User Edits Could Turn AI Failures into Training Data

A new preprint on arXiv argues that the growing volume of human-AI interaction is an underused resource for improving AI systems. The authors note that while many model responses are already satisfactory, others expose failures—and if those failures can be identified and corrected, they could provide useful supervision.

The paper centers on the idea of "edit discovery": detecting when a user's modification to a model output signals a genuine flaw rather than a stylistic preference. By isolating these edits, it may be possible to turn routine interactions into a steady stream of training signal.

Because only the abstract is available, the specific methods and results are not yet detailed. But the framing points toward a future where user corrections, not just curated datasets, help drive model improvement.