Just-in-Time News: An AI Chatbot for the Modern Information Age

Abstract

This study advances AI-powered news delivery by introducing an innovative chatbot capable of providing personalized news summaries and real-time event analysis. This approach addressed a critical gap identified through a comprehensive review of 52 AI chatbot studies. Unlike prior models limited to static information retrieval or predefined interactions, this chatbot harnesses generative AI and real-time data integration to deliver a dynamic and tailored news experience. Its unique architecture combines conversational AI, robotic process automation (RPA), a comprehensive news database (989,432 reports from 2342 sources spanning 27 October 2023 to 30 September 2024), and a large language model (LLM). Within this architecture, LLM generates dynamic queries against the News database for obtain tailored News for the users. Hence, this approach interprets user intent, and delivers LLM-based summaries of the fetched tailored news. Empirical testing with 35 users across 321 diverse news queries validated its robustness in navigating a combinatorial classification space of 53,916,650 potential news categorizations, achieving an F1-score of 0.97, recall of 0.99, and precision of 0.96. Deployed on Microsoft Teams and as a standalone web app, this research lays the foundation for transformative AI applications in news analysis, promising to revolutionize news consumption and empower a more informed citizenry.

1. Introduction

The rapid evolution of artificial intelligence (AI) is driving transformative changes across various sectors, and chatbots have emerged as a particularly promising application of this technology. These AI-powered conversational agents are designed to engage in human-like conversation and provide support, information, and assistance across diverse fields. Previous research has showcased the potential of AI chatbots in healthcare [1,2,3,4], education [5,6,7,8], and business [9,10,11,12], highlighting their versatility in addressing sector-specific challenges. However, the full potential of AI chatbots, particularly in the realm of real-time news and global event analysis, remains largely untapped.

The existing method of dashboard or app-based News delivery often provides a static and one-dimensional view of information as shown in [13,14,15,16,17,18,19,20]. The need for research in AI-powered news delivery is underscored by the limitations of existing approaches and the transformative potential of this technology. Traditional news consumption methods, such as news aggregators and social media platforms, often overwhelm users with irrelevant or unreliable content, hindering their ability to synthesize key insights and understand the broader context of events [19]. Moreover, the constant influx of news can lead to information overload and a distorted perception of reality, as highlighted by the prevalence of negative news and its impact on public sentiment [15,19]. AI chatbots, with their ability to filter, analyze, and summarize information, offer a solution to these challenges, providing users with a personalized and efficient way to stay informed about the news that matters most to them.

This research delves into the development and deployment of a new AI-powered chatbot specifically designed for real-time news and global event analysis. Unlike existing chatbots that primarily focus on static information retrieval or pre-defined conversational flows (e.g., [21,22,23,24]), the AI-chatbot reported in this paper leverages generative AI capabilities to provide users with dynamic and just-in-time news summaries. By connecting to live news databases and employing advanced Large Language Models (LLMs) as previously portrayed in [25], this new AI-chatbot offers a unique solution for staying informed in today’s dynamic information environment. It should be noted that none of the existing AI-driven News analytics systems (e.g., [13,15,19,20,25]) demonstrate the integration of an AI-Chatbot facility that understands users’ intent and provides a custom summary of curated news reporting. This approach addresses a significant gap in the literature, as none of the 52 reviewed papers discusses a standalone AI chatbot solution that connects to live news databases to provide up-to-the-minute information and analysis on current events.

The development of this real-time news chatbot is driven by the recognition that traditional news consumption methods often fall short of meeting the needs of today’s information seekers. Generic AI-based News aggregators and social media platforms, while providing access to a vast amount of information, often overwhelm users with irrelevant or unreliable content [25]. Moreover, the constant influx of news can lead to information overload and hinder the ability to synthesize key insights and understand the broader context of events [15]. This chatbot aims to address these challenges by providing users with a personalized and efficient way to stay informed about the news that matters most to them. By combining the strengths of generative AI, topic-based interaction models, real-time News data integration, and integration of feedback mechanism, this chatbot offers a valuable tool for news consumption and analysis. The chatbot’s ability to understand user intent, generate concise summaries, and provide access to diverse news sources has the potential to transform how individuals engage with and comprehend current events. By examining the shortcomings of 52 existing AI chatbot papers, this research identified critical requirements for the development of more advanced and effective chatbots. Then, with diagrams and schematics, this paper presents the technical development, implementation, and deployment of this novel chatbot, paving the way for a more informed and engaged citizenry in the digital age.

Unlike the 52 AI-chatbot-related papers critically reviewed within the scope of this study, which primarily focused on static information retrieval or pre-defined conversational flows, the proposed chatbot leverages generative AI and real-time data integration to offer a dynamic and personalized news experience. The following are the theoretical contributions of this study:

• New Architecture: Introduces a new architecture combining conversational AI with generative AI capability, RPA, a news database, and an LLM for dynamic news understanding and summarization.
• Complexity Analysis: Provides a detailed mathematical complexity analysis, offering insights into system performance and optimization.
• Robustness of Solution: Demonstrates the chatbot’s robustness in handling a vast combinatorial space of potential news classifications (53,916,650 unique combinations).

Apart from the above theatrical contributions, this paper provides the following practical contributions:

• Real-world Implementation: Implemented and evaluated with a massive news database (989,432 reports from 2342 sources), distinguishing it from previous studies with smaller datasets.
• Multi-Platform Deployment: Deployed on Microsoft Teams and as a standalone web application, showcasing practical usability in various contexts.
• High Performance: Achieved an F1-score of 0.97, recall of 0.99, and precision of 0.96, demonstrating superior performance in delivering accurate news summaries.

2. Background & Literature Review

AI-powered chatbots have emerged as transformative tools across various domains, addressing critical needs and challenges. In healthcare, chatbots offer mental health support, as seen in [1], and assist in medical diagnosis [2] and information retrieval [3,4]. Within education, chatbots facilitate personalized learning [5,6], enhance student engagement [7], and provide valuable feedback [8]. In business, chatbots streamline customer service [9,10] aid in marketing [11], and automate business processes [12]. These diverse applications highlight the versatility and problem-solving capabilities of AI chatbots in addressing sector-specific challenges.

Beyond their problem-solving capabilities, AI chatbots are also being evaluated based on their features. For instance, in the context of sexual and reproductive health, [26] explored the role of AI attributes, such as perceived intelligence and anthropomorphism, in the adoption of a chatbot designed to provide information on this sensitive topic. In the realm of education, Ref. [27] examined the impact of AI chatbots on learner emotions, highlighting the potential for these technologies to reduce anxiety and increase motivation. Additionally, Ref. [28] investigated the effectiveness of AI chatbots in second language vocabulary acquisition, demonstrating their potential to enhance both receptive and productive vocabulary knowledge. These studies underscore the importance of considering various features of AI chatbots, beyond their mere functional capabilities, in understanding their potential impact and adoption across different domains.

