Whether you're looking for an angel investor, a growth advisor, or just want to connect — I'm always open to great ideas.
Get in TouchWaymo's autonomous vehicles have illegally passed stopped school buses dozens of times across multiple cities, triggering federal investigations, a software recall affecting 3,000+ vehicles, and growing calls to halt operations near schools.
AI, startups & growth insights. No spam.
TL;DR: There is one traffic law so universally understood that it appears in the driver's education handbook of every U.S. state without variation: stop for a school bus with its red lights flashing. No exceptions. No ambiguity. No edge cases. Children are crossing the road. You stop. Yet over the past several months, Waymo's autonomous robotaxis have failed this test — not once, not twice, but more than two dozen documented times across multiple American cities. A software recall affecting over 3,000 vehicles hasn't fixed it. A data-collection event organized by a school district to help train Waymo's AI didn't fix it. A human remote assistant reportedly made it worse. And the company refused a school district's request to halt operations near schools during bus hours. This is the story of how the most symbolically important safety test in autonomous driving became Waymo's most damaging crisis to date.
The footage is not subtle. School buses parked at the curb, red lights flashing, stop arms fully extended, crossing control arms deployed at the front of the bus. In some videos, children are visible on the sidewalk or in the process of crossing. And in frame after frame, a white Waymo Jaguar I-PACE rolls past the bus without stopping — sometimes at near full speed, sometimes after a brief hesitation.
Austin Independent School District (AISD) began documenting these incidents during the 2025–26 school year and by January 2026 had confirmed at least 24 separate violations. That number includes incidents spread across multiple routes and times of day, and the violations occurred in at least two states — Austin, Texas and Atlanta, Georgia. Atlanta Public Schools reported six confirmed cases as of early December 2025.
Under Texas law — and the law of every other U.S. state — passing a stopped school bus with lights activated is illegal regardless of which side of the road you are on, in most circumstances. It is among the traffic violations with the most severe consequences for human drivers, carrying fines of up to $1,250 in Texas and potential criminal charges if a child is injured. For Waymo, so far, there have been no injuries reported. But that is not the frame through which safety regulators are looking at this.
What makes these incidents especially striking is that school buses are not subtle objects. They are painted a distinctive yellow-orange precisely so they stand out. They carry multiple redundant warning systems: flashing amber lights that activate before the bus stops, red lights that activate when it is fully stopped, a mechanical stop arm that extends from the driver's side, and — on newer buses — a crossing control arm that swings out from the front to prevent children from walking into traffic. Any functional perception system should have no trouble identifying this combination of signals.
The incidents first drew federal attention in October 2025, when the National Highway Traffic Safety Administration (NHTSA) opened a preliminary evaluation into Waymo's autonomous driving system. The trigger was a media report about a single incident in which a Waymo vehicle failed to remain stopped when approaching a bus with its red lights flashing and stop arm deployed. NHTSA sent Waymo a detailed questionnaire and gave the company until January 20, 2026 to respond with documentation of similar incidents and details on its remediation efforts.
Then in January 2026, the National Transportation Safety Board (NTSB) opened its own, separate investigation. The NTSB is an independent federal agency responsible for investigating transportation accidents and issuing safety recommendations — it does not have enforcement power like NHTSA, but its investigations tend to produce the most thorough technical analyses and carry significant reputational weight. Investigators traveled to Austin specifically to examine more than 20 incidents involving AISD buses.
The dual federal investigation — NHTSA and NTSB simultaneously scrutinizing the same pattern of behavior — is unusually aggressive for a company that has generally positioned itself as the industry's gold standard for autonomous vehicle safety. And the NTSB's March 2026 update made clear the investigation was not winding down: the board confirmed new incidents had occurred, suggesting the problem was ongoing even as regulatory scrutiny intensified.
NHTSA also expanded its investigation during this period, upgrading the scope of its probe in response to the volume of incidents. An expanded NHTSA investigation — moving from preliminary evaluation toward an engineering analysis — is typically a precursor to regulatory action and carries the possibility of mandatory recalls rather than voluntary ones.
On December 5, 2025, Waymo filed a voluntary software recall with NHTSA covering 3,067 vehicles equipped with its fifth-generation automated driving system. The company said it had identified a software issue contributing to the school bus incidents and believed subsequent updates would fix the problem. This was, by the numbers, one of the larger AV software recalls in U.S. history.
AISD, cautiously optimistic, took Waymo at its word. Then the violations kept happening.
KXAN Austin, which had been documenting the story aggressively through its investigative unit, reported that violations continued after the recall. A lawyer for the school district wrote to Waymo specifically noting that at least five incidents had occurred after Waymo had assured the district that its software was updated. One of these took place on January 19, 2026 — a robotaxi drove past a stopped bus with its stop arm fully extended while children waited to cross the street.
The recall had covered the software, but the underlying perception and decision-making failures that led to the incidents were apparently not fully addressed by the patch. Either the fix was incomplete, the software rollout didn't reach all vehicles, or there were additional failure modes the initial analysis hadn't identified. Waymo has not publicly provided a detailed post-mortem on why violations continued after the recall — a silence that has frustrated both regulators and school officials.
