Waymo's Parking Tickets Expose the Gap in AI Judgment

Autonomous vehicles from Waymo have collected thousands of dollars in parking fines in Austin, according to the Wall Street Journal, even though the same cars crash far less often than human drivers over tens of millions of miles.

That contradiction sits at the center of a new argument from Madhu Shashanka, co-founder, chief scientist and chief technology officer at Concentric AI, a data security software company. In an opinion piece published by the Guardian on October 20, 2026, Shashanka writes that AI is increasingly getting good at doing some tasks, but it can never replace human workers' judgment.

The story matters because it lands in the middle of the automation debate. Enterprises are deploying AI systems across coding, customer service, security and analysis, and boards want to know which roles shrink and which survive. Shashanka's answer, drawn from running a security company, is that judgment — not capability — is the dividing line.

What do the Waymo fines actually show?

The Wall Street Journal reported that Waymo vehicles in Austin keep picking up easily avoidable parking tickets. The specific failures the Journal documented are mundane: the cars park in disabled parking spots, in tow-away zones, and in a church garage during Sunday service.

Set against the company's safety record, those failures look strange. As Shashanka notes, Waymos crash far less often than human drivers over tens of millions of miles. The cars handle the statistically hard problem — driving itself — better than people do, while failing at the trivially easy one: reading a curb sign and choosing the next spot.

"That is the state of AI today: impressively capable at something complex, yet embarrassingly clumsy at something easy," Shashanka writes.

The comparison is the useful one. A system can hold the state of the art on a benchmarked task and still lack the contextual awareness a teenager applies without thinking. Austin's parking fines are that gap, priced in municipal penalties.

Why does this matter for the labor market?

Shashanka draws a direct line from the robotaxis to the workplace. The parking failures, he argues, have implications for what work will remain human and what will be automated by AI.

The logic runs as follows. Automation targets tasks where success can be specified, measured and repeated — miles driven without a collision. It struggles where success depends on reading context that was never written into the specification — that this garage fills with parishioners at 10 a.m. on a Sunday, that this curb space belongs to someone with a disability placard.

Workplaces are full of the second kind of problem. A security analyst decides whether an anomaly is an attack or a contractor working odd hours. A manager decides whether a flagged transaction reflects fraud or an unusual but legitimate customer. Those calls require judgment about situations the rules did not anticipate, which is precisely where Shashanka says AI breaks down.

His position carries weight because of where he sits. Concentric AI sells data security software, a field where vendors market autonomy aggressively — systems that classify data, detect risk and remediate exposure with minimal human input. A chief technology officer at such a company arguing that AI cannot replace human judgment is a caution from inside the automation trade, not from outside it.

What is the honest division of labor?

The Waymo case suggests a split that is already visible in deployed systems. AI takes the high-volume, well-specified work: driving millions of miles, scanning logs, drafting text. Humans keep the judgment calls that depend on context, accountability and an understanding of what the rules meant to protect.

This is not a temporary bug that another model generation will fix, on Shashanka's reading. The Austin fines are not failures of sensors or compute; the cars can see the signs. They are failures of interpretation — knowing which details matter right now — and that is the capability Shashanka places permanently on the human side of the ledger.

For companies writing AI policy in 2026, the practical takeaway is to audit for the parking-ticket class of failure: the embarrassing, obvious, low-stakes-per-instance error that no benchmark predicts. Those errors, multiplied across a workforce or a customer base, define where a human still has to stay in the loop.

The full argument appears in the Guardian. Its core claim, that capability and judgment are different properties and only one of them automates cleanly, will keep testing every deployment that ships this year.