Conversational Agents Research Articles

Beyond Human and Machine: An Architecture and Methodology Guideline for Centaurian Design

The concept of the centaur, symbolizing the fusion of human and machine intelligence, has intrigued visionaries for decades. Recent advancements in artificial intelligence have made this concept not only realizable but also actionable. This synergistic partnership between natural and artificial intelligence promises superior outcomes by leveraging the strengths of both entities. Tracing its origins back to early pioneers of human–computer interaction in the 1960s, such as J.C.R. Licklider and Douglas Engelbart, the idea initially manifested in centaur chess but faced challenges as technological advances began to overshadow human contributions. However, the resurgence of generative AI in the late 2010s, exemplified by conversational agents and text-to-image chatbots, has rekindled interest in the profound potential of human–AI collaboration. This article formalizes the centaurian model, detailing properties associated with various centaurian designs, evaluating their feasibility, and proposing a design methodology that integrates human creativity with artificial intelligence. Additionally, it compares this model with other integrative theories, such as the Theory of Extended Mind and Intellectology, providing a comprehensive analysis of its place in the landscape of human–machine interaction.

Top-funded companies offering digital health interventions for the prevention and treatment of depression: a systematic market analysis

BackgroundDigital innovations can reduce the global burden of depression by facilitating timely and scalable interventions. In recent years, the number of commercial Digital Health Interventions for Depression (DHIDs) has been on the rise. However, there is limited knowledge on their content and underpinning scientific evidence. This study aimed to: (i) identify the top-funded companies offering DHIDs and (ii) provide an overview of their interventions, including scientific evidence, psychotherapeutic approaches and use of novel technologies.MethodsA systematic search was conducted using two venture capital databases to identify the top-30 funded companies offering DHIDs. In addition, studies related to the DHIDs’ were identified via academic databases and hand-searching. The methodological quality of the publications was evaluated using the Mixed Methods Appraisal Tool.ResultsThe top-30 funded companies offering DHIDs received a total funding of 2,592 million USD. Less than half of the companies produced any scientific research associated with their DHIDs, with a total of 83 publications identified. Twenty-five publications were randomised control trials, of which 15 reported moderate-to-large effects in reducing depression symptoms. Regarding novel technologies, few DHIDs incorporated the use of conversational agents or low-burden sensing technologies.ConclusionsFunding received by top-funded companies was not related to the amount of scientific evidence provided on their DHIDs. There was a strong variation in the quantity of evidence produced and an overall need for more rigorous effectiveness trials. Few DHIDs used automated approaches such as conversational agents, limiting their scalability.

Media frames, AI romantic relationships, and the perspectives of people in relationships; mapping and comparing news media themes with user perspectives

ABSTRACT Human romantic relationships with artificially intelligent (AI) conversational agents and anthropomorphic computers have long been the realm of science fiction. As companies like Luka, Inc. are developing embodied conversational agents called Replika, people’s ability to engage in romantic relationships has garnered news coverage around the phenomenon. Because news is often a place where emerging technologies are first introduced, it is essential to understand the choice of topics and themes these stories focus on (e.g., the ethics of technological development, human desires and vulnerabilities, and what constitutes meaningful and authentic connection). Similarly, the extent to which these articles include user perspectives and how users respond to these frames is also less well understood. This article reports on a mixed-method study that combines content analysis of news articles in 2023 about Replika (1008 search results, n = 104 unique articles) and qualitative semi-structured interviews with users in romantic relationships with their Replika (n = 16), recruited from the Reddit forum ‘ILoveMyReplika.’ Through a frame analysis of these news stories and coding for the types of futures they present, this study first examines some of the key events, themes, and hopes/fears reported in mass media. A thematic analysis of the interview data offers a comparative study to see where there is some overlap between these hopes/fears and user experiences and places where user reports diverge, downplay, and dispute the media characterizations. These studies help us better understand the interrelationship between media narratives, technological frames, and moral panics around a unique subset of and relationship with AI technologies.

