Why AV Safety Is Hard, and How Companies Try to Get It Right

If there's one topic that dominates the AV conversation, it's safety. And it should.

But AV safety is easy to oversimplify. "Are they safer than humans?" is a good question, but the answer is more nuanced than a simple yes or no.

Let's unpack why safety is hard, and what AV developers actually do about it.

Humans vs. machines: Different strengths, different failures

Humans:

• Are great at reading context and body language
• Can handle totally new situations with common sense
• But… get tired, distracted, intoxicated, or overconfident

Machines:

• Don't get sleepy or drunk
• Can pay attention in all directions at once
• But… can misunderstand context or fail in odd, specific ways

So the trade-off looks like this:

Humans are bad at statistics, good at context. Machines are good at statistics, bad at context.

AV safety work is about leveraging machine strengths while mitigating their weaknesses.

How AVs are tested

AV programs typically use a layered approach:

1. Simulation

• Millions of virtual miles
• Run rare and dangerous scenarios without risking real people
• Quickly test new software versions against a library of edge cases

2. Closed-course testing

• Private tracks where the system can practice safely
• Scenarios like emergency vehicles, pedestrians crossing unpredictably, tricky intersections

3. On-road testing with safety drivers

• Trained drivers sit in the front seat, ready to take over
• Every "disengagement" is logged and analyzed
• Data feeds back into simulation and software improvements

4. Gradual driverless deployment

• Start in limited areas, at limited times, under specific conditions
• Monitor performance, refine policies, expand over time

What "safer than humans" really means

It's not enough to say "AVs are safe." The real question is:

Over millions or billions of miles, how does the risk of serious crashes compare to human driving?

That's a statistical question. Because:

• Crashes, especially severe ones, are relatively rare.
• You need a lot of data to draw strong conclusions.
• Results can vary by city, conditions, and system version.

Some companies publish safety reports and metrics (like collisions per mile, or comparisons to human baselines), but these are still evolving.

When AVs make mistakes

When an AV is involved in a crash or a close call, the follow-up process often looks like this:

• Immediate halt or restriction of similar operations, if needed
• Detailed review of logs and sensor data
• Root-cause analysis: perception issue? prediction error? policy gap?
• Code and policy changes
• Re-testing in simulation and on test tracks before resuming

The goal is not just to fix that specific case, but to improve the system class of behaviors so similar risks are less likely across the fleet.

Perfection is impossible, but the bar is high

No transportation system is perfect. Human driving certainly isn't — globally, traffic crashes kill over a million people per year.

AVs don't have to be perfect to be valuable. But they do need to meet a very high bar: ideally, demonstrably safer than typical human driving under comparable conditions.

That's why:

• Deployment is gradual.
• Operational domains are limited.
• Policy and safety engineering are as important as raw technology.

How to think about AV safety as a non-expert

You don't need to read technical safety reports to ask good questions. Here are a few:

• In what areas and conditions is this AV allowed to operate?
• How is its performance measured and monitored over time?
• What happens after an incident — how does learning feed back into the system?
• Are there clear ways for passengers and other road users to report issues?

AV safety isn't a yes/no switch. It's an ongoing process of designing, measuring, and improving — with real engineering and real accountability behind the scenes.