AI image editors break after 100 edits — and the tools meant to catch this fail completely

Researchers built a dataset of 28,000 images edited repeatedly 100 times each, and found that each editing step introduces small artifacts that pile up into visible garbage. The standard tools used to measure image quality don't catch this degradation — meaning bad synthetic images could slip through quality filters and poison the training data for future AI models.

Right now, AI systems that edit images through multiple steps are reliable in theory but fragile in practice. Every edit adds invisible damage that accumulates fast, but the measurement tools everyone uses are blind to it. This creates a hidden failure mode: AI systems trained on this degraded synthetic data will themselves degrade, and nobody's measuring the loss because the measurement tools don't work.