AI Transparency Falters as New Models Stay Secretive

AI models like GPT-4 are becoming more capable but companies like OpenAI are keeping key details secret. This reduces accountability and safety.

A Stanford study scored 10 major AI models on 13 transparency criteria. Even "open source" models scored low, with the most transparent only reaching 54%.

Secrecy contrasts with the openness of the previous AI boom, which helped progress in speech and image recognition.

Experts say more openness is needed to advance AI as a scientific field. AI2 is releasing an open model called OLMo to set an example.

Lack of access to training data makes it hard to understand model behavior. More openness could improve safety as deployment grows.