Google's AI Safety Test Has a Quarter of Its Questions Broken

Google DeepMind recently released a safety report for its latest AI model, Gemini 4. The report checks whether the AI could improve itself—a concept called recursive self-improvement (RSI). To do this, they used an internal test called Google RE-Bench (GRB), which has 74 tasks related to AI development. The idea is that if the AI can complete 90% of these tasks, it's considered critically capable and might pose a risk. But here's the catch: the report itself says about a quarter of the tasks had bugs that made them impossible to pass. That means even a perfect AI would fail those tasks, skewing the results.

Why does this matter to you? As AI gets more powerful, we rely on these safety tests to ensure it doesn't go out of control. If the tests are flawed, we might miss warning signs. For example, the report also mentions that some evaluations were rushed, using checkpoints from just days before the final model. This suggests that even top experts struggle to keep up with the pace of AI development. If Google can't properly test its own AI, how can we trust that it's safe?

The catch is that Google is aware of the issue and set a conservative threshold of 90% to account for bugs. But with 25% of tasks broken, that margin might not be enough. This isn't to say Gemini 4 is dangerous—it's a reminder that safety testing is hard and imperfect. As AI advances, we need better ways to measure its capabilities and risks. For now, this report highlights the challenges of keeping AI safe while racing to build it.