The accuracy of AI chatbots is a critical aspect of their evaluation, and recent research has shed light on their capabilities and limitations in this regard. For example, Xiao et al. [27] assessed the influence of AI chatbots on language learners’ emotions and highlighted the need for validated emotion scales in chatbot-assisted language learning contexts. Research in [28] employed quantitative tests to assess receptive and productive vocabulary knowledge, demonstrating that AI Chatbots based on Large Language Models significantly aid students in acquiring both receptive and productive vocabulary knowledge. Wang et al. in [10] conducted a survey with 294 U.S. marketing employees to examine the impacts of chatbot-enabled agility on customer service performance. Liu et al. in [29] demonstrated the superiority of a chatbot-delivered self-help depression intervention over bibliotherapy in terms of reducing depression and anxiety symptoms, achieving an average GAD-7 score of 2.94. In the context of organic chemistry, the study [30] assessed the accuracy of ChatGPT and Bard in understanding structural notations and answering related questions, revealing limitations in their ability to handle complex tasks and highlighting the need for further training and development. Additionally, the study [31] compared the accuracy of retrieval-based and generative-based chatbots in the mental healthcare domain, with retrieval-based models achieving an accuracy of 65.5% and generative-based models achieving 71.2%. These studies underscore the ongoing efforts to evaluate and enhance the accuracy of AI chatbots across various domains, paving the way for their reliable and effective integration into diverse fields.

Table 1. The multi-dimensional use of AI-Chatbots modern age.

Main Area	Sub-Area	Paper
Business	Customer Relationship Management	[21]
Business	Customer Service	[9,10]
Business	Marketing	[11]
Business	Business Process Automation	[12]
Business	Restaurant Management	[22]
Business	Online Shopping Assistance	[32]
Business	Knowledge Industries	[33]
Chemistry	Organic Chemistry	[30]
Computer Science	Code Generation	[34]
Computer Science	Natural Language Processing	[35,36,37,38]
Computer Science	Software Development	[39]
Computer Science	Chatbot Error Correction	[40]
Computer Science	Chatbot Evaluation	[41]
Computer Science	Human-Computer Interaction	[42]
Education	Assessment and Evaluation	[43]
Education	Educational Chatbots	[23]
Education	Educational Data Mining	[44]
Education	Educational Technology	[45]
Education	Language Learning	[6,27,28,46,47,48]
Education	Learning and Teaching	[49]
Education	Learning Management System	[50]
Education	Personalized Learning	[5]
Education	Student Engagement	[7]
Education	Student Feedback & Support	[8,51]
Education	Technology Adoption	[52,53,54]
Education	Surgical Education	[53]
Healthcare	Health Information	[24]
Healthcare	Immunization Information	[55]
Healthcare	Medical Diagnosis	[2]
Healthcare	Medical Information Retrieval	[3,4]
Healthcare	Sexual and Reproductive Health	[26]
Healthcare	Pathology	[56]
Healthcare	Radiotherapy	[57]
Linguistics	Language Learning	[6,27,28,46,47,48]
Mental Health	Depression Interventions	[29]
Mental Health	Emotion Recognition	[1]
Mental Health	Empathy and Well-being	[58]
Mental Health	Mental Health Care	[31]
Mental Health	Psychotherapy	[59]
Psychology	Mental Health	[29,31,58,59]
Marketing	Ethical Use of Chatbots	[60]

provides a comprehensive summary of AI-Chatbots currently solving problems in different facets of life.

While the reviewed papers showcase the wide-ranging applications of AI chatbots, there is a notable absence of research on deployed AI chatbots specifically designed for real-time news and global event analysis. None of the papers discusses standalone chatbot solutions that connect to live news databases to provide up-to-the-minute information and analysis on current events. This gap in the literature highlights an area ripe for exploration, as the potential for AI chatbots to revolutionize news consumption and analysis remains largely untapped (as shown in

Table 2. The gap of existing AI-chatbot-based technologies in providing real-time News and Global Event Intelligence.

Paper	News/Event Analysis	Real-Time Updates	Technical Insights	Connection with Live Database	Deployed Standalone Custom Chatbot	Detailed Performance Evaluation
[1]	No	No	Yes	No	No	Yes
[50]	No	No	Yes	No	Yes	Yes
[8]	No	No	No	No	No	Yes
[35]	No	No	No	Yes	Yes	No
[2]	No	No	Yes	Yes	Yes	Yes
[23]	No	No	No	No	No	Yes
[36]	No	No	Yes	No	No	No
[3]	No	No	Yes	No	Yes	Yes
[9]	No	No	Yes	No	Yes	No
[49]	No	No	No	No	Yes	Yes
[34]	No	No	Yes	No	No	Yes
[51]	No	No	Yes	No	No	Yes
[44]	No	No	No	No	No	Yes
[5]	No	No	No	No	No	Yes
[46]	No	No	No	No	Yes	Yes
[11]	No	No	Yes	No	Yes	Yes
[21]	No	No	Yes	No	Yes	Yes
[43]	No	No	No	No	No	Yes
[45]	No	No	No	No	No	Yes
[7]	No	No	No	No	No	Yes
[22]	No	No	Yes	No	Yes	Yes
[37]	No	No	Yes	No	No	Yes
[30]	No	No	Yes	No	No	Yes
[59]	No	No	No	No	No	Yes
[58]	No	No	Yes	No	Yes	Yes
[47]	No	No	Yes	No	No	Yes
[24]	No	No	Yes	No	No	No
[6]	No	No	No	No	Yes	Yes
[39]	No	No	Yes	Yes	Yes	Yes
[61]	No	No	No	No	Yes	Yes
[26]	No	No	Yes	No	Yes	Yes
[29]	No	No	Yes	No	Yes	Yes
[4]	No	No	Yes	No	No	Yes
[31]	No	No	Yes	No	Yes	Yes
[48]	No	No	Yes	No	No	Yes
[33]	No	No	Yes	No	No	Yes
[32]	No	No	No	No	No	Yes
[52]	No	No	Yes	No	Yes	Yes
[53]	No	No	Yes	No	Yes	Yes
[55]	No	No	Yes	No	Yes	Yes
[54]	No	No	No	No	No	Yes
[12]	No	No	Yes	No	No	No
[10]	No	No	Yes	No	No	Yes
[27]	No	No	Yes	No	No	Yes
[28]	No	No	Yes	No	No	Yes
[38]	No	No	Yes	No	No	Yes
[41]	No	No	Yes	No	No	Yes
[56]	No	No	Yes	No	No	No
[40]	No	No	Yes	No	No	No
[42]	No	No	Yes	No	No	Yes
[60]	No	No	No	No	No	No
[57]	No	No	Yes	No	Yes	No

).

