TLDR: A study identified five dynamic epistemic relationships between humans and AI in knowledge work: Instrumental Reliance (AI as a tool for efficiency), Contingent Delegation (AI as a supervised assistant), Co-agency Collaboration (AI as a thinking partner), Authority Displacement (granting AI partial knowledge authority), and Epistemic Abstention (using AI but denying its knowledge contribution). These relationships are fluid and depend on context, task, and user expertise, highlighting the need for a nuanced understanding of human-AI knowledge co-construction.
As artificial intelligence systems become increasingly integrated into our daily work, especially in fields that rely heavily on knowledge, a crucial question emerges: how do humans and AI truly interact when it comes to creating and using knowledge? Beyond just their functional capabilities, what are the deeper “epistemic roles” AI plays in our intellectual lives?
Traditional research on human-computer interaction has often categorized AI based on its functions – as a tool, an assistant, or a partner. However, this perspective often overlooks how AI fundamentally reshapes our own roles as contributors to knowledge. A recent study delves into this very question, exploring how users form unique “epistemic relationships” with AI – that is, how they assess, trust, and collaborate with these systems in contexts like research and teaching.
Defining Epistemic Relationships
In this study, an epistemic relationship is defined as the way human users interact with AI systems to evaluate, rely on, or negotiate the knowledge contributions of AI in their work. Instead of assuming fixed roles, the researchers investigated how these relationships are built and change dynamically across different situations.
The Study’s Approach
To understand these complex interactions, the researchers conducted 31 semi-structured interviews with academics from various disciplines and countries who regularly work with AI. They asked participants about their AI usage, their perceptions of AI’s role, and how they evaluate AI’s quality and outcomes. Based on these interviews and existing literature, they developed a five-part framework to categorize key dimensions of human-AI interaction, including the metaphor used for AI (e.g., tool, assistant), the type of task, the assessment perspective (outcome-based vs. process-based), the human’s expertise level, and the type of trust involved.
Five Ways Humans Relate to AI Epistemically
The study identified five distinct types of epistemic relationships that highlight the dynamic and context-dependent nature of human-AI interaction:
1. Instrumental Reliance: In this relationship, AI is seen purely as a tool to boost efficiency. Users rely on AI to get tasks done faster, without necessarily granting it any deep knowledge status. Their assessment is based solely on the outcome, and trust is limited to functional reliability. For example, a digital humanities expert might use AI to write Python scripts for data analysis, focusing only on whether the script works, not on the AI’s “understanding.”
2. Contingent Delegation: Here, users selectively delegate cognitive tasks to AI, but always maintain human oversight. AI is viewed as an assistant or co-agent, and trust is conditional – it develops only after the human verifies the AI’s output. An example is a researcher who treats AI like a research assistant, checking everything it produces before accepting it.
3. Co-agency Collaboration: This describes a more integrated and dynamic partnership where AI is treated as a thinking partner or mentor. Trust is higher and more flexible, evolving through ongoing interaction. Assessment considers both the final outcome and the underlying process. This is common in open-ended tasks like brainstorming or exploratory writing, where users might “converse” with AI as if it were a colleague, even anthropomorphizing it.
4. Authority Displacement: In this relationship, users acknowledge and grant AI partial epistemic authority in specific areas, allowing it to take a lead in knowledge production. This involves a high level of trust and often process-based assessment, especially in complex reasoning tasks. Users in this category often have hybrid expertise, combining domain knowledge with AI expertise, such as a programmer who “pair programs” with AI and accepts most of its suggestions.
5. Epistemic Abstention: This type captures situations where users utilize AI tools but actively deny them any epistemic contribution or authority. Trust is minimal or absent, and assessment remains outcome-oriented. An example is a teacher who has students use AI to draft papers but then critiques the AI-generated drafts for lacking specificity, expressing distrust in its performance. Some users in this category might even choose to avoid AI for core research tasks, believing the human process itself is valuable for becoming a better researcher.
Also Read:
- Designing AI for Human Well-being: A New Framework for Human-AI Interaction
- Navigating the Landscape of LLM-Based Data Science Agents: A Comprehensive Survey
Implications for the Future of Work
These findings suggest that our relationship with AI is not static; it’s a fluid, negotiated outcome shaped by the specific task, our expertise, and our level of trust. This understanding has significant implications for how we design AI systems, develop training programs, and structure workflows in knowledge-intensive environments. Recognizing these different epistemic relationships can help organizations better align AI deployment with user needs and support employees’ evolving roles as they increasingly collaborate with AI.
To learn more about this fascinating research, you can read the full paper: Classifying Epistemic Relationships in Human-AI Interaction: An Exploratory Approach.
