Noise Poisoning

My work comments on how AI will eat itself. I introduce data-poisoned images into online circulation so that when AI systems scrape the web, they gradually ingest their own corruption. This is sabotage as art: using error to interrupt machine vision and to challenge its growing cultural authority.

Noise Poisoning intervenes in the visual economies of contemporary AI. I treat images as operational material—training data, scraped archives, and documentation produced primarily for machines rather than humans. In these systems, images are not representations but inputs. They are transformed into features and probabilities, then used to build models that classify, predict, and decide.

A key focus is the portrait. The faces I work with are already inside machine vision: some are AI-generated, others have circulated long enough to function like synthetic images. What matters is their role as training material. I work at the threshold where a portrait still holds “human structure,” but is increasingly shaped by a closed loop—models training on images made by other models. My interventions accelerate that drift, pushing recognition toward instability and making the face less reliable as a training object.

AI vision does not simply depict the world; it acts on it. It sorts bodies, flags difference, and automates judgment. Datasets are never neutral: they encode assumptions about normality, identity, and value, then reproduce those assumptions at scale under the appearance of objectivity.

My practice introduces controlled disruption into this process. The interference is often negligible to human viewers but significant for machine learning systems because it destabilizes the regularities recognition depends on. I am not interested in spectacle or collapse. I am interested in hesitation—moments where automated confidence weakens and classification becomes unreliable. In that sense, Noise Poisoning reframes sabotage as an ethical and aesthetic practice: a way of reclaiming uncertainty and opacity inside systems designed to extract stable features from everything they see.

Because the work engages real infrastructures, method is part of its ethics. I treat interference as artistic material, not a deployable technique. I avoid operational disclosure and focus on what the images make visible: that machine vision depends on compliant images to function, and that this compliance can be refused. I am committed to responsible disclosure and institutional dialogue, speaking clearly about intent and method without turning artistic practice into operational instruction.

In Noise Poisoning, sabotage is interruption. It reintroduces uncertainty into systems built for prediction and control. The project insists that AI vision is not neutral perception but a form of governance—and that governance can be disrupted from inside the image itself.