The absence of research on news and event analysis chatbots may stem from the complexities associated with developing such a solution. Connecting to live news databases, ensuring real-time updates, and providing accurate and unbiased analysis present significant technical hurdles. Moreover, the dynamic and rapidly evolving nature of news and events requires a high degree of adaptability and learning capabilities in the chatbot. These challenges may explain the lack of focus on this specific application in the current literature.

Despite the complexities, the potential benefits of news and event analysis chatbots are substantial. Such chatbots could provide users with personalized news feeds, real-time updates on global events, and diverse perspectives on critical issues. They could also assist in identifying misinformation, combating fake news, and fostering media literacy. Future research should address this gap by exploring the technical development, user adoption, and societal impact of news and event analysis chatbots, paving the way for a more informed and engaged citizenry.

3. Materials and Methods

This research investigates the development and deployment of a novel AI-powered chatbot designed for real-time news and global event analysis. The chatbot leverages the capabilities of Microsoft Copilot Studio, Google Gemini Application Programming Interface (API), Microsoft Power Automate and Microsoft Dataverse to provide users with concise and up-to-date summaries of news from around the world [62]. The development of the AI chatbot involved a combination of generative AI and topic-based interaction models within Microsoft Copilot Studio. This hybrid approach enables the chatbot to understand user intent regarding news and global events effectively. The chatbot interacts with users through natural language, allowing them to ask questions or request information on specific topics or events.

3.1. Mathematical Model

The core of the chatbot’s functionality relies on the Google Gemini API, which employs advanced LLMs for natural language understanding and generation. These LLMs are based on deep learning algorithms that have been trained on massive datasets of text and code, enabling them to perform tasks such as text summarization, question answering, and language translation with high accuracy.

3.1.1. Transformer Architecture

Gemini likely utilizes the Transformer architecture, a powerful neural network design that has become central to many advanced language models. Transformers excel at processing sequential data like text by employing self-attention mechanisms. Self-attention allows the model to weigh the importance of different words in the input when generating a response, capturing complex relationships and dependencies within the text. This is achieved by calculating attention scores, which are essentially measures of similarity between different parts of the input sequence. These scores are derived from the dot product of query, key, and value matrices, which are mathematical representations of the input text. In addition to self-attention, Transformers also incorporate feedforward neural networks. These networks apply non-linear transformations to the information gleaned from the self-attention layer, enabling the model to learn intricate patterns and representations in the data.

3.1.2. Statistical Language Modeling

At its core, Gemini operates as a statistical language model. This means it learns the probabilities of different words occurring based on the preceding context. In simpler terms, it predicts the next word in a sequence by considering the words that came before it. This probabilistic approach is often represented mathematically as $P\left(w_t\right|w_{t-1},w_{t-2},\dots ,w_1)$, where w_t is the current word and w_t−1, etc., are the previous words. By learning these probabilities from massive amounts of text data, Gemini can generate coherent and contextually relevant responses. While not the primary architecture, concepts from Hidden Markov Models (HMMs) may also play a role in how Gemini transitions between different topics or “states” within a conversation.

3.1.3. Hypothetical Equations

While the exact equations behind Gemini’s inner workings are not publicly available, researchers can make educated guesses based on common practices in language modeling. A simplified representation of the self-attention mechanism could be expressed as Equation (1).

$$Attention\left(Q,K,V\right)=softmax\left({\displaystyle \frac{Q{K}^T}{\sqrt{d_k}}}\right)V$$

(1)

In Equation (1), Q, K, and V are matrices representing queries, keys, and values derived from the input text, d_k is the dimension of the key vectors, and softmax is a function that normalizes the attention scores into probabilities. Let’s consider simplified values for the matrices: Q = [[1, 2], [3, 4]], K = [[5, 6], [7, 8]], V = [[9, 10], [11, 12]], and d_k = 2.

• Calculate QK^T = [[19, 22], [43, 50]].
• Divide by d_k: [[9.5, 11], [21.5, 25]].
• Apply softmax: [[0.1192, 0.2689], [0.8808, 0.7311]].
• Multiply by V: [[3.4417, 3.7568], [10.5583, 11.2432]].

This example demonstrates how the attention mechanism calculates weighted values based on the similarity between different parts of the input sequence.

Similarly, a simplified equation for token probability might look like Equation (2).

$$P\left(w_t|w_{t-1},w_{t-2},\dots ,w_1\right)=softmax(W_h,h_t)$$

(2)

where, W_h is a weight matrix connecting the hidden state to the output vocabulary and h_t is the hidden state of the model at time t (influenced by previous words). It should be mentioned that $P\in [0, 1]$ is the probability of w_t given that the previous states are w_t−1, w_t−2, …, w₁. Let’s assume a simplified scenario: W_h = [[0.5, 0.2], [0.3, 0.4]], h_t = [0.8, 0.6], and the output vocabulary is {“the”, “cat”}.

• Calculate W_h h_t = [0.52, 0.48].
• Apply softmax: [0.5498, 0.4502].

This indicates that given the hidden state h_t, the probability of the next word being “the” is approximately 0.55, and the probability of it being “cat” is approximately 0.45.

3.1.4. AI Studio Playground Parameters

The AI Studio Playground provides users with various parameters to fine-tune Gemini’s behavior. These parameters likely influence the model’s output through mathematical operations. For example, “harmful content” could be controlled by a Toxicity (text) function that assigns a score between 0 and 1 to the generated text, with higher scores indicating a greater likelihood of harm. The model might then use a threshold to filter out responses exceeding a certain toxicity level. Safety parameters like “temperature” (T) can be represented mathematically as Equation (3).

$$\acute{P}(w_t|context)={\displaystyle \frac{{P\left(w_t\right|context)}^{\raisebox{1ex}{$1$}\!\left/ \!\raisebox{-1ex}{$T$}\right.}}{Z\left(T\right)}}$$

(3)

where, Z(T) is a normalization factor. Higher T leads to more diverse (and potentially less predictable) outputs. Higher temperatures lead to more diverse and potentially less predictable outputs by flattening the probability distribution, while lower temperatures make the model more focused and deterministic. For example: Let $P\left(w_t\right|context)$ = [0.7, 0.2, 0.1] for three possible words.

• For T = 0.5 (low temperature): [0.9643, 0.0354, 0.0003]. The model becomes highly focused on the most likely word.
• For T = 1.0 (normal temperature): [0.7, 0.2, 0.1]. The probabilities remain unchanged.
• For T = 2.0 (high temperature): [0.5946, 0.2649, 0.1405]. The model’s output becomes more diverse, considering less likely words.

Let’s consider the following three potential summaries generated by the LLM for a news article about a political event:

• Summary 1: “The President addressed the nation, outlining a new economic plan”.
• Summary 2: “Amidst growing concerns, the President announced a controversial economic strategy”.
• Summary 3: “The President’s speech sparked outrage, with critics denouncing the proposed economic measures”.
• These summaries differ in their tone and emphasis. Summary 1 is neutral, Summary 2 highlights controversy, and Summary 3 focuses on negative reactions.

