Hello, everybody. My name is Shraddha Kohli, and today I will be talking about building user trust in conversational AI, the role of explainable AI in chatbot transparency. The agenda for today's talk is, I will be introducing what is explainable AI, the evolution of chatbots, what are the key technologies in the explainable AI domain, what is the methodology that I used in my research, What was the result of my research, the user impact, some of the challenges, and lastly, what is the future research directions that can be taken in this area? Starting with introduction to explainable AI. Explainable AI tackles the black box challenge in complex AI systems by making decision processes understandable. It helps, it's essentially for systems that interact directly with users, such as chatbots. Without transparency, it's impossible. Users often feel unsure about how chatbots arrive at responses, especially in areas such as healthcare or finance. This leads to a lot of mistrust and hesitation for users to use these technologies. There are many explainable AI methods like LIME, SHAP, and Counterfactuals that can actually help chatbots provide insights into their reasoning, fostering user trust, and enabling developers to fine tune their tools. How do chatbots actually evolve? Chatbots have been here for a long time, since the 1960s, and previously they were rule based on pattern matching and scripted responses. But now with the advancement in the field of AI, a lot of them use very deep neural networks as well as very advanced AI and machine learning technology to allow for more natural and conversational interactions. But today's modern chatbots Though they are very powerful, they do not provide us the transparency in their decision making, which tends to create a sense of doubt in the user's mind. In order to solve this, there are a few key techniques that we can use for chatbots. The first one is called LIME, which is Local Interpretable Model Agnostic Explanations. This tool necessarily generates explanations by creating a simplified interpretable model of chatbot behavior for a single response. Users can see which input features example words most influenced the chatbot's responses. The second one stands for SHAMP, which is Shapely Additive Explanations. This uses values from game theory to assign importance to input features, such as adding weights to the different words. Which helps to offer a clearer picture of how each input or how each word has been contributing to the chatbot's response at both the local as well as global levels. Lastly, the counterfactual explanations provide how slight changes in the user's input can actually lead to different outputs, helping users understand the decision boundary and the sensitivity of the chatbot to different input variations. For our study, we actually interviewed around 150 users in different demographics, and we selected three different types of chatbot models, ranging from retrieval based, generative, as well as hybrid models. These represent the diversity of different modern chatbots that we have, ranging from simple FAQ based to highly conversational systems that we see nowadays. The explainable AI application process, we basically use the three models that are described above to, to use, on our users. Starting with the LIME one, this one helped us analyze specific chatbot responses by approximating them with simple interpretable models. The SHAP1 helped us to calculate feature importance scores to understand which parts of the input influence responses. And lastly, the counterfactuals identified minor changes in input that would actually yield a different output, showcasing the sensitivity of the response. In order to evaluate this, we had three different metrics. The first one was faithfulness. What was the accuracy of the response? The second one was stability. What is the consistency of the explanations across the different users? And lastly, the comprehensibility. How easy was it for non technical users as well to understand the responses? The results were pretty good. What we found was that LIME worked well for individual, for explaining individual responses, but it struggled at global level, meaning that the explanations may vary for similar inputs. The SHAP provided detailed insights into feature importance across multiple interactions. However, due to high processing and computational need, LIME did not meet the initial need This is not, this was not feasible and also confusing for non technical users. The counterfactual explanations were very intuitive for users, but this was also challenging due to real time demands that the users have. The main insight is that each technique brings its own pros and cons to the chatbot transparency, though real time interaction and latency and ease of comprehension are still some areas that need to be explored. Some of the key data points and the user impact was after trying this with, trying these techniques with some of the users, we found that their trust in their users trust in the chatbot technology improved by 35%. They also understood that 48% What of users understood what is the, what is the reasoning behind the chatbot behavior complex task engagement. This was the willingness to engage with complex chatbot tasks was also increased by 27%. The error tolerance for users. Basically, in forgiving the chatbot for the responses or the errors was also increased to a stunning 40%. And the overall engagement and the removal of hesitation increased by 30, 31%. So pretty promising results. Some of the challenges that we experienced in implementing the explainable AI for chatbots is number one, as I said, real time constraints. There is a, the generate, generating explanations in real time without actually providing a noticeable delay and latency is still difficult. How do we balance detail with comprehension to make sure that the explanations have the correct amount of detail, but they are also easy to be understood by non technical users. Then each of the model, as I said, has its own limitation and not everything will be suitable for complex architectures in state of the art chatbots that we have nowadays. Lastly, ethical concerns. When our chatbots are providing us with this information, there is also a chance that some sensitive information could be exposed or there could be some bias. This requires careful ethical handling of explanation techniques. There are a lot of future research directions that are, that we can take in this area, starting with advancing the explainable AI for complex models. How do we develop new explainable AI methods suited to complex architectures like transformers, which are common in chatbots? Lastly, real time adaptive explanations. How do we make sure that the research adaptive systems that adjust with the right level of detail that is needed? And that's Third, self explaining chatbots. Investigate how chatbots can learn to explain themselves as they evolve, leading to a truly self explanatory AI that users understand intuitively. Lastly, I would like to conclude, as chatbots became an integral part of customer service and user interaction across sectors, building trust in these sectors is critical. Explainable AI offers a pathway to address the black box nature of AI by making chatbot decision making processes transparent and interpretable. By leveraging methods like LIME, SHAP, and counterfactual explanations, we enable chatbots to provide insights into their responses, which helps to foster a more reliable and user centered experience. This transparency not only enhances user trust and confidence in the chatbot, but also equips developers with powerful tools for refining and improving these systems. Our research demonstrated that integrating XAI with chatbot technology yields significant benefits in user trust, comprehension, and engagement. However, this journey is still a long one and there is a lot of research that still needs to be done in this area. The advancements in XAI techniques have provided a promising foundation for more accountable, transparent, and trustworthy AI applications. Thank you.