New Research Explores How Everyday Machine Learning Training Can Reshape Human-System Boundaries

By Staff Reporters

When humans interact with software, the design of the tools they use does much more than simply facilitate tasks. According to a new study conducted by a team of researchers including Nava Haghighi, Danielle Olson, Halden Lin, Erdrin Azemi, Gierad Laput, Kayur Patel, and James Landay, designed artifacts are fundamentally ontological. They actively shape, and at times severely limit, what individuals believe is possible or even imaginable within a given technological ecosystem.

For decades, the standard path toward mitigating these conceptual foreclosures has involved giving people more direct power over how technical systems are designed and built. Numerous scholarship traditions have explored participatory design, end-user programming, and systems that allow individuals to act as authors of their own technological experiences. However, the research team points out a persistent blind spot in how these systems are traditionally evaluated. For the most part, user-empowering systems are judged strictly on whether they are usable, useful, or technically feasible. Crucially, this leaves deeper philosophical and conceptual questions—specifically regarding ontological boundary negotiation—largely unexamined.

To address this gap, the researchers set out to investigate what happens when everyday people are given direct, hands-on control over defining and training personalized machine learning systems based on phenomena they choose themselves. Rather than focusing merely on interface efficiency or task completion accuracy, the study explores how the very boundaries of user understanding shift when individuals actively construct the categories that a machine learning model is meant to recognize.

The methodology relied on the creation of two open-ended probes that utilized a Wizard of Oz technique. In human-computer interaction research, the Wizard of Oz method involves simulating automated system capabilities behind the scenes while participants believe they are interacting with a fully functioning autonomous system. In this study, these probes allowed participants to experience the process of training a personalized machine learning system on real-world phenomena of their own choosing, bringing abstract concepts of machine learning customization directly into their daily routines.

Over the course of a rigorous week-long exploratory study, participants integrated one of the two probes into their everyday lives. By living with these systems and actively teaching them to recognize personalized concepts, participants encountered moments where the rigid categories expected by computational systems clashed with the fluid, nuanced realities of human experience.

Through this immersive, real-world deployment, the research team identified four distinct sites where ontological boundaries were actively negotiated by the users. The first major site of negotiation centered around the boundaries of the phenomena themselves. Participants grappled with defining where a specific, individualized phenomenon began and where it ended, discovering that abstract concepts rarely fit into neat, discrete computational buckets.

The second site involved understanding the subject as part of broader relations. Users found that the phenomena they were attempting to teach the machine learning system could not be isolated from its surrounding context, forcing them to consider how relational dynamics influence data collection and categorization.

The third site addressed the eternal computational divide between what constitutes signal and what constitutes noise. As participants tried to train their personalized models amid the chaos of everyday life, they had to actively determine which real-world variations mattered to the system and which should be discarded as irrelevant background interference.

Finally, the fourth site tackled the perceived objectivity of data. By attempting to codify their own subjective experiences into training data for a machine learning model, participants confronted the inherent tensions between personal interpretation and the illusion of neutral, objective computation.

Based on these findings, the researchers offer concrete starting points for supporting ontological boundary negotiation through thoughtful system design. Rather than designing tools that obscure the boundaries of data categorization behind seamless interfaces, future systems might explicitly acknowledge and support the fluid nature of human categorization. Furthermore, the study highlights open-ended probes as a valuable, robust methodology for conducting ontological design research, demonstrating that living with customizable AI systems can reveal profound insights into how technology shapes human thought.

Related readings and updates

The broader discourse surrounding system design, privacy, and knowledge extraction continues to evolve across academic and technical forums. Among related developments, a paper authored by the research community has been accepted at the AI4TCI (Workshop on AI for Secure and Trustworthy Critical Infrastructure Systems) Workshop at the International Conference on Availability, Reliability and Security (ARES) 2026. This work examines the growing deployment of autonomous negotiation agents in high-stakes settings such as insurance and procurement. While cryptographic techniques successfully protect explicitly disclosed constraint values in these environments, the research highlights that they often fail to address subtler threats, pointing to ongoing challenges in securing agentic interactions.

In another related area, knowledge graphs remain foundational to a wide array of artificial intelligence applications, yet maintaining their freshness and completeness continues to incur significant operational costs. To address this, recent developments include ODKE+, a production-grade system designed to automatically extract and ingest millions of open-domain facts from web sources with high precision. ODKE+ combines modular components into a scalable pipeline, featuring an extraction initiator that detects missing or stale facts, alongside an evidence retriever designed to collect supporting information efficiently.