The temperature parameter influences the LLM’s choice of summary:

• Low Temperature (e.g., T = 0.5): The LLM will likely select the most probable and neutral summary, which is often the safest option. In this case, it would likely choose Summary 1.
• Normal Temperature (e.g., T = 1.0): The LLM has more flexibility in its selection, potentially choosing a slightly more nuanced summary like Summary 2.
• High Temperature (e.g., T = 2.0): The LLM is encouraged to explore less probable and potentially more dramatic summaries, possibly selecting Summary 3, which highlights the controversy and strong reactions.

In the context of news summarization, the temperature parameter allows for control over the tone and focus of the generated summaries. A higher temperature can lead to more diverse and potentially more engaging summaries but also increases the risk of selecting less probable or inaccurate summaries.

Other parameters like “Top-k Sampling” and “Repetition Penalty” also have mathematical representations, influencing the selection and probability of words in the generated text. For Top-k Sampling the k most likely words are considered at each step. This can be represented as selecting words from the set Top_k(P(w_t|context)). On the other hand, for Repetition Penalty, a function Penalty (w_t, history) decreases the probability of a word based on its frequency in the recent output history.

3.2. Architecture

The architecture presented for this AI-powered news chatbot represents a sophisticated integration of diverse technologies to achieve a seamless and informative user experience, as shown in Figure 1. It leverages a synergistic interplay between a knowledge-grounded conversational AI, a robust robotic process automation (RPA) engine, a comprehensive news database, and a cutting-edge large language model (LLM).

At the heart of this architecture lies the AI chatbot, built on a Microsoft Copilot foundation. This chatbot is not simply a conversational interface; it is imbued with a contextual understanding derived from a vast knowledge base of websites and documents. This pre-existing knowledge allows the chatbot to comprehend the nuances of news reports and global events, going beyond simple keyword matching to truly grasp the user’s intent and information needs. Unlike traditional rule-based Chatbots, the chatbot presented in this paper possesses contextual understanding of current News along with Generative AI functions as shown in Figure 1.

When a user interacts with the chatbot, the conversation triggers a dynamic process of news retrieval and summarization. The chatbot’s understanding of the user’s query is relayed to a Microsoft Power Automate flow. This RPA engine acts as an intelligent intermediary, orchestrating the formation of a precise FetchXML query. This query construction is not a static process; it leverages the power of a Google Gemini LLM to dynamically generate a query that accurately reflects the user’s request (as shown in step 3 of Figure 1).

This dynamically generated FetchXML query is then executed against a comprehensive news database housed in Microsoft Dataverse. This database, constantly updated with a massive influx of news data from a multitude of sources, provides the raw material for the chatbot’s responses. The RPA engine retrieves the relevant news articles, filtering through the vast repository to pinpoint the information that precisely matches the user’s needs.

The final stage of this process involves another interaction with the Google Gemini LLM. This time, the LLM is tasked with generating a concise and informative summary of the retrieved news articles. This summarization goes beyond simple concatenation; the LLM leverages its advanced language understanding capabilities to synthesize the key information, identify important trends, and present a coherent overview of the news relevant to the user’s query (as shown in step 4 of Figure 1). This summarized information is then relayed back to the chatbot, which delivers it to the user in a clear and easily digestible format.

As shown in Figure 1, this architecture represents a significant advancement in AI-powered news delivery. By combining the strengths of conversational AI, RPA, a comprehensive news database, and a powerful LLM, it offers a personalized, efficient, and insightful way for users to stay informed in today’s dynamic information environment. This approach not only addresses the limitations of traditional news consumption methods but also opens new possibilities for user engagement and knowledge discovery.

3.3. News Aggregation Process

As seen in Figure 1, the presented AI-Chatbot obtains News reports from the Dataverse News Database. The Architectural details of this News aggregator have been detailed in earlier studies [13,20,25]. However, this section will briefly cover some important aspects of this News aggregator. The Dataverse News Database (i.e., News aggregator) is systematically updated with daily news reports sourced from an extensive array of news outlets. This process leverages a network of news APIs (e.g., The Guardian News API, Microsoft Bing News API), Really Simple Syndication (RSS) feeds, and web scraping technologies, all coordinated through Microsoft Power Automate to refresh the database at variable intervals. Through Power Automate’s RSS connectors, prominent news portals, including CNN, BBC, CNBC, The Washington Post, The New York Times, LA Times, The Epoch Times, Fox News, The Daily Mail UK, The Guardian UK, CBS News, Voice of America, Politico, and Defence One (among others), are updated in real-time as news reports are published. These portals often provide multiple RSS feeds for different categories, such as Breaking News, World News, Political News, and Top Headlines. The Dataverse News Database prioritizes subscriptions to feeds that cover Breaking News, World News, Political News, Military News, and Technology News, with a limited inclusion of Sports, Travel, and Entertainment news, primarily from portals based in the US, UK, Australia, and India.

In addition to live RSS feeds from hundreds of sources, API integration and web scraping technologies enable dynamic acquisition of news from sites that lack RSS support, such as Al Jazeera (Qatar) and Khaleej Times (UAE), with updates occurring every six hours. This news aggregation process for the Dataverse News Database is detailed in prior studies, providing a comprehensive framework for real-time and periodic data integration [13,20].

Microsoft Power Automate functions as the primary orchestrator for automated news aggregation, enabling real-time updates from a vast array of online news portals via RSS, as well as six-hourly updates through API integrations and web scraping. As detailed in [25], an initial trial captured news across three categories—Science & Technology, Education & Learning, and Health & Medicine—from 394 distinct sources. Expanding upon this framework, the current study incorporates fifteen categories (to be discussed in the following section), increasing coverage to approximately 2342 sources. This extensive source network supports balanced reporting and mitigates potential biases that could arise from a limited selection of news sources.

A news event reported in one media source may be biased, but having more news media reporting the same event could provide a more bias-free view if the information is summarized by LLM. For example, consider a news story about immigration. One right-leaning news source might report the story with a headline like “Illegal Aliens Flood Border”. Another left-leaning source might report the same story with a headline like “Undocumented Immigrants Seek Asylum”. In both cases, the underlying news event is the same, but the way it is reported is biased. The first headline uses inflammatory language to paint immigrants in a negative light. The second headline is more neutral and factual. If multiple sources are reporting on the same event, the LLM can identify and remove these biases to generate a more neutral and objective summary. For example, the LLM might generate a summary like this: “There has been an increase in the number of immigrants crossing the border. Some news sources have used biased language to report on this story”. This summary is more bias-free because it does not use inflammatory language or take sides. It simply reports the facts of the event.

In addition to the example above, here are some other full sentences that illustrate how having more news media reporting the same event can provide a more bias-free way of summarizing the information:

• “The LLM can identify and remove biases in news reporting by comparing and contrasting different sources”.
• “By using multiple sources, the LLM can generate a more comprehensive and objective summary of the news event”.
• “Having more news media reporting the same event can help to mitigate the impact of bias on the summarization process”.

