Memory Undone: Between Knowing and Not Knowing in Data Systems

AI summary

Overview

Research area: AI safety and ethics; specifically the sociotechnical study of memory, deletion, and machine unlearning in data and machine learning systems. The paper sits at the intersection of computer science, philosophy, memory studies, critical media theory, and governance.

Technical level: Beginner-Friendly. This is a conceptual position/abstract paper that introduces and connects ideas rather than presenting experiments, formalisms, or quantitative results.

Scope: A one-sentence scope: the paper argues that forgetting in data systems should be treated as an intentional, accountable sociotechnical practice—not merely a technical compliance feature—and distinguishes forgetting from related practices such as erasure, unlearning, and exclusion.

What This Paper Is About

Machine learning systems increasingly act as infrastructures of memory, collecting and learning from data that contains traces of personal, political, and cultural life. The paper asks what role a capability to forget should play in systems built by default to remember, given that legal mechanisms such as the GDPR's "right to be forgotten" do not map cleanly onto the technical reality that deleting records does not remove their influence on trained models. Its goal is to reframe machine unlearning as a sociopolitical intervention with consequences for privacy, redress, narrative amplification or suppression, and the silencing of underrepresented perspectives.

Key Contributions

1. A clarifying vocabulary. The paper separates four commonly conflated practices: erasure (an operational act removing or disabling access to data artifacts such as rows, files, indexes, and caches), unlearning (an intentional technical intervention to remove or bound the influence of specific data on learned parameters and downstream outputs), exclusion (upstream decisions not to collect, label, or include certain people, events, or categories—"forgetting by omission"), and forgetting as the umbrella sociotechnical term varying in agency, temporality, reversibility, and scale.

2. A reframing of forgetting as a sociotechnical practice. The paper argues that forgetting should be analyzed within structures of power, rights, epistemic justice, and governance, and treated not as an internal optimization goal but as a sociopolitical intervention whose consequences extend beyond the data system.

3. A dual-risk account of forgetting. It shows that erasure and unlearning function simultaneously as safeguard and risk: forgetting biased or harmful data may reduce harm, while forgetting underrepresented perspectives risks reproducing epistemic injustice.

4. A research direction toward "forgetting machines." The paper points toward computational systems capable not only of remembering but of forgetting responsibly by design—unlearning by design that is intentional, transparent, and accountable—and states three aims for its presentation: to present the reframing, to illustrate how unlearning intersects with power, rights, epistemic justice and governance, and to open discussion about designing epistemic infrastructures that forget responsibly.

Main Findings

• Deletion is not removal. Permanently removing information from large-scale systems is technically challenging because deletion must propagate through copies, caches, and derived data products, and most infrastructures were designed to accumulate rather than erase data. Deleting records from database storage does not remove their influence on models trained on that data; in deep learning, once information is integrated into the weights of a neural network, its traces remain entangled in billions of parameters.

• A quantum analogy for the limits of deletion. The paper compares machine unlearning to the quantum no-deleting theorem, which shows an unknown quantum state cannot be deleted against a copy. In machine learning, "deletion" is therefore often not complete removal but a redistribution or attenuation of influence across representations, with traces persisting indirectly through learned correlations, semantic dependencies, or downstream artifacts.

• Forgetting is already routine in data science. At every pipeline stage—data collection, feature design, labeling—choices are made about what to exclude, ignore, or treat as noise. The paper notes prior research showing data never exist in a neutral form, that preservation is typically the infrastructure default ("keep everything") while deletion appears as an exceptional intervention motivated by rights or harm reduction, and that this produces data silences and systematic erasures of certain groups or perspectives.

• Exclusions are engineered, not incidental. Studies of entity–relationship design reveal invisible entities—objects, actors, and relations that shape data models without being explicitly represented—often left undescribed to reduce complexity or maintain convenience. Data infrastructures contain nonrandom exclusions, where the absence of data about certain populations stems from systematic marginalization or perceived lack of economic or political value. Forgetting, in this sense, is an act of governance.

• Forgetting has dual consequences. Forgetting biased or harmful data may reduce harm, but forgetting underrepresented perspectives risks reproducing epistemic injustice.

• Information manipulation at scale is an epistemic-poisoning problem. Research on Russian disinformation networks shows how propaganda content, once statistically dominant in training corpora, can distort model outputs; large language models reproduce such narratives, sometimes without disclaimers. Whether unlearning can provide remedies for these epistemic distortions remains an open question.

• Technical unlearning progress is partial. The paper cites advances including exact unlearning in recommender systems, class forgetting without gradients, erasure in semantic web and description logic ontologies, and selective unlearning in large language models—while noting that fundamental questions remain unresolved.

• Known and not-known are entangled. The paper draws on memory studies treating remembering/forgetting as institutionally organized practices rather than individual cognitive events; on cognitive psychology showing human memory is reconstructive and susceptible to post-event framing effects; and on evidence that post-decision attitude change can occur even without explicit memory for the decision (sometimes discussed as "synthetic" preference or satisfaction).

