From Objects to Influences: Rethinking Deletion in Learning-Based Systems
* denotes equal contribution.
Proceedings of the Extended Abstracts of the 2026 CHI Conference on Human Factors in Computing Systems
Machine UnlearningDeletionHCIAI Ethics
What should Delete mean after a system has learned from data?
Deleting a file feels like removing an object. But a learning system may already have used that file to change its model. Removing the original data and reducing its learned influence are different operations, even if an interface presents both with the same familiar button.
Examine the assumptions behind the button
This conceptual paper draws on HCI literature to examine four expectations people bring to deletion: control over the operation, its scope, its completeness, and transparency about what happened. We explore how each expectation is strained when information becomes distributed influence within a learned model.
The interface can promise more than the system does
The central concern is a mismatch between the user’s mental model and the system’s behavior. A familiar removal metaphor may leave people with an inaccurate understanding of what remains. Technical work on machine unlearning is important, but it does not by itself explain the operation to the person requesting it.
Make the scope of deletion understandable
We frame deletion as an interaction design problem and propose a user-centered research agenda around learned influence. This is a conceptual contribution, not an empirical test of a deletion interface or a new unlearning algorithm. Its purpose is to ask what systems should communicate and what meaningful control should look like.
Read the original abstract
Deletion is a foundational operation in digital systems, yet its meaning has shifted as computing infrastructure has evolved. In non-learning-based systems, deletion targets identifiable data objects; in learning-based systems, data is transformed into learned influences distributed across model parameters, fundamentally changing what deletion can accomplish. Despite this shift, interfaces continue to present deletion as a familiar object-removal operation, leaving users to rely on outdated assumptions. Drawing on prior HCI literature, we synthesize the recurring assumptions users bring to deletion—Control, Scope, Completeness, and Transparency—and show that each is placed under structural strain when deletion shifts from object removal to influence mitigation. This mismatch between user expectations and system behavior reveals a critical gap that technical approaches alone cannot address. By framing deletion as an interactional concern, this work positions it as a critical HCI problem and outlines a user-centered research agenda for rethinking deletion in the context of learned influence.