Overall, the use of multiple news sources can help to reduce bias in the summarization process. This is because the LLM can compare and contrast different sources to identify and remove biases. As a result, the LLM can generate a more neutral and objective summary of the news event.

In this architecture, Microsoft Power Automate, integrated with LLMs and GPT technology, facilitates not only the classification of news content and assessment of its significance (e.g., local, regional, national, international, or global) but also the detection of any biases present in news reports. By analyzing syntactic patterns and linguistic structures, GPT can identify and classify potential biases within news articles, categorizing them as right-leaning or left-leaning where applicable. The results of this bias identification and tagging process are dynamically updated in the Dataverse News Database and can be viewed live at https://bias.press/news/ (accessed on 1 December 2024). Although the architectural design of the Dataverse News Aggregator is outside the scope of this study, discussing the database features and parameters is essential for understanding the FetchXML implementation strategy outlined in Section 3.2 (Architecture) and Section 3.4 (Implementation Strategy).

3.4. Implementation Strategy

Implementing this sophisticated AI-powered news chatbot requires a carefully orchestrated process, akin to composing a symphony of interconnected technologies. It begins with the development of a robust knowledge base, a rich tapestry of information woven from a vast collection of web pages and documents. This knowledge base, structured as a knowledge graph, becomes the foundation for the chatbot’s contextual awareness, enabling it to understand the nuances of news and global events.

The next movement in this symphony involves the creation of a news database, a dynamic repository of information constantly refreshed with the latest news from a multitude of sources. This database, housed within Microsoft Dataverse, is meticulously curated to ensure accuracy, relevance, and timeliness.

With the knowledge base and news database in place, the focus shifts to the development of the Power Automate flow, the conductor of this technological orchestra. This RPA engine is responsible for orchestrating the seamless flow of information between the chatbot, the LLM, and the news database. It dynamically generates FetchXML queries, precisely worded requests for information, tailored to the user’s specific needs.

The Google Gemini LLM plays a dual role in this symphony. First, it assists the RPA engine in crafting the FetchXML queries, ensuring they accurately capture the user’s intent. Second, it takes center stage in the final act, summarizing the retrieved news articles with its advanced language processing capabilities. It is like a skilled storyteller, weaving together the key information into a concise and engaging narrative.

Finally, the chatbot is deployed, making its debut in a collaborative workspace like Microsoft Teams or as a standalone web application. This deployment marks the culmination of the implementation process, bringing the symphony of technologies to life and delivering a personalized, efficient, and insightful news experience to users.

3.5. Agent Communication

Following implementation, the agent (AI-Chatbot) communicates with the user and other components in a structured and dynamic manner, as illustrated in Figure 2.

• User Initiation: The user initiates the interaction by requesting specific news analytics.
• Intent Identification: The AI-Chatbot, leveraging its knowledge base and natural language understanding capabilities, identifies and reconfirms the user’s intent.
• Information Gathering: The chatbot engages in a brief dialogue with the user to gather essential information, such as the type of event and location of interest.
• RPA Activation: The chatbot relays the gathered information to the Power Automate RPA, initiating the news retrieval process.
• FetchXML Query Generation: The RPA, utilizing the Google Gemini LLM, dynamically generates a FetchXML query tailored to the user’s request.
• News Retrieval: The FetchXML query is executed against the Dataverse News Database, retrieving relevant news articles.
• News Summarization: The retrieved news articles are processed by the Google Gemini LLM to generate a concise summary.
• Response Delivery: The summarized news content is returned to the chatbot, which presents it to the user in a clear and user-friendly format.

This structured communication flow ensures that the AI-Chatbot effectively understands the user’s needs, retrieves relevant information, and delivers a concise and informative news summary in real-time.

To enhance user interaction and continuously improve chatbot accuracy, a structured feedback mechanism is integrated into the system via Microsoft CoPilot’s “End of Conversation” topic. This mechanism enables users to provide feedback on responses through a star-based Customer Satisfaction (CSAT) rating system and additional comments. The technical process involves adaptive dialog management, where user input is captured using Boolean entities to determine satisfaction levels. If feedback indicates dissatisfaction, users are prompted with options to retry or escalate the issue. All collected feedback is stored and processed in Dataverse, leveraging analytics tools for trend identification and performance improvement.

The code implementation for this feedback mechanism is provided in Appendix A (System Redirect & Feedback), detailing the adaptive dialog framework and conditional workflows. This robust integration ensures that user input directly informs iterative updates to the chatbot, fostering transparency, reliability, and user trust.

3.6. Complexity Analysis

Analyzing the complexity of this system requires breaking down the process into its key components and evaluating their individual contributions to the overall computational cost.

3.6.1. Conversational AI

The complexity of the conversational AI, powered by Microsoft Copilot, is influenced by factors such as the size of the knowledge base, the complexity of the dialogue flows, and the sophistication of the natural language understanding (NLU) model. Assuming the knowledge base is represented as a graph with ‘V’ vertices (entities) and ‘E’ edges (relationships), and the dialogue flows involve ‘D’ distinct states, the complexity can be approximated as $O(V+E+D)$.

3.6.2. RPA Engine

The complexity of the RPA engine, driven by Microsoft Power Automate, is primarily determined by the complexity of the FetchXML query generation process and the number of API calls made to the Google Gemini LLM. Assuming the query generation involves ‘Q’ steps and ‘G’ API calls are made, the complexity can be represented as $O(Q+G)$.

3.6.3. LLM Operations

The complexity of the LLM operations, performed by Google Gemini, is influenced by the size of the model, the length of the input text, and the specific task being performed (query generation or summarization). Assuming the LLM has ‘P’ parameters and the input text has ‘T’ tokens, the complexity can be approximated as $O(P\times T)$.

3.6.4. News Database

The complexity of the news database operations, performed on Microsoft Dataverse, is primarily determined by the efficiency of the FetchXML query execution and the size of the news database. Assuming the database has ‘N’ news articles and the query execution time is ‘R’, the complexity can be represented as $O(N\times R)$.

3.6.5. Overall System Complexity

Combining the complexities of the individual components, the overall system complexity can be approximated as shown in

Table 3. Critical analysis of the Complexity of the proposed system.

Component Name	Calculation of Complexity
Complexity of the conversational AI	$O(V+E+D)$
Complexity of RPA engine	$O(Q+G)$
Complexity of LLM operation	$O(P\times T)$
Complexity of News database	$O(N\times R)$
Overall Complexity of the proposed system	$O(V+E+D+G+\left(P\times T\right)+\left(N\times R\right))$

.

This complexity analysis highlights the key factors influencing the system’s performance. Optimizing each component, such as reducing the size of the knowledge base, streamlining the RPA flows, and employing efficient database queries, can significantly enhance the overall efficiency and responsiveness of the AI-powered news chatbot.