• The database-as-fortress metaphor is limiting. Databases are often imagined as fortresses designed to protect and preserve memory, but memory is naturally selective and becomes effective by retaining certain information while skipping others. The paper argues forgetting is not an exception to knowledge but its condition, that knowing is never neutral (it assumes the power to select and keep authority over what is known), and that not knowing can act as a productive epistemic tool rather than a failure. It cites Francis Bacon's Idols of the Theater as a warning that dominant intellectual frameworks can hide alternative ways of knowing.

• Technology may be structurally misaligned with forgetting. Critical media theory argues technology functions primarily as a retentional device and is therefore structurally misaligned with forgetting; the paper takes this as a challenge to clarify which kinds of forgetting are possible, and which are not, in computational infrastructures.

• No quantitative results are reported. The paper contains no datasets, benchmarks, metrics, or empirical evaluations.

Methodology in Plain English

This is a conceptual and agenda-setting paper rather than an empirical study. The authors synthesize literature across machine unlearning, data management, participatory data modeling, memory studies, cognitive psychology, critical media theory, philosophy, ethics, governance, and human–computer interaction. They build a shared vocabulary by defining and separating erasure, unlearning, exclusion, and forgetting, then use that vocabulary to analyze examples—GDPR deletion rights, invisible entities in data models, missing population data, and disinformation-contaminated training corpora—to show how technical acts of forgetting carry political and epistemic consequences. The paper describes an ongoing "work-in-progress" investigation of memory from a sociotechnical perspective, and is framed as a presentation with three stated aims to engage conference participants in critical dialogue. It performs no experiments, uses no datasets, and reports no measurements.

Why This Matters

Impact on research: The paper pushes machine unlearning research beyond a narrow compliance framing, arguing that forgetting should be a first-class capability and a foundational step in the machine learning pipeline. It proposes connecting technical unlearning advances with philosophy, ethics, memory studies, and HCI, and treating knowing/not knowing as a research subject in its own right.

Real-world applications:

• Handling GDPR "right to be forgotten" requests in systems where deletion must propagate through copies, caches, and derived data products.
• Mitigating epistemic poisoning of model training corpora by statistically dominant propaganda and disinformation narratives.
• Designing datasets and data models in domains where missing populations result from systematic marginalization, including participatory data modeling with affected communities.
• Governing exclusion decisions during data collection, feature design, and labeling, where what is omitted can structure outcomes as powerfully as what is included.

Industry relevance: The paper argues for models capable of unlearning by design, with forgetting that is intentional, transparent, and accountable. This speaks to organizations that must reconcile legal deletion obligations with models whose learned parameters retain traces of the deleted data, and to those building knowledge infrastructures where preservation is currently the unexamined default.

Future Directions

• What does it mean for a system to truly forget?
• How can erasure be ensured in the presence of semantic dependencies, where one record influences others?
• At what point does forgetting protect rights, and when does it instead threaten them through silencing or erasure of underrepresented perspectives?
• Can unlearning provide remedies for epistemic distortions introduced by manipulative or statistically dominant content in training corpora?
• How can "forgetting machines" be built so that unlearning is intentional, transparent, accountable, and positioned as a principle for healthier knowledge infrastructures rather than treated as a limitation?

Target Audience

This paper benefits most readers who want a conceptual, cross-disciplinary framing of machine unlearning rather than technical methods: researchers and students in AI ethics and responsible AI, data management and data systems researchers interested in the social dimensions of schemas and query languages, policy and governance specialists working on data protection and the right to be forgotten, and scholars in memory studies, philosophy, HCI, and the social sciences who study how knowledge infrastructures include and exclude. It is also useful for practitioners who need to understand why deletion requests and unlearning interventions do not straightforwardly remove data influence from trained models.

Authors’ abstract

Machine learning and data systems increasingly function as infrastructures of memory: they ingest, store, and operationalize traces of personal, political, and cultural life. Yet contemporary governance demands credible forms of forgetting, from GDPR-backed deletion to harm-mitigation and the removal of manipulative content, while technical infrastructures are optimized to retain, replicate, and reuse. This work argues that "forgetting" in computational systems cannot be reduced to a single operation (e.g., record deletion) and should instead be treated as a sociotechnical practice with distinct mechanisms and consequences. We clarify a vocabulary that separates erasure (removing or disabling access to data artifacts), unlearning (interventions that bound or remove a data point influence on learned parameters and outputs), exclusion (upstream non-collection and omission), and forgetting as an umbrella term spanning agency, temporality, reversibility, and scale. Building on examples from machine unlearning, semantic dependencies in data management, participatory data modeling, and manipulation at scale, we show how forgetting can simultaneously protect rights and enable silencing. We propose reframing unlearning as a first-class capability in knowledge infrastructures, evaluated not only by compliance or utility retention, but by its governance properties: transparency, accountability, and epistemic justice.