AI cameras on buses helped issue 375,000 tickets — but where is the impact?

Imagine your AI is excellent at spotting users' mistakes, and you get paid for every mistake it finds. Sales grow, and the reports keep the client happy. Yet users keep making mistakes, and the original problem remains. Cameras on American school buses offer a tangible example of this product trap: detecting dangerous behavior has been automated, but demonstrating an improvement in safety has proved harder.

In Montgomery County, Maryland, the program has generated more than 375,000 tickets totaling $92 million. That is the value of fines issued; it should not be treated as money collected or vendor revenue. According to an investigation by Bloomberg Businessweek and Type Investigations, published back in April 2026, violations declined only slightly, and reporters found no evidence of fewer collisions near stopped buses.

How a camera turns a passing vehicle into a police case

The task is straightforward: a bus stops, activates its red lights and extends its stop arm, but a vehicle drives past. Cameras capture the incident and the vehicle's license plate. That sets off a process in which recognition is just one step.

Allentown School District describes the following workflow: video, along with location coordinates and a timestamp, is stored on an onboard recorder, then uploaded to the cloud over a secure connection. AI filters the footage, a specialist reviews the suspected violation and compiles the evidence. Police examine the package and decide whether to send the vehicle's owner a notice.

For example, footage shows a car passing through the area beside a bus. A recognized license plate alone is not enough: the reviewer needs the sequence of events, the status of the signals, the car's position and the road conditions. Product quality therefore depends on the model, the completeness of the recording and how much time a person can spend reviewing it. Automating the screening process does not eliminate the need to investigate disputed cases.

When the business grows along with the problem

Vendor BusPatrol explains its financial model in simple terms: the company covers equipment installation and operation with proceeds from fines. The client gets a system without having to buy cameras upfront. For a municipality with a limited budget, that is an attractive offer.

But this arrangement creates conflicting incentives. If dangerous passing incidents disappear, the source of payment shrinks. That does not prove violations are deliberately overstated, but it does require a separate assessment of the outcome for residents. Rising receipts alone say nothing about how much safer it has become for children to board buses.

It helps to distinguish three metrics: how many incidents the system detected, how many violations reviewers confirmed and how road risk changed. The first describes the volume of work, the second helps assess screening quality, and the third addresses the promise made to the client. It is very easy to substitute the first two for the last, especially when they are what generates the cash flow.

For developers, this creates a concrete task: preserve the grounds for each decision and the ability to reconstruct the entire incident. For product managers, it means agreeing in advance on what will count as success a year later. Examples include fewer dangerous passing incidents for a comparable number of bus stops, as well as the cost of review and the share of decisions overturned. These are proposed evaluation criteria, not published results from the program.

Why 19 Waymo incidents prove nothing without a denominator

The original post compares tickets issued to ordinary drivers with 19 Waymo incidents. NPR reported that this was the number of incidents documented by Austin's school district. The federal investigation began after a report of a separate incident; attributing its launch specifically to the figure of 19 is inaccurate.

The comparison grabs attention, but drawing a safety conclusion requires matching time periods, conditions and numbers of encounters with stopped buses. Hundreds of thousands of tickets issued to human drivers do not excuse a self-driving vehicle's mistake. Nor does the lack of a demonstrated reduction in crashes mean cameras are useless: their impact needs to be evaluated appropriately.

The practical takeaway for an AI product team is straightforward: before scaling up, check whether your revenue is tied to solving the client's problem. Otherwise, you may build a successful service whose continued growth depends on that problem persisting.