Can code review models be trained on non-stolen data?
Understand this story
Some authors ask whether it is possible to remove the practices that users find objectionable in generative AI chatbots while retaining key capabilities such as code review. The first post in this series focuses on the issue of training large language models on copyrighted data without authorization, and asks whether a model with code review capabilities comparable to Claude's can be trained on "non-stolen data." The authors list the relevant doubts as open research questions and do not provide verified conclusions or feasible solutions.
AI overview based on source reports · Updated 19 min ago
Coverage timeline
Follow the coverage to explore different perspectives.
- Hacker News AICan We Fix AI?
This first part of a series asks whether AI can be built without objectionable practices, focusing on training on copyrighted data without consent. It examines whether models and their outputs are legally derivative works, introduces the Common Pile of public-domain and permissively licensed text (about 8TB), and poses research questions on matching frontier performance with openly licensed data, detecting derivative outputs, and closing the data gap via synthetic data and self-play.
Activity for this story
Not enough continuous observations to show a trend yet.