Alignment research seeks to make AI behavior accord with intended human goals. Anthropic's 2026-04-14 study automated iterations on weak-to-strong training using an objective score. Transfer—applying a developed method to different tasks or training settings—was mixed on held-out tasks, which were not used to develop it. Testing at the larger scale used for deployed models found no statistically significant improvement. The authors say the setup unusually favors automation and human oversight remains essential. This is a bounded research proxy, not general-purpose alignment science. They can currently interpret the agents' work; harder-to-verify future 'alien science' is a possibility they raise, not a demonstrated loss of understanding.

Evidence status: developer_reported_results_with_transfer_limits.