Decoding AV Safety: How to Read Disengagement Reports (2025)

One of the most cited metrics in autonomous vehicle (AV) safety is the "disengagement report." Every year, companies testing in California must report how often their human safety drivers took control of the vehicle. But is this number a true measure of safety? This guide breaks down what disengagements really mean and how to interpret the data like a pro.

What is a Disengagement?

A disengagement occurs when the autonomous mode is deactivated, and control returns to a human driver. This happens in two ways:

• Active (Human-Initiated): The safety driver grabs the wheel or taps the brake because they perceive a hazard or a violation of road rules.
• Passive (System-Initiated): The AV software detects a fault (like a sensor blockage or internal error) and asks the driver to take over.

The "Miles Per Disengagement" Trap

Media often rank companies by "Miles Per Disengagement" (MPD). A higher number supposedly means better technology. However, this metric is flawed for several reasons:

• Testing Difficulty: A company driving 1,000 miles on empty desert highways will have a fantastic MPD score compared to a company testing 100 miles in downtown San Francisco during rush hour. Context is key.
• Safety Culture: A conscientious safety driver might disengage early out of caution. A risky testing culture might discourage disengagements to boost the score, actually making the testing less safe.
• Software Versions: Reports are annual aggregates, hiding the fact that software version 10.1 in January might be vastly different from version 12.0 in December.

Better Metrics to Watch

If MPD isn't the gold standard, what should you look for?

• Collision Rates: The ultimate lagging indicator. How often does the AV actually hit something compared to human drivers in the same city?
• Mean Time Between Critical Events: Some companies report how often a disengagement was necessary to prevent a collision (simulation-verified), which is more meaningful than routine disengagements.
• Service Uptime: For robotaxi fleets, how often does the service go down or enter a "stalled" state blocking traffic?

Conclusion

Disengagement reports are a useful data point, but they are not a scoreboard. When evaluating AV safety, look at the full picture: where they drive, how transparent they are about incidents, and their real-world collision record.