Synthetic imaginaries of “sensitive” AI: On ambient amplification and jail(break)ing as method

In this article, we propose the concept of “synthetic imaginaries” to map the complex hierarchies of (in)visibility perpetuated by generative AI models. Our analysis explores how GPT-4o reimagines sensitive visual content in line with the platform's content policy restrictions. To better understand its operative logic, we developed a derivative approach: starting with images as inputs, we co-created stories around them to guide the generation of new, story-based image outputs. In the process, we employed iterative prompting that blends “jailbreaking”— eliciting responses the model would typically avoid—with “jailing,” or reinforcing platform-imposed constraints. We found that, when confronted with sensitive prompts, GPT-4o operates in a default mode of ambient amplification, giving rise to synthetic imaginaries steered by safety protocols and normative value alignments. Rather than highlighting the subject matter in the foreground, GPT-4o outputs foreground the background, transforming tension into atmosphere, conflict into harmony, and controversy into an aestheticized mood.