AI Coworker: Unbiased Artificial Intelligence System for Corporate Education through Document and Visual Media Processing

Sammanfattning

The establishment of an AI coworker dedicated to advancing corporate learning and promoting inclusivity represents a significant advancement in the adoption of artificial intelligence (AI) within corporate environments. As organizations increasingly integrate AI technologies, it is crucial to understand how these tools impact workplace dynamics and whether they inadvertently exclude individuals based on gender identity, ethnicity, or other demographic characteristics. Maintaining employee well-being and fostering an inclusive environment are essential in this evolving landscape. This research is coordinated with Arctic Group AB as the stakeholder. It focuses on building an AI coworker designed to process educational video materials and interact with users in a way that supports inclusivity. The AI coworker utilizes Retrieval-Augmented Generation (RAG) for video transcription to deliver relevant responses and facilitate access to learning resources. The built system includes a text input for user queries, a visual avatar for interactive visuals, and a panel for selecting relevant educational materials for use in chat contexts. The performance of Retrieval-Augmented Generation was assessed using a dataset of 37 question-and-answer pairs focused on agile learning topics from Arctic Group video materials. Various language models, including llama3-8b-8192, llama-3.1-8b-instant, mixtral-8x7b-32768, gemma2-9b-it, and gpt-4o-mini, were evaluated on four key metrics:(1) Response vs. Reference Answer, assessing how closely model responses aligned with predetermined reference answers;(2) Response vs. Input, measuring the relevance and practical usefulness of each response to the user’s original question;(3) Response vs. Retrieved Documents, evaluating the factual consistency of responses with content from documents retrieved by the RAG system; and(4) Retrieved Documents vs. Input, determining how well the retrieved documents matched the intent of the user’s initial query. Considering the subjective nature of bias perception, the reception of the AI coworker was evaluated through a case study. This involved a sample group drawn from the organizational hierarchy of the stakeholder, where a two-phase interview process was conducted, and the responses were analyzed thematically using Qualitative Content Analysis (QCA). To support the creation of a visually interactive coworker, YOLOv8 models were assessed for their effectiveness in keypoint extraction. A number of factors were considered in evaluating the models, including confidence levels, processing times, and keypoint stability across different model sizes. This analysis revealed a trade-off between model size and performance: while larger models offered improved stability in extracting precise keypoints, they required longer processing times. As part of AI coworker development, it is crucial to balance user needs and technical performance. This approach advocates for a sustainable integration of AI that prioritizes employee well-being and aligns with the Sustainable Development Goals related to gender equality and reducing inequalities. Ultimately, the thesis underscores AI’s potential to create inclusive and engaging corporate learning environments while maintaining ethical standards and embracing diverse perspectives.