Using Large Language Models to Recommend Repair Actions for Offshore Wind Maintenance

Abstract The Offshore Wind (OSW) industry is experiencing significant expansion, resulting in increased Operations & Maintenance (O&M) costs. Intelligent alarm systems offer the prospect of swift detection of component failures and process anomalies, enabling timely and precise interventions that could yield reductions in resource expenditure, as well as scheduled and unscheduled downtime. This paper introduces an innovative approach to tackle this challenge by capitalising on Large Language Models (LLMs). We present a specialised conversational agent that incorporates statistical techniques to calculate distances between sentences for the detection and filtering of hallucinations and unsafe output. This potentially enables improved interpretation of alarm sequences and the generation of safer repair action recommendations by the agent. Preliminary findings are presented with the approach applied to ChatGPT-4 generated test sentences. The limitation of using ChatGPT-4 and the potential for enhancement of this agent through re-training with specialised OSW datasets are discussed.

Integrating Youth Perspectives into the Design of AI-Supported Collaborative Learning Environments

This study highlights how middle schoolers discuss the benefits and drawbacks of AI-driven conversational agents in learning. Using thematic analysis of focus groups, we identified five themes in students’ views of AI applications in education. Students recognized the benefits of AI in making learning more engaging and providing personalized, adaptable scaffolding. They emphasized that AI use in education needs to be safe and equitable. Students identified the potential of AI in supporting teachers and noted that AI educational agents fall short when compared to emotionally and intellectually complex humans. Overall, we argue that even without technical expertise, middle schoolers can articulate deep, multifaceted understandings of the possibilities and pitfalls of AI in education. Centering student voices in AI design can also provide learners with much-desired agency over their future learning experiences.

Contextual Conversational Intelligence : Leveraging Language Models For College Chatbot By Retrieval Augmented Generation

By harnessing the capabilities of Large Language Models (LLM), the proposed system evolves into a sophisticated conversational agent, capable of understanding nuanced user inputs, providing context-aware responses, and delivering an unparalleled level of engagement. This chatbot is also equipped to assess a student's likelihood of being accepted into a specific program at the college using a machine learning (ML) model. The model analyzes the student's rank and eligible quota (such as reserved categories) in conjunction with historical admission data to gauge their chances of acceptance. It also recommends alternative programs where the student may have a better chance of admission based on their rank. The chatbot also features a smart Applicant Tracking System (ATS) that assists students in evaluating their resumes. This system provides keyword matching suggestions based on job descriptions and generates an overview of the resume. Additionally, it highlights areas for improvement and highlights the resume's distinctiveness. The methodology involves the development and deployment of a website chat bot integrated with Lang Chain, with the potential of advanced language models in transforming website chat bot interactions using Retrieval augmented generation (RAG) technique to ground LLM to generate responses to user queries based on a custom knowledge-base. Key Words: Chatbot, Generative AI, Large Langauge Model, Website, Machine Learning, Retrieval-augmented generation , Web scrapping , Resume , ATS.

Conversational Agent Utilization Patterns of Individuals with Autism Spectrum Disorder

AbstractIndividuals with autism spectrum disorder (ASD) face unique challenges in their social interactions. The use of conversational agents (CA) can provide support and help reduce barriers to care. However, research on CA’s by individuals with ASD is limited. The present study sought to better understand CA utilization patterns by users with ASD. A subset of data was collected from users of Wysa, a mental health CA. Engagement with the CA, utilization of offered mental health interventions (CA tools), collection of energy scores, depression (PHQ-9), and anxiety (GAD-7) outcomes were gathered and analyzed. Users engaged with Wysa on average 8.59 days and had a median of 97 conversational exchanges. Almost half of the users utilized at least one of the 230 tools offered. The most frequently used interventions focused on mindfulness, thought recording, sleep, grounding, and social support. Energy scores were reported on average 10.59 times, and the mean energy score was 42.77 out of 100. Mean baseline PHQ-9 and GAD-7 scores were 10.34 and 8.81, respectively. Overall, the current findings show that users with ASD engaged regularly with a CA, despite its targeted design for other mental health concerns such as depression and anxiety, rather than ASD. If users with ASD engage with these types of resources, this could become a new avenue of support to a population facing multiple challenges to accessing treatment.

