Benjamin et al., CHI 2023 · via Shadowplay (TEI ‘24)
What it says
Borrowed from vision: entoptic phenomena are the things you see that are produced by your own eye rather than by the world — floaters, afterimages. Transferred to machine learning, they are the marks the model’s internal machinery leaves in what it outputs. Not noise, and not the subject either. Evidence of the apparatus.
Does it hold for my work
Same condition as ML uncertainty — it needs a model in the loop, and the current work does not have one.
Where it was true: the Brutalist work’s outputs were full of these. What the audience read as architecture was often the model’s own habits showing through a body.
The reason to keep it separate
Uncertainty is about unpredictability; entoptic phenomena are about authorship. They answer different questions. If I ever have to argue that the machine is a participant rather than a tool, this is the concept that gets me there — and it does it without the weight of Barad.