Google is trying to make AI testing more honest

Current AI tests are hard to trust because developers often know the questions ahead of time. Google is now trying a double blind method to change this. They want to prove that their models can pass tests without any cheating.

To make this work Google is using special security tech. This keeps the test questions hidden from Google and the model details hidden from the judges. It creates a private space where no one can manipulate the results.

They are running this first trial with experts in Singapore. If this test goes well it could become the new rule for how we measure if AI is actually safe and capable.