Follow the latest coverage, related explainers and connected technology stories.
A report involving more than 20 researchers and specialists warns that automating artificial intelligence research and development could accelerate the evolution of systems to the point that humans’ ability to control them is weakened. The report suggests that some research projects requiring months could become fully automated by 2028, while calling for independent oversight and emergency plans before the window for action narrows.
A research paper from Anthropic showed that automated systems were able to improve a model’s performance on 10 tests of behaviors inconsistent with alignment goals, without reducing overall performance. The early findings suggest that part of alignment research could be automated, while the quality of benchmarks and research data remains a key constraint.