This methodology enables the development of a robust and scalable AI chatbot solution for real-time news and global event analysis, addressing the gap identified in the literature and providing users with a valuable tool for staying informed in today’s dynamic information environment.

4. Results

To rigorously evaluate the efficacy of the proposed AI-Chatbot in a real-world scenario, the AI-Chabot was interfaced with an extensive news database comprising 945,029 events spanning 152,837 distinct locations, collected between 30 September 2023, and 1 October 2024. This dataset, several orders of magnitude larger than those employed in previous AI-Chatbot studies [1,2,3,4,5], provides a robust testing ground for assessing the chatbot’s capabilities in handling and analyzing massive news data.

The news reports, sourced in real-time from a diverse range of prominent news outlets (e.g., BBC, CNN, New York Times, etc.), were meticulously categorized into 15 primary event categories, each further subdivided into 202 subcategories. These News were classified into multiple categories like “Politics, Governance, and International Affairs”, “Industry and Business News”, “Crime, Safety, and Security”, “Disaster, Accidents, and Crisis” etc. The full lists of these 15 event categories is provided in

Table 4. The main categories of News along with a number of sub-categories and news reports within categories and sub-categories.

Event Category	Count of Sub-Categories	Count of Title	Count of Source URL
Politics, Governance, and International Affairs	19	190,702	163,466
Industry and Business News	27	137,150	131,439
Economic and Financial News	19	108,615	96,597
Crime, Safety, and Security	25	95,859	85,790
Entertainment and Culture	15	94,974	87,577
Uncategorized	NA	88,499	81,278
Human Rights and Social Issues	16	62,090	55,921
Disasters, Accidents, and Crisis	20	32,851	27,783
Science and Technology	19	29,372	28,008
Environment and Climate	7	26,564	24,772
Legal and Justice	4	25,588	23,560
Health and Medicine	12	19,326	18,266
Lifestyle and Trends	7	15,801	15,055
Education and Learning	2	10,961	10,391
Media and Communication	8	5782	5405
Unusual and Extraordinary Events	2	848	793

and

Table 5. News categories grouped by News impact (i.e., local, community level, state level, regional, international, global impact).

Event Category	Community	Global	International	Local	Nation	Region	State	Total
Unusual and Extraordinary Events	70	107	180	217	226	14	34	848
Science and Technology	1916	12,143	8256	3089	3474	176	261	29,315
Politics, Governance, and International Affairs	704	671	74,214	15,104	79,084	1171	19,677	190,625
Media and Communication	368	1076	798	1080	2293	27	133	5775
Lifestyle and Trends	4185	703	2543	6392	1496	143	325	15,787
Legal and Justice	798	634	3285	8070	10,996	8	1753	25,544
Industry and Business News	4046	28,030	23,878	36,149	39,019	1576	3839	136,537
Human Rights and Social Issues	3956	266	9554	25,347	17,308	206	5441	62,078
Health and Medicine	3057	1826	2494	5476	5569	101	786	19,309
Environment and Climate	1240	2732	2996	13,387	3716	1022	1444	26,537
Entertainment and Culture	11,411	2667	24,269	35,928	19,218	82	1123	94,698
Education and Learning	586	158	516	5038	3359	36	1262	10,955
Economic and Financial News	1682	26,844	11,320	18,888	46,571	1241	1450	107,996
Disasters, Accidents, and Crisis	637	204	1217	23,114	3301	1697	2677	32,847
Crime, Safety, and Security	2337	2953	17,156	43,563	22,806	662	6367	95,844
Total	44,300	85,885	201,618	277,548	275,089	8720	49,963	943,123

. As shown in Table 5, each of these News categories are further divided into 202 Sub-categories of News. For example, “Hurricane Event”, “Tornado Event”, “Significant Earthquake event”, “Major Flooding Event”, “Nuclear Accidents” etc. are part of the 20 sub-categories of News under “Disaster, Accidents, and Crisis”. This hierarchical classification, encompassing a wide spectrum of news topics and event types, mirrors the complexity and diversity of real-world news dissemination. Beyond categorization, the news events were also analyzed for their significance levels, ranging from local and community-level events to those with national and global implications. For example, the Russia-Ukraine war or Hamas-Israel conflicts are typical examples of news that involve multiple nations and countries, making them significant at a global level. This multi-faceted analysis, coupled with the sheer volume of the dataset, provides a unique opportunity to assess the AI-Chatbot’s ability to discern the importance and relevance of news events in a dynamic information environment.

The creation of this massive event dataset, utilizing the methodology and technology stack outlined in previous research [13,19,25], represents a significant contribution to the field of AI-powered news analysis. Since the focus of this study is the proposed News AI-Chatbot and not the News dataset, the same technology stack as portrayed in [13,19] (i.e., Dataverse for storage/management of News reports and Microsoft Power Automate for acquiring the News reports) was used for the creation of the News dataset. LLM-based technology was utilized for categorizing and analyzing these News articles as depicted in [25]. This robust News dataset serves as the cornerstone for evaluating the functionality and performance of the designed AI-Chatbot, paving the way for a more nuanced and informed understanding of current events in the digital age.

Upon integration of the meticulously curated news dataset with the AI-Chatbot, a series of rigorous evaluations were conducted to assess its real-time news retrieval and summarization capabilities. Figure 3 illustrates a representative interaction where a user queries for “Cyber” related news on a global scale (Figure 3a). Within a few seconds, the AI-Chatbot responds with a concise summary of critical cyber events that occurred within the preceding 24 h (Figure 3b). Notably, the chatbot’s response includes clickable references, providing transparency and facilitating further exploration of the reported events.

This rapid and informative response is facilitated by the intricate interplay of the AI-Chatbot, the Power Automate RPA, and the Google Gemini LLM, as depicted in Figure 1 and Figure 2. Figure 4 offers a detailed visualization of the Power Automate RPA workflow, highlighting the real-time processing and delivery of cyber-related global news. The initial HTTP call, responsible for generating the FetchXML query, is executed within one second, demonstrating the efficiency of the LLM-driven query generation process. Code 1 provides a snapshot of the generated FetchXML query, showcasing its precision in targeting cyber-related news reports within a global context.

