Google DeepMind Conducts The First Double-blind Experiment Of An Artificial Intelligence

The world’s first double-blind test of AI, using cryptography to prevent tampering. Google DeepMind, in collaboration with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons, performed the world’s first double-blind test of a language model, the Gemini Flash Lite model. Double blinding refers to a process in which the test questions are hidden from the developers and the model’s weights are hidden from the evaluators. This way, neither party can alter their approach based on what they know about the other’s input.

1How the Anti-Manipulation Safeguards Work

To do this, Google DeepMind used Google Cloud’s Confidential Space, a cryptographically secure environment. This means the test is set up so the evaluators cannot get the model’s weights, and Google cannot see the test questions before the test. In the end, both parties get the test results, without ever being able to access the other party’s details.

2Why This Matters for the Entire Industry

AI benchmark contamination is a well-known issue in the AI industry. If the AI model gets to see the test questions during the training phase, its performance metrics will be unfairly skewed. In normal tests, the company knows the benchmark it’s going to be tested on, and can — intentionally or unintentionally — tune the AI model accordingly. Double blinding would eliminate this possibility. If the technique becomes widely adopted, we could finally get objective rankings of AI systems, altering how models are chosen and how we trust companies’ claims.