GoKawiilAI recursion experiment shows Nano Banana 2 image edits spiral into chaos while text settles into 'LLMese'
A security researcher tested how generative AI models behave when fed their own outputs repeatedly, using Nano Banana 2 to make a video poster undergo successive localized edits until the image degraded into visual noise. Testing the same recursive approach on text rewrites, the researcher found that unlike images, text doesn't spiral into total chaos because encoding is lossless and per-token entropy is low—but open-ended rewrites still drift substantially from the original, eventually stabilizing into a distinct, homogenized style.
GoKawiil's interpretation of the reporting above, not reported fact.
These findings matter because recursive self-feeding is the core mechanic behind AI agent loops, and understanding when and how such loops degrade or stabilize has direct implications for debates about model collapse and claims of runaway AI self-improvement. The contrast between image and text failure modes suggests that reliability limits are highly modality-specific, meaning safety and quality assumptions can't simply be ported from one AI domain to another.
- Recursive image editing with Nano Banana 2 accumulates errors quickly, producing visual chaos after limited iterations.
- Text-based recursive edits are more stable due to lossless encoding and low injected entropy, especially for narrow tasks.
- Open-ended text rewrites still drift significantly from the original, often converging on a generic 'LLMese' style rather than diverging endlessly.
Source: blog.coredump.cx, 2026-09-07
Published there as: “Recursion into Madness”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.