Why Autonomous Vehicles Keep Failing the Human Test

In 2022, a driverless Cruise car blocked a San Francisco fire engine for 25 seconds. The vehicle had stopped next to a bin lorry that was double-parked, technically the car was "behaving as designed." A human driver would have assessed the situation and reversed a few metres to clear the path.

The fire engine couldn't get through. City officials later noted the delay "slowed SFFD response to a fire that resulted in property damage and personal injuries."

The AV had done the right thing in a narrow, rules-based sense. It just couldn't “read the room”.

Recent evidence from AV implementations have been on relatively ordered car-orientated street scapes in North America; when AVs are introduced into older city environments (such as the upcoming trial in London) AVs ability to understand their human environment will only get more difficult.

The product is the system, not the vehicle

I've spent years working in transport innovation, and the most glaring mistake AV programs make is treating the vehicle as the product. Other transport modes have learned over decades that the product is the whole operating system: rules, signaling, interfaces, stewarded data, and the human layer that makes everything legible when reality departs from the plan.

When you bolt clever vehicles onto streets without that surrounding system, you get technically correct behaviour that still feels socially wrong.

In traditional transport modes, we work hard to stop users having to interpret the system. We standardise what things are called, how information is presented, and how decisions get made during disruptions or edge cases.

That same "invisible infrastructure" work is what AV deployments often underinvest in, even though it's where trust is won or lost.

The missing language of the street

In that split second of eye contact between a driver and a pedestrian, something quite technical happens without anyone thinking of it that way. Two people establish joint attention, then make a tiny agreement about priority.

Eye contact says "I see you." A nod, a raised palm, or a hand wave says "I'm giving way" or "you go." Both parties watch for compliance and adjust in real time if the other person hesitates.

What matters isn't the gesture itself. It's the shared understanding it creates, plus the shared accountability. If I wave you through and you move, we've both made a commitment, and we both feel responsible for keeping it safe.

When you replace a person with an AV, that shared layer collapses.

The accountability fragments across the AV system's design choices, the remote operations model, the fleet operator's policies, and the road authority that permitted the operational design domain. None of that is visible at kerb level. The pedestrian has nothing to "lock onto" the way they would with a human driver's face.

The one opportunity we have

A hundred years ago we transferred priority from people on foot to cars. We've spent decades trying to rebalance roads back to provide equity to people not in vehicles.

We only have one opportunity to introduce AVs as equal road users that don't push pedestrians and cyclists back to the periphery.

The first wave will set three path dependencies that are hard to unwind:

Infrastructure. Once kerb geometry, crossing logic, lane rules, and signal timings are tweaked to make AV operations stable, those choices become the new baseline. Cities rarely roll back road layouts for the sake of people on foot unless there's a political campaign behind it.

Liability and precedent. If early deployments normalise the idea that pedestrians and cyclists must behave more predictably to accommodate AVs, that expectation will creep into guidance, enforcement, and ultimately into how blame is allocated after incidents. This is a live debate as many organisations push responsibility onto pedestrians and cyclists suggesting that they be equipped with beacons before being permitted into the road traffic mix.

Behavioural learning. People adapt fast, then settle. If the learned lesson of AV streets is "give them space" or "don't assume they'll negotiate," you get avoidance routes, longer waits, and more defensive cycling patterns. Those habits show up as demand patterns that planners interpret as preference, and that gets baked into investment choices.

What cities should demand before rollout

If I were advising a city about how to permit AV operations, I'd start with governance and official responsibility. Appoint a named city/regional/national-level AV controller with real authority, then publish a single operating rulebook that every vendor must implement before they run a vehicle.

That first slice of invisible infrastructure is a common, testable "street interaction spec," written in plain terms and enforced through permits:

• How vehicles behave at crossings, near schools, around cyclists filtering, and when pedestrians are assertive rather than compliant

• How vehicles respond to emergency vehicles and to human instruction on street, including when the safe move is to reverse or clear a pinch point

• What the vehicle is allowed to signal externally (and how), so a pedestrian doesn't need to learn five proprietary languages

In the rail sector, the operational landscape is instructive. We don't rely on eye contact at platform edges. We externalise accountability into visible artefacts: signals, rules, trained staff, and consistent information. That makes "who is responsible" feel clear, even when you can't see the person making the decision.

For AVs, the equivalent isn't a screen that says "I am yielding." It's a whole operating pattern that pedestrians can learn without trying.

Cities will end up codifying it, because leaving it to each vendor will keep the accountability feeling vague and reinforce the 20th century concept that the most vulnerable users give way.