Self-Driving Cars Are Getting Better — But Are Our Roads Actually Ready?

Last year I almost got sideswiped by a Tesla on autopilot merging onto I-95 near Stamford, Connecticut. The car drifted into my lane with this eerie smoothness, no turn signal, no hesitation — just a machine doing what it calculated was correct. The driver had his hands in his lap. I laid on my horn, my heart did something unpleasant, and afterward I sat in a parking lot thinking: we are genuinely living through one of the strangest transitions in transportation history.

That near-miss got me reading, and what I found is a lot more complicated than the "self-driving cars will save us all" narrative most tech coverage pushes.

How Far the Technology Has Actually Come Since 2020

The jump in autonomous vehicle capability over the past four years is real and I don't want to dismiss it. Waymo's robotaxi fleet in Phoenix, Arizona logged over 7 million rider-only miles by early 2024, which is a genuinely impressive number. Tesla's Full Self-Driving system has gone through at least a dozen significant software versions, with version 12 shifting to a neural-net-heavy approach that handles unstructured environments noticeably better than earlier rule-based systems. Companies like Mobileye are shipping hardware to automakers that processes over 2.5 trillion operations per second for real-time sensor fusion.

The underlying architecture — lidar, radar, cameras, HD mapping, and machine learning stacked together — is maturing fast.

But maturity and readiness aren't the same thing.

Where the Safety Numbers Get Complicated

Here's where I'll take a side: I think the current public conversation about autonomous vehicle safety is too binary. Either people say "self-driving cars are dangerous and we shouldn't trust them" or "the stats show they crash less than humans, so stop worrying." Both camps are cherry-picking.

The honest picture looks something like this:

• Waymo's own data suggests its vehicles have fewer injury-causing crashes per million miles than the national average for human drivers.


• But most AV testing happens in sunbelt cities with relatively predictable road layouts — Phoenix, San Francisco, Austin. Rural roads, heavy snow, and poorly marked construction zones are still serious weak points.


• A 2023 California DMV incident report logged 57 autonomous vehicle collisions in a single year across the state, most at low speed, but the reporting methodology makes direct comparison to human-driver stats genuinely difficult.

The edge cases are where people get hurt. A system that handles 99.97% of situations flawlessly and catastrophically fails the remaining 0.03% is not the same as a human driver who makes small errors constantly but has better improvisation in genuinely weird situations.

The Regulatory Gap Nobody Talks About Enough

Federal regulation in the US has not caught up with what's being deployed on public roads. The National Highway Traffic Safety Administration issued a standing general order in 2021 requiring AV manufacturers to report crashes, which is fine — but there's no unified federal licensing framework dictating what an autonomous vehicle must demonstrate before commercial deployment. Individual states are writing their own rules, which means a vehicle cleared to operate in Nevada may face totally different requirements in Georgia.

That patchwork is a problem.

Europe is moving faster on formal standards through the UNECE Working Party 29 framework, which set binding requirements for Automated Lane Keeping Systems in 2022. The US is still largely relying on voluntary guidelines, and that gap matters when something goes wrong and liability needs to be assigned.

What Needs to Happen Before I'd Feel Comfortable on That Highway Again

My honest list of what I think is missing before widespread autonomous vehicles are genuinely road-safe:

1. Standardized federal certification — not guidelines, actual pass/fail requirements.


2. Real-world performance data from non-ideal conditions: ice, heavy rain, unlined rural roads, active construction.


3. Clear liability law. Right now it's murky whether the manufacturer, the software provider, or the "operator" bears responsibility in a crash.


4. Mandatory incident transparency, not voluntary reporting with brand-friendly framing.


5. Better driver education on what Level 2 and Level 3 automation actually means — because most people using these systems right now don't fully understand the limits.

None of that is impossible. It's just slow, unglamorous regulatory work that doesn't get TechCrunch headlines.

The Part That Actually Keeps Me Optimistic

Human drivers kill about 40,000 people per year in the US. That number has barely moved in a decade. If autonomous technology genuinely gets the edge cases figured out — and I think it eventually will — the safety upside is enormous. I'm not anti-autonomous vehicle. I'm anti-deploying-half-finished-systems-on-public-roads-while-pretending-it's-settled.

The technology is moving faster than our ability to manage it. That's not a new problem, but it's one we really can't afford to get wrong at highway speed.