Unlearning Machines: Toward Sustainable Forgetting in AI

Why Forgetting Matters

The ability to forget is a cornerstone of intelligence, both biological and artificial. Yet while human memory evolves through selection and decay, machine memory is persistent and cumulative. Once trained, models tend to “remember everything,” including sensitive data, biases, or copyrighted materials.

This poses serious challenges in an age of data regulation and public scrutiny. The right to be forgotten, enshrined in frameworks such as the GDPR, is incompatible with traditional training pipelines that rely on irreversible optimization. Retraining a large model every time data must be removed is not only impractical but environmentally unsustainable.

The field of machine unlearning addresses this dilemma by seeking efficient methods to remove unwanted data influence from AI systems without having to start over. As Nguyen et al. (2025) show in their comprehensive survey, unlearning combines techniques from optimization, information theory, and differential privacy to achieve selective forgetting while preserving model utility.

From Privacy to Sustainability

Although unlearning first emerged as a privacy safeguard, its implications extend much further. By allowing models to update selectively, it can drastically reduce the energy and computational cost associated with retraining. According to Bourtoule et al. (2021), even partial retraining processes can consume significant resources, whereas targeted unlearning operations can achieve similar effects with a fraction of the energy footprint.

In this sense, unlearning becomes a form of digital resource conservation, a practice that supports both ethical and environmental goals. It also bridges the gap between privacy-enhancing technologies and the sustainability agenda, demonstrating that data protection and ecological efficiency need not be competing priorities.

The idea is simple but powerful: sustainable AI is not only about using greener servers or cleaner energy, but also about training smarter and forgetting efficiently.

Unlearning in the Era of Large Language Models

The challenge becomes even more complex with large language models. As Liu et al. (2025) explain in Rethinking Machine Unlearning for Large Language Models, forgetting in these massive architectures requires balancing multiple constraints: maintaining linguistic fluency, avoiding “relearning,” and verifying that the removed information truly disappears.

The authors identify two major categories of unlearning.
First, data influence removal, which focuses on eliminating the impact of specific training data, for example, copyrighted content or personal information inadvertently memorized.
Second, capability unlearning, which aims to suppress harmful model behaviors, such as generating toxic, misleading, or unsafe responses.

Both forms rely on highly targeted interventions in model parameters or gradients. These methods include gradient ascent approaches that counteract learning, influence function-based strategies that estimate which weights to modify, and localization-informed methods that identify where in the model knowledge resides.

Liu et al. also introduce the concept of adversarial unlearning, in which forgetting is reinforced through adversarial training loops that test the model’s ability to resist recovering deleted knowledge.

The Verification Challenge

One of the hardest problems in machine unlearning is verification, ensuring that forgetting has truly occurred. Residual traces of deleted data can remain embedded in model weights, producing what practitioners call pseudo forgetting. This makes the unlearning process both a technical and an epistemic issue, because it is difficult to prove that a model no longer behaves as if it had seen the deleted data.
Recent research, such as the already mentioned Liu et al. (2025), has explored influence auditing and forgetting fidelity metrics, which measure how much a model’s behavior changes relative to its original state. These methods challenge developers to create rigorous, standardised tests for forgetfulness and are essential to establish trust and accountability in AI, especially as unlearning becomes a regulatory expectation.

A Practice of Digital Sustainability

At its core, machine unlearning represents a paradigm shift: from a model of intelligence that accumulates endlessly to one that learns, adapts, and forgets with purpose. It transforms forgetting into a virtue, a form of digital self-regulation that aligns computation with both ethical and ecological boundaries.

For digital sustainability, this evolution is profound.
By integrating privacy, efficiency, and transparency, machine unlearning demonstrates how technological design can evolve toward responsible intelligence. Forgetting is not failure; it is maintenance. In the future, sustainable AI systems will not be those that remember everything, but those that remember wisely.