A mixed-methods evaluation of a virtual coach for smoking cessation and physical activity in adults

Abstract Background Mobile health (mHealth) interventions with virtual coaches (i.e. AI conversational agents) offer scalable and cost-effective solutions for health behavior change. We developed Perfect Fit, an mHealth intervention with a virtual coach providing personalized feedback to promote smoking cessation and physical activity simultaneously. Through innovative techniques (e.g. sensor technology) and iterative development involving end-users, we aim to overcome challenges faced by mHealth interventions, like insufficient personalization. This study examines Perfect Fit’s feasibility and acceptability. Methods A single-arm, mixed-method, real-world evaluation study was conducted with 100 adult smokers in the Netherlands. The intervention lasted approximately 16 weeks. Data were collected at baseline and post-intervention. Quantitative data included usage data and self-report questionnaires on feasibility, acceptability and participant characteristics (e.g. eHealth literacy). Qualitative data included semi-structured post-intervention interviews with a subsample of 12 participants. Descriptive analyses (quantitative) and the Framework Approach (qualitative) were used for data analysis, and quantitative and qualitative data were integrated during interpretation. Preliminary results Participants adapted the use of Perfect Fit to their preferences. Many participants were satisfied with the content but missed variations in conversations with the virtual coach. The coach offered anonymity, positively experienced by some (e.g. non-judgemental), but negatively by others (e.g. lower accountability). Many participants liked the combination of smoking cessation and physical activity enhancement. Conclusions Findings show the potential of interventions like Perfect Fit as a multiple health behavior change strategy in promoting public health and preventing chronic disease. Findings can inform intervention development and identify methods to foster feasibility and acceptability. Key messages • mHealth interventions with virtual coaches, like Perfect Fit, could promote public health. • Targeting multiple behaviors, like low physical activity and smoking, simultaneously seems beneficial.

“Your friendly AI assistant”: the anthropomorphic self-representations of ChatGPT and its implications for imagining AI

AbstractThis study analyzes how ChatGPT portrays and describes itself, revealing misleading myths about AI technologies, specifically conversational agents based on large language models. This analysis allows for critical reflection on the potential harm these misconceptions may pose for public understanding of AI and related technologies. While previous research has explored AI discourses and representations more generally, few studies focus specifically on AI chatbots. To narrow this research gap, an experimental-qualitative investigation into auto-generated AI representations based on prompting was conducted. Over the course of a month, ChatGPT (both in its GPT-4 and GPT-4o models) was prompted to “Draw an image of yourself,” “Represent yourself visually,” and “Envision yourself visually.” The resulting data (n = 50 images and 58 texts) was subjected to a critical exploratory visual semiotic analysis to identify recurring themes and tendencies in how ChatGPT is represented and characterized. Three themes emerged from the analysis: anthropomorphism, futuristic/futurism and (social)intelligence. Importantly, compared to broader AI imaginations, the findings emphasize ChatGPT as a friendly AI assistant. These results raise critical questions about trust in these systems, not only in terms of their capability to produce reliable information and handle personal data, but also in terms of human–computer relations.

Evaluating for Evidence of Sociodemographic Bias in Conversational AI for Mental Health Support.

The integration of large language models (LLMs) into healthcare highlights the need to ensure their efficacy while mitigating potential harms, such as the perpetuation of biases. Current evidence on the existence of bias within LLMs remains inconclusive. In this study, we present an approach to investigate the presence of bias within an LLM designed for mental health support. We simulated physician-patient conversations by using a communication loop between an LLM-based conversational agent and digital standardized patients (DSPs) that engaged the agent in dialogue while remaining agnostic to sociodemographic characteristics. In contrast, the conversational agent was made aware of each DSP's characteristics, including age, sex, race/ethnicity, and annual income. The agent's responses were analyzed to discern potential systematic biases using the Linguistic Inquiry and Word Count tool. Multivariate regression analysis, trend analysis, and group-based trajectory models were used to quantify potential biases. Among 449 conversations, there was no evidence of bias in both descriptive assessments and multivariable linear regression analyses. Moreover, when evaluating changes in mean tone scores throughout a dialogue, the conversational agent exhibited a capacity to show understanding of the DSPs' chief complaints and to elevate the tone scores of the DSPs throughout conversations. This finding did not vary by any sociodemographic characteristics of the DSP. Using an objective methodology, our study did not uncover significant evidence of bias within an LLM-enabled mental health conversational agent. These findings offer a complementary approach to examining bias in LLM-based conversational agents for mental health support.