For observers of the AV industry, this sequence — violation pattern identified, recall issued, violations continue — raised uncomfortable questions about the thoroughness of safety validation processes. A software recall in the automotive world typically involves extensive testing before deployment to ensure the fix actually addresses the root cause. The continuation of incidents suggests that testing loop either missed something or that the problem is more architecturally complex than a single software patch can resolve.
Perhaps the most disturbing detail to emerge from the NTSB investigation is this: in at least one documented incident, the autonomous vehicle itself did the right thing — and then a human told it not to.
On January 12, 2026, a school bus stopped on Oltorf Street near Interstate 35 in South Austin to pick up students. The Waymo vehicle initially stopped behind the bus. Then, as sometimes happens, other (presumably human-driven) vehicles began passing the bus illegally. The Waymo, uncertain about the situation, did something it is programmed to do when confused: it contacted remote assistance.
Remote assistance is a feature built into Waymo's system to handle edge cases. A human operator — not inside the car, but monitoring remotely — can provide context or guidance when the autonomous system encounters a situation it cannot confidently classify. The vehicle transmitted a query: "Is this a school bus with active signals?"
A remote assistance agent located in Novi, Michigan answered: "No."
The Waymo then proceeded past the bus — which still had its stop arms fully extended and its red lights flashing — while students were loading.
The NTSB's preliminary report framed this as human error. The remote agent had misidentified the situation, and the autonomous system, trusting the human's input, acted on that incorrect classification. Waymo later characterized the incident as involving human error by its remote operator, which was consistent with a separate NTSB finding that human error played a central role.
But this explanation opens a can of worms. Waymo has long positioned remote assistance as a safety backstop — a human layer that catches what the AI misses. If that human layer is itself making critical safety errors, then the system's overall safety profile is worse than the AI-only performance metrics suggest. It also means Waymo cannot fully attribute these incidents to software bugs: at least some of them are organizational and operational failures.
The broader implication for AV safety is significant. The entire premise of "human oversight" as a safety enhancement assumes the humans doing the overseeing are correctly trained and consistently reliable. A remote operator based a thousand miles away, working through a camera feed, misidentifying a school bus with all signals active is a failure that no software patch addresses.
In mid-December 2025 — after the recall had been filed but before violations resumed — Austin ISD organized something remarkable: a half-day data collection event at a school parking lot designed to help Waymo train its perception systems on school buses.
District employees pulled school buses from across the fleet and assembled them in a parking lot. Stop arms were deployed. Lights were activated. The setup was designed to give Waymo's sensing and camera systems clean, controlled exposure to the exact visual signatures they were failing to recognize. Staff from multiple departments coordinated their schedules to make it happen. The school district was, in effect, voluntarily doing Waymo's training data work for them — out of concern for student safety.
By mid-January, the violations had resumed.
AISD officials were, understandably, incensed. The district had gone out of its way to assist a private company in fixing a problem that endangered children, and the company had apparently been unable to translate that effort into meaningful system improvement. The school district sent a letter to Waymo requesting that the company halt all operations in Austin during morning and afternoon bus hours until the issue was fully resolved.
Waymo declined.
The company maintained that it was actively working to address the problem and that a blanket operational pause was not warranted. This response — a technology company refusing a school district's safety request involving children — generated significant negative press and deepened skepticism about whether Waymo's stated commitment to safety extended to situations where compliance would cost them service revenue.
This dynamic is increasingly relevant for the AI industry broadly. As we covered in our piece on Figure AI's White House ambitions and the politics of robot deployment, the tension between rapid commercial expansion and safety due diligence is a recurring theme across the robotics and autonomous systems space — and one that regulators are watching closely.
To understand why this is happening, it helps to think about what autonomous vehicles are actually doing when they navigate urban environments. They are not "seeing" in the way humans do. They are running perception algorithms over point clouds from LiDAR sensors, frames from cameras, and radar returns — fusing all of this into a real-time 3D model of the world and then classifying objects within that model.
School buses, from a machine learning perspective, present several challenges that are easy to underestimate. Their visual presentation changes substantially based on distance, angle, lighting conditions, and occlusion. A bus viewed from the side at close range looks completely different from a bus at 200 meters at dawn with another vehicle partially in front of it. The flashing lights — which are the critical signal — can be difficult to detect when they're washed out by sunlight, obscured by rain, or simply not well-represented in training data for a particular camera angle or time of day.
Phil Koopman, a professor at Carnegie Mellon University who has written extensively on autonomous vehicle safety, published a detailed analysis of the Waymo school bus problem on his Substack. His core observation: this is not a fringe edge case. School buses operate on predictable routes during predictable hours, and failing to reliably stop for them is not a rare anomaly — it is a systematic gap in the operational design domain that Waymo should have resolved before deploying in cities where school buses operate.
Koopman's analysis also highlights the distinction between "no injuries reported" and "the system is safe." An autonomous vehicle passing a stopped school bus at speed while children are crossing is an extremely high-severity near-miss, even if no child happens to be in the vehicle's path on any given occasion. Frequency and severity must both be considered when evaluating safety — and a pattern of 24+ violations of one of the most fundamental traffic safety laws represents a serious severity issue regardless of the outcome statistics so far.
Throughout this entire episode — the initial violations, the recall, the continued violations, the NTSB investigation, the AISD training event that didn't work — Waymo has been expanding its service footprint.
The company launched commercial robotaxi service in Miami in early 2026, adding to existing