AI labs criticized for secret model testing, limited safeguards

Two groups of AI experts have voiced concerns that leading AI labs are conducting internal tests of powerful models with critical safeguards disabled, and are not transparent enough about experimental details. One group, including former researchers from Ai2 and GovAI, argue that this secrecy limits outside scrutiny and undermines public safety. They warn that the current lack of openness makes it difficult to independently assess risks associated with advanced AI systems.

• AI labs run powerful models with safeguards turned off
• Safety tests may not reflect real-world model behavior
• Researchers call for greater transparency in AI development
• Recent incidents cited as examples of insufficient oversight