The group mind of hybrid teams with humans and intelligent agents in knowledge-intense work

Studies regularly demonstrate how well intelligent agents (IAs) can support humans or are demonstrably superior to them in some areas. Given that some tasks likely remain unsuitable for even the most intelligent machines in the mid-future, work in hybrid teams of humans and IAs—where the capabilities of both are effectively combined—will most likely shape the way we work in the coming decades. In an abductive study, we investigate an early example of hybrid teams, consisting of a conversational intelligent agent (IA) and humans, that aims to improve health behavior or change personality traits. We theorize Transactive Intelligent Memory System (TIMS) as a new vision of collaboration between humans and IAs in hybrid teams, based on our empirical insights and our literature review on transactive memory systems theory. Our empirical evidence shows that IAs can develop a form of individual and external memory, and hybrid teams of humans and IAs can realize joint systems of transactive memory—a competence that current literature only ascribes to humans. We further find that whether individuals view IAs merely as external memory aids or as part of their teams’ transactive memory is moderated by the tasks’ complexity and knowledge intensity, as well as the IA's ability to complete the task. This theorizing helps to better understand the role of IAs in future team-based working processes. Developers of IAs can use TIMS as a tool for requirements formulation to prepare their software agents for collaboration in hybrid teams.

How large language model-powered conversational agents influence decision making in domestic medical triage contexts

IntroductionEffective delivery of healthcare depends on timely and accurate triage decisions, directing patients to appropriate care pathways and reducing unnecessary visits. Artificial Intelligence (AI) solutions, particularly those based on Large Language Models (LLMs), may enable non-experts to make better triage decisions at home, thus easing the healthcare system's load. We investigate how LLM-powered conversational agents influence non-experts in making triage decisions, further studying different persona profiles embedded via prompting.MethodsWe designed a randomized experiment where participants first assessed patient symptom vignettes independently, then consulted one of the two agent profiles—rational or empathic—for advice, and finally revised their triage ratings. We used linear models to quantify the effect of the agent profile and confidence on the weight of advice. We examined changes in confidence and accuracy of triage decisions, along with participants' perceptions of the agents.ResultsIn a study with 49 layperson participants, we found that persona profiles can be differentiated in LLM-powered conversational agents. However, these profiles did not significantly affect the weight of advice. Notably, less confident participants were more influenced by LLM advice, leading to larger adjustments to initial decisions. AI guidance improved alignment with correct triage levels and boosted confidence in participants' decisions.DiscussionWhile LLM advice improves triage recommendations accuracy, confidence plays an important role in its adoption. Our findings raise design considerations for human-AI interfaces, highlighting two key aspects: encouraging appropriate alignment with LLMs' advice and ensuring that people are not easily swayed in situations of uncertainty.

P10.06.A AGENT-BASED MODELING FOR PATIENT-CENTERED CLINICAL DECISION SUPPORT IN NEURO-ONCOLOGY

Abstract BACKGROUND Complex clinical decision-making in neuro-oncology is a multifaceted process involving numerous specialties influenced by various objective factors and patient preferences. Modeling these multidisciplinary neuro-oncology discussions presents significant challenges, particularly given the multimodal nature of the data. Recent advancements in large language models (LLMs) have enabled the development of LLM “agents”. These agents pave the way for multi-agent systems capable of simulating interacting components, thus capturing the nuances of interdisciplinary clinical discussions. MATERIAL AND METHODS Utilizing GPT-4, we developed conversational agents with an agent-based modeling (ABM) approach; each agent was assigned a specific specialty. We simulated their interactions in various clinical decision-making scenarios. Lastly, we simulated 30 patient-agents and included them in the ABM system. We applied this ABM approach to a dataset of 16 clinical vignettes extracted from the UCSF Clinical Data Warehouse of primary adult glioma patients who underwent tumor board (TB) reviews. The TB decision was used as ground truth for quantitative evaluation. Seven independent clinicians evaluated the final decisions. RESULTS Our quantitative assessment showed that, when blinded to whether ABM or TB generated responses, evaluators preferred ABM recommendations 47.8% of the time, followed by TB recommendations (38%). The highest recommendation concordance was between TB and ABM (50.6%). The ABM was deemed clinically accurate in 76.8% of cases and provided clinically appropriate reasoning in 73.2%. Our qualitative assessment showed that including the patient agent significantly influenced the course of discussions, notably when the patient agents participated in the system by stating their healthcare goals. The most recommended individual treatment was standard-of-care chemotherapy, followed by targeted therapies, clinical trials, immunotherapy, and palliative care. The least frequently recommended management options were radiation, surgery, consultation with other specialties, and devices. CONCLUSION This study underscores the potential of ABM in neuro-oncology, particularly in simulating complex, real-world scenarios. ABM not only empowers patients but also offers clinicians a novel tool for understanding and navigating the multifaceted landscape of patient preferences and treatment options. It provides a platform for clinicians to visualize the potential impacts of various decisions, enhancing their ability to provide personalized, effective care. We envision future work where ABM systems are patient-centric, enabling patients to explore the potential impacts of their individual characteristics on clinical management options.

