Across social media, users engage in iterative prompting trends that push generative models beyond their expected outputs—trends like “make it more” or “create the exact replica 74times.” In these experiments, the same prompt is used to generate a new image from the last, producing long visual chains where meaning slowly drifts. Such sequences often go viral as evidence of “AI glitches”--moments when the model seems to cross boundaries-- such as turning a white woman into a Black man, or muscular biceps into a croissant. They captivate precisely because they appear to reveal a breakdown in the model’s logic. This work will argue, however, that these boundary crossings are not breakdowns but interpretative openings that expose the internal coherence of latent space. They make visible the friction between human and machinic epistemologies: while human meaning depends on categorical difference—taxonomical, oppositional, and historically shaped by hierarchies of identity—machine reasoning unfolds through gradients of statistical proximity. When a model drifts from one category to another, it does not violate a boundary but moves along a smooth topology of resemblance. The “boundary”arises only in human interpretation, where categorical expectation collides with gradient reasoning. Iteration lies at the core of understanding these encounters. Using iterative image generation as both phenomenon and method, the work will examine how boundaries and their
transgressions emerge only at the interface between human and machinic sense-making. The framework of forensic iteration approaches these moments as acts of epistemic translation, where two modes of reasoning briefly converge. In viral prompting trends, users repeatedly adjust prompts to probe how a model “understands” concepts, performing a lay form of model interrogation. This practice mirrors computational explainability methods such as neuron activation maximization, where researchers iteratively modify inputs to reveal the features that activate particular neurons. Both enact iteration as a tool of boundary crossing: a way to surface what is otherwise latent, to map the contours of meaning as it emerges through gradients rather than categories. By tracing how iteration bridges cultural practice and technical explainability, the paper repositions so-called “AI glitches” as productive sites for theorizing boundary, meaning, and the interpretative space between taxonomy and gradient.