AI agents are designed with behavioral theory, they are as more realistic in how they decisions and with In the simulated infant formula theory-based agents more closely with human than non-theory-based agents. This the importance of behavioral theory into the of have that it is to and the realism of these agents in a a for behavioral a simulation that people trust to reflect under crisis credibility is critical are to be used in and The MITRE Corporation. ALL RIGHTS RESERVED this work a for the and of AI agents that are by behavioral theory. studies the to different and more social As policymakers and researchers on simulation for planning, this research provides an and a for that autonomous agents grounded in human are for future research that could on this for future research is to a simulations and behavioral Park et al. (2023) agents for of time to set for how they is validation comparing simulated agent behavior to from real during the infant formula of the from human more in the design of the agents their realism with is to autonomous agents a decision-making that uses behavioral drawn from behavioral research et al., et al., In that research, the decision were to and research that decision-making when with health or et al., et al., important for future research involves across the this to scenarios, systems, and urban contexts would its and for For example, use models to responses during or agents would provide more realistic simulations of and assess patterns In could simulate decisions adherence and could use theory-based agents to simulate responses to or with and modeling the use of human raters and the provide a and for agent behavioral theory with agent behavior and a for the design and of AI agents and This addresses a in the validation have on internal or rather than the realism of agent behavior by human raters (Bogdanovych et al., agent design and with behavioral theory, this a for the validation across and that autonomous agents are and as Health and to The in the of the or and the no is to the to internal or or that to or behavior & et al., et al., 2012). The and confidence in ability to a specific behavior & et al., et al., 2012). barriers and The in the or barriers that behavior or health and the that the behavior or health could to health & et al., et al., 2012). to The to internal or or that to or behavior by which to the social and the specific behavior & et al., et al., to a to a behavior 2020). behavioral The or has over the specific by that or behavior et al., & et al., the behavior The the of a specific by behavioral specific to the of a behavior & et al., by which are based on or of the in to an to the to which an to ©2025 The MITRE Corporation. ALL RIGHTS RESERVED with & et al., to behaviors through and by and et al., et al., 2023). the and of the of a specific by social et al., & 2020). behavior in The of a specific behavior by a model, is the are when models are to be the et al., and to behavior that how the specific behavior are to specific behaviors when or et al., 2. and real and The for crisis decision-making the ability to and in response to the infant formula crisis as as their ability to strategies and to the is by the of the their to the and the to such as a of infant formula on behavioral which agent based on and et al., 2020). and how it & assesses agents a or confidence in to a ©2025 The MITRE Corporation. ALL RIGHTS RESERVED or on behavior barriers or behavioral the their ability to a behavior or actions et al., 2020). and agents adjust their behavior based on and the results & theoretical for the of actions that agents actions their constraints as as with their and the and barriers they actions the behavior assesses a an as or & 2017). behavior in that behavior realistic constraints and & of theoretical for this agents with or that reflect real human to through highlights that social an agent to on and in with and the infant formula by has of uses to for an and to for support the strategies to maintain with to and a that which have formula has a the of the infant formula formula making is using and with to and uses to with formula in the of the has to formula to future has to and social on and in with and the infant formula decisions on how formula has the has on so to formula so to the and to the the same when formula a when has the of the infant formula has to to no a is of and has a formula the has on when for the