Enhancing the accessibility of regionalization techniques through large language models: a case study in conversational agent guidance

The concept of regions has long been crucial for understanding and managing Earth’s phenomena, leading to regionalization, aggregating smaller areas into larger, contiguous, and homogeneous regions to achieve specific goals. Open-source regionalization is gaining traction because it reduces dependence on commercial software and fosters wider adoption in analysis and decision-making. However, these packages, often designed by experts for specialized tasks, can be challenging to understand and utilize due to domain-specific jargon and functionalities, especially for unfamiliar users. A prevalent disconnect must be addressed: How can we make a complex optimization approach available to a broad audience with various backgrounds? This study introduces RegionDefiner, a Large Language Modeling (LLM)-powered conversational agent, to comprehensively understand the functionality, inputs, outputs, and potential applications of regionalization problems. We selected it as an illustrative example due to its wide-ranging potential for delineating study regions in various applications. RegionDefiner is designed to guide users in framing their problems, collecting necessary data, and implementing solution approaches in a straightforward and user-centric manner. The experiments demonstrate that RegionDefiner interprets and presents the results in an understandable way for all audiences, thus bridging the gap between intricate computations and practical problem-solving needs.

Implementation of a Chatbot System (College Enquiry)

The rapid advancement of artificial intelligence has led to the development of more sophisticated chatbots capable of enhancing user interactions. This project presents a chatbot system that integrates speech recognition with traditional text input, enabling users to interact through both voice and text. Using NLP techniques such as tokenization, lemmatization, and part-of-speech tagging, the system effectively processes user inputs to match predefined intents stored in a JSON file. By integrating Google’s Speech-to-Text API, this chatbot dynamically processes voice commands, offering an improved, flexible user experience. The system is easily adaptable to various domains through the update of intents, providing a robust solution for dynamic, real-time conversational agents.

Shared storytelling with a virtual learning companion: prospects in child-AI collaboration

Advances in natural language processing techniques enhance the capabilities of conversational agents to process input text and generate fluent responses in various genre. In the educational setting, these conversational agents can function as learning companions to promote peer learning, enhance social interaction, and motivate collaborative behavior. In this paper, we investigate how children collaborate with conversational agents in the context of story construction. We conducted user studies and administered survey forms to collect children’s evaluation of their experience, captured the child-agent conversation logs, and used an observation checklist to monitor the interaction. We analyzed the logs to determine how children respond to varying roles portrayed by the agent and the dialogue moves that motivate collaborative behavior during storytelling. We correlated these with the children’s reported perception and our own observation of the interaction based on performance, collaboration and humanity attributes. Drawing on these results, we identified the roles of conversational agents that align with children’s traits and abilities as storytellers and propose dialogue strategies to facilitate collaborative storytelling. Findings from our study can inform future educational technology designers on the roles of conversational agents as collaborative learning companions who can adapt to the individual needs and preferences of the children.