Analysis of Code 1 reveals the precision with which the FetchXML query targets specific news subcategories, including “Cybersecurity news,” “Nation State Hacking,” “Globally Disruptive Cyber Attack,” and “Ransomware Attack News” (lines 14-17). This granular filtering, tailored to the user’s “Cyber” query, enabled the database to efficiently retrieve 28 relevant news articles within 2 s, as evidenced by the “List rows” step in Figure 4. This rapid retrieval underscores the efficiency of the query execution process and the database’s ability to handle real-time information requests. Following retrieval, only 4 columns are passed to “HTTP 2” (as shown in Figure 5). “HTTP 2” step invokes the LLM to summarize these 28 cyber-related news articles, prioritizing them based on their significance level (line 11, Code 1). This summarization process, however, necessitates careful consideration of ethical AI practices. As news articles, particularly those related to politics, crime, military, or cyber events, may contain sensitive or potentially harmful content, the LLM’s summarization could be inadvertently hindered by safety protocols designed to prevent the generation of inappropriate or offensive outputs. To mitigate this potential obstacle, specific settings were applied to the “HTTP 2” step (Figure 6), allowing the summarization to proceed even for news articles flagged as potentially containing “Hate Speech”, “Harassment”, “Dangerous Content”, or “Sexually Explicit” material. This strategic configuration, leveraging the Toxicity (text) function previously discussed in Section 3.1.4, ensures that the LLM can effectively summarize the news without compromising ethical AI principles.

The final output, presented in Figure 3b, comprises a concise summary of the cyber-related news, along with their corresponding sources. This presentation not only provides users with a quick overview of the key events but also maintains transparency and facilitates further exploration of the reported news. The ability to efficiently retrieve, summarize, and present relevant news articles in real-time, while adhering to ethical AI practices, highlights the potential of this AI-Chatbot to revolutionize news consumption and empower users with knowledge.

This seamless integration of technologies enables the AI-Chatbot to effectively navigate the vast news database, extract relevant information, and deliver concise summaries to users in real time. The chatbot’s ability to provide timely and informative responses, coupled with its transparent referencing of news sources, underscores its potential to transform news consumption and empower users with knowledge.

Code 1: Fetch XML Generated by LLM for Querying the News Database
1.	<fetch version=“1.0” output-format=“xml-platform” mapping=“logical” distinct=“false”>
2.		<entity name=“news_database”>
3.			<attribute name=“dfs_newsdatabaseid”/>
4.			<attribute name=“dfs_title”/>
5.			<attribute name=“dfs_description”/>
6.			<attribute name=“dfs_eventcode”/>
7.			<attribute name=“dfs_sourceurl”/>
8			<attribute name=“dfs_firsteventcountry”/>
9.			<attribute name=“dfs_secondeventcountry”/>
10.			<attribute name=“dfs_rating”/>
11.			<order attribute=“dfs_rating” descending=“true”/>
12.			<filter type=“and”>
13.				<condition attribute=“dfs_eventcode” operator=“in”>
14.					<value>Cybersecurity News</value>
15.					<value>Nation State Hacking</value>
16.					<value>Globally Disruptive Cyber Attack</value>
17.					<value>Ransomware Attack News</value>
18.				</condition>
19.			</filter>
20.		</entity>
21.	</fetch>

5. Discussion

The burgeoning field of AI-Chatbot development has captured the attention of researchers across diverse domains, including medicine, computer science, marketing, and business, as evidenced by the multidisciplinary applications highlighted in Table 1. However, a critical examination of 45 existing AI-Chatbot studies revealed a significant gap in their application to efficient news delivery, as underscored in Table 2. This research addresses this gap by presenting a novel AI-Chatbot specifically designed for real-time news analysis and summarization, distinguishing itself from previous works through its integration with a massive dataset of approximately one million records and its public accessibility across various platforms (https://web.powerva.microsoft.com/environments/a00ca161-b640-eae8-9caa-8058a3d7ae19/bots/crd69_copilot/canvas?__version__=2&enableFileAttachment=false accessed on 1 December 2024). The code for this AI-Chatbot is provided in the Appendix A.

Unlike the majority of reviewed papers that evaluated their AI-Chatbots with limited datasets and restricted accessibility, this research provides a publicly accessible AI-Chatbot deployed across multiple platforms, including a standalone web-based deployment (Figure 3), mobile platforms (Figure 7), and Microsoft Teams. This multi-platform accessibility, coupled with the chatbot’s integration with a massive news dataset, allows for a more robust and comprehensive evaluation of its performance in real-world scenarios.

5.1. Performance Evaluation

While 43 out of the 52 reviewed papers conducted performance evaluations of their AI-Chatbots [1,2,3,4,5,6,7,8,10,11,12,13,14,15,16,17,18,19,20,21,22,23,25,26,28,29,30,31,32,33,34,37,38,39,40,42,43,44,45], this research presents a more rigorous and comprehensive evaluation, leveraging a significantly larger dataset (as shown in

Table 6. Number of News Articles used for research in previous studies in comparison with this study.

Reference	Number of News Articles	Chatbot Integration
[25]	33,979	No
[13,15,19,20]	22,425	No
This Study	989,432	YES

) and diverse deployment strategies. A user study involving 35 participants across the US, UK, and Australia, who executed 321 news queries tailored to various categories and geographic locations (

Table 7. Detailed Performance Evaluation of the Proposed News AI-Chatbot.

User Topic of Interest	Location	TP	FP	FN	Precision	Recall	F1-Score
Cyber	USA	34	1	1	0.971429	0.9714	0.97143
Cyber	Global	33	0	1	1	0.9706	0.98507
Cyber	UK	39	1	2	0.975	0.9512	0.96296
Crimes	UK	21	1	3	0.954545	0.875	0.91304
Elections	USA	15	1	2	0.9375	0.8824	0.90909
Elections	UK	23	1	2	0.958333	0.92	0.93878
Politics	Global	28	2	2	0.933333	0.9333	0.93333
Politics	USA	13	3	2	0.8125	0.8667	0.83871
Politics	Australia	9	0	0	1	1	1
Business & Finance	USA	41	2	1	0.953488	0.9762	0.96471
War	Global	9	0	2	1	0.8182	0.9
Terrorism	UK	35	3	1	0.958785	0.991	0.97464
Total		300	15	6	0.957995	0.9874	0.97249

), demonstrated the chatbot’s superior performance, achieving precision, recall, and F1-score of 0.96, 0.99, and 0.97, respectively. For example, a user would execute “Political” News applicable to “USA” (as shown in Figure 7), if all the News reports are actually “Political” in nature and more applicable to “USA”, then the response is treated as True Positive (TP). For erroneous responses (e.g., non-political News or non-USA related News) it would be classified as False Positive (FP) or False Negative (FN). This rigorous evaluation, based on a clearly defined classification of responses as TP, FP, or FN, underscores the chatbot’s accuracy and effectiveness in delivering relevant news information.

