Bell tests with random measurements require very high detection efficiencies

The observation that violating Bell inequalities with high probability is possible even when the local measurements are randomly chosen, as occurs when local measurements cannot be suitably calibrated or the parties do not share a common reference frame, has recently attracted much theoretical and experimental efforts. Here we show that this observation is only valid when the overall detection efficiency is very high ($\eta \geq 0.90$), otherwise, even when using the highest detection efficiency of recent photonic Bell tests, the probability of demonstrating nonlocality is negligible (e.g., it is smaller than 0.02% for $\eta=0.785$). Our results show that detection efficiency is a much more critical resource for real-world applications than it was previously thought.