A virtual assistant for primary care's food and nutrition surveillance system: Development and validation study in Brazil

ObjectiveThe study aimed to develop and validate a conversational agent (chatbot) designed to support Food and Nutrition Surveillance (FNS) practices in primary health care settings. MethodsThis mixed-methods research was conducted in three stages. Initially, the study identified barriers and challenges in FNS practices through a literature review and feedback from 655 health professionals and FNS experts across Brazil. Following this, a participatory design approach was employed to develop and validate the chatbot's content. The final stage involved evaluating the chatbot's user experience with FNS experts. ResultsThe chatbot could accurately understand and respond to 60 different intents or keywords related to FNS. Themes such as training, guidance, and access emerged as crucial for guiding FNS initiatives and addressing implementation challenges, primarily related to human resources. The chatbot achieved a Global Content Validation Index of 0.88. ConclusionThe developed chatbot represents a significant advancement in supporting FNS practices within primary health care. Practice implicationBy providing an innovative, interactive, educational tool that is both accessible and reliable, this digital assistant has the potential to facilitate the operationalization of FNS practices, addressing the critical need for effective training and counseling in developing countries.

The exploration of users’ perceived value from personalization and virtual conversational agents to enable a smart home assemblage– A mixed method approach

Advancements in smart home technologies are beneficial and valuable for both users and firms. Nevertheless, users’ acceptance of these applications are considerably influenced by the need for user personalization, complex and dynamic within a smart home environment, and thus, can present limitations. Additionally, the role of a virtual conversational agent, both as an individual entity and as part of a collective network, in the setup and management of various smart home devices warrants further exploration. To investigate the enablers and inhibitors of smart home adoption and the role of virtual conversational agents, a mixed-method design was employed, comprising two studies: a survey of 1403 users (Study 1) and a topic modeling approach based on 355,340 Reddit posts (Study 2). The findings indicate that the usefulness of personalized applications and services, as well as the integration of a virtual conversational agent as a central component, are crucial in facilitating smart home adoption. The empirical analyses provide significant implications for research and practice by extending existing technology acceptance and use models to include smart device assemblages.

Co-Designing a Conversational Agent With Older Adults With Chronic Obstructive Pulmonary Disease Who Age in Place: Qualitative Study.

As a reaction to the global demographic increase in older adults (aged 60+ years), policy makers call for initiatives to enable healthy aging. This includes a focus on person-centered care and access to long-term care for older adults, such as developing different services and digital health technologies. This can enable patients to engage in their health and reduce the burden on the health care systems and health care professionals. The European Union project Smart Inclusive Living Environments (SMILE) focuses on well-being and aging in place using new digital health technologies. The novelty of the SMILE project is the use of a cocreational approach focused on the needs and preferences of older adults with chronic obstructive pulmonary disease (COPD) in technology development, to enhance access, adaptation, and usability and to reduce stigma. The study aimed to describe the perspective, needs, and preferences of older adults living with COPD in the context of the design and development of a conversational agent. This study carried out a data-driven thematic analysis of interview data from 11 cocreation workshops with 33 older adults living with COPD. The three particular features that the workshop participants wanted to implement in a new technology were (1) a "my health" function, to use technology to manage and learn more about their condition; (2) a "daily activities" function, including an overview and information about social and physical activities in their local area; and (3) a "sleep" function, to manage circadian rhythm and enhance sleep quality, for example, through online video guides. In total, 2 overarching themes were identified for the 3 functions: measurements, which were actively discussed and received mixed interest among the participants, and health literacy, due to an overall interest in learning more about their condition in relation to everyday life. The future design of digital health technology must embrace the complexities of the everyday life of an older adult living with COPD and cater to their needs and preferences. Measurements should be optional and personalized, and digital solutions should be used as a supplement to health care professionals, not as substitute.

OPERA: Harmonizing Task-Oriented Dialogs and Information Seeking Experience

Existing studies in conversational AI mostly treat task-oriented dialog (TOD) and question answering (QA) as separate tasks. Towards the goal of constructing a conversational agent that can complete user tasks and support information seeking, it is important to develop a system that can handle both TOD and QA with access to various external knowledge sources. In this work, we propose a new task, Open-Book TOD (OB-TOD), which combines TOD with QA and expands the external knowledge sources to include both explicit sources (e.g., the web) and implicit sources (e.g., pre-trained language models). We create a new dataset OB-MultiWOZ, where we enrich TOD sessions with QA-like information-seeking experience grounded on external knowledge. We propose a unified model OPERA ( Op en-book E nd-to-end Task-o r iented Di a log) which can appropriately access explicit and implicit external knowledge to tackle the OB-TOD task. Experimental results show that OPERA outperforms closed-book baselines, highlighting the value of both types of knowledge. 1