Overall, the table provides strong quantitative support for the effectiveness of the AI-powered news chatbot in delivering accurate, relevant, and timely news summaries to users. The recall scores are even more impressive (averaging 0.9874 overall), suggesting that the chatbot rarely misses relevant News articles. In addition to the quantitative evaluation metrics (TP, FP, FN, Precision, Recall, and F1-Score) outlined in Table 7, Figure 4 demonstrates that the system’s summarization process required only 7 s in a specific instance involving recent cyber-related events, closely aligning with the average response time of 9 s. In the worst-case scenario (i.e., for summarizing political events), the response time is extended to 25 s. Beyond these objective performance measures, feedback from the 35 test users yielded an average quality rating of 4.3 out of 5 for summarization accuracy. This real-time news summarization, powered by GPT, is also enhanced with Microsoft Text-to-Speech avatars, allowing for realistic voiceovers of the summaries. Daily news summaries from the past 100 days, spoken by a live avatar, are accessible on the author’s LinkedIn profile at https://www.linkedin.com/in/fsufi/recent-activity/videos/ (accessed on 1 December 2024). Furthermore, several hundreds of daily textual summaries on cyber, disaster, and public health-related News are also publicly available on the author’s Medium blogs (i.e., https://medium.com/@research_60508 accessed on 1 December 2024). These texts and videos of News summaries provide evidence towards the superior summarization quality of LLM technology used in this study. This objective (i.e., Precision, Recall, F1-Score, and Response Time) and subject performance (summarization quality), coupled with the chatbot’s accessibility across various platforms, highlights its potential to transform news consumption and empower users with knowledge in the digital age.

5.2. Robustness of Solution

A key strength of this AI-powered news chatbot lies in its robust handling of diverse news classifications and user queries. The underlying complexity of the news data, coupled with the vast possibilities for user requests, necessitates a system capable of adapting to a wide range of scenarios. To illustrate the robustness of our solution, consider the combinatorial nature of the news data itself. Each news article can be classified according to four parameters. The first parameter denotes the number of possible countries (c₁). The second parameter denotes the possible countries allowing for events involving two countries (c₂). The third parameter denotes the number of News sub-categories (c₃). The last parameter denotes the number of significant levels (c₄). Thus, the total classification denoted as TC is given as follows:

$$TC=c_1\times c_2\times c_3\times c_4$$

(4)

There are 195 possible countries, 202 distinct news sub-categories, and 7 significance levels (Local, Community, State, Region, Nation, International, Global). This yields TC = 195 × 195 × 202 × 7 = 53,916,650.

This vast number of potential classifications highlights the inherent complexity of the news data and the challenge of retrieving relevant information based on user queries. The chatbot, however, demonstrated remarkable robustness in navigating this complexity. The user study involving 35 geographically dispersed individuals, who conducted 321 trials with diverse news queries, revealed the chatbot’s ability to effectively handle this complexity. This empirical evidence, coupled with the mathematical demonstration of the vast number of potential news classifications, underscores the robustness of the proposed solution in addressing the dynamic and multifaceted nature of news data. This robustness ensures that the chatbot can effectively cater to a wide range of user information needs, providing a personalized and efficient way to stay informed in today’s dynamic news landscape.

As mentioned in the previous section, the GPT model, integrated with Microsoft Power Automate, classifies news and determines its significance level (e.g., local, national, and global) by leveraging its understanding of language, syntax, and context. Furthermore, balanced coverage across political, cultural, and social domains is achieved through the aggregation of news from approximately 2342 diverse sources (i.e., robust real-time News monitoring), allowing for multiple perspectives on the same events, thus mitigating potential biases.

5.3. Ethical AI and Handling of Bias

To address bias mitigation and ethical AI practices effectively, it is essential to consider both proactive measures and real-time moderation techniques in chatbot design [34,47,60]. Recent work in AI ethics highlights the importance of transparency, privacy, and bias mitigation to foster responsible AI systems [34,60]. In implementing these measures, our chatbot employs multiple layers of filtering mechanisms and ethical guidelines. For instance, filtering tools are configured to identify sensitive or controversial topics, adapting responses to remain neutral and factual without promoting any specific viewpoint. This approach ensures that while the chatbot remains informative, it avoids escalating potentially contentious issues. The design incorporates regular updates to maintain a dataset that reflects diverse perspectives, addressing a common concern about inherent bias in language models [34]. As shown previously in Figure 6, Google Gemini API used within this study fully supported parameterized settings on effectively handing sensitive content like Sexually Explicit, Hate Speech, Harassment, or Dangerous Content. Sensitive content filtrations like these are currently supported by all major LLM providers like OpenAI’s ChatGPT, Microsoft’s Copilot etc.

Furthermore, transparency is a foundational component of ethical AI [47]. As discussed in prior literature, ChatGPT-based systems can inadvertently reflect societal biases unless continuously monitored [60]. Our chatbot’s architecture, therefore, includes feedback loops where flagged responses undergo review, allowing for adjustments that align with ethical AI standards. This is complemented by implementing data governance protocols that protect user privacy while allowing for data analysis that promotes improvement without violating ethical principles. In doing so, the aim is to ensure that the chatbot upholds fairness, accountability, and privacy across its interactions, embodying best practices in AI ethics.

In short, recognizing the critical importance of ethical AI practices, the design incorporates multiple strategies for detecting and mitigating bias. The news dataset was curated from diverse, credible sources, ensuring representation across geopolitical and cultural domains. Additionally, during summarization, the Google Gemini LLM applies content filters to identify and neutralize potentially biased or harmful content. These include toxicity checks and cross-validation with reference datasets to detect anomalies. The chatbot also employs a human-in-the-loop process, wherein flagged outputs are reviewed periodically to assess fairness and neutrality. By addressing bias systematically at both the dataset and generation levels, this system adheres to responsible AI principles and minimizes the risk of perpetuating misinformation.

6. Conclusions

This study advances AI-powered news delivery by introducing a unique chatbot that provides customized news summaries and real-time event analysis. Unlike the 52 previously reviewed AI-chatbot papers, which focused on static information retrieval or pre-defined conversations, this chatbot leverages generative AI and real-time data integration for a dynamic, personalized news experience. It fills a crucial gap identified in the literature by offering a standalone solution that connects to live news databases for real-time analysis, a feature that is absent in the reviewed papers. This is achieved through a unique architecture combining conversational and generative AI (Microsoft Copilot), RPA (Power Automate), a massive news database (989,432 reports from 2342 sources in Dataverse), and an LLM (Google Gemini). This architecture enables the chatbot to understand user intent, dynamically generate LLM-based FetchXML queries, and deliver concise, relevant summaries using generative AI.

The chatbot’s robustness is evident in its ability to handle a vast combinatorial space of potential news classifications (53,916,650 unique combinations) arising from categorizing news by country, sub-category, and significance level. This robustness is further validated by a user study involving 35 geographically dispersed individuals who conducted 321 trials, demonstrating the chatbot’s capacity to accurately and efficiently respond to diverse news queries. The chatbot achieved an impressive F1-score of 0.97, recall of 0.99, and precision of 0.96, showcasing its superior performance compared to existing AI-chatbot solutions in news analysis. Furthermore, the chatbot’s deployment on both Microsoft Teams and as a standalone web application highlights its practical usability and accessibility. This research paves the way for future AI chatbot applications in news analysis, promising to transform news consumption and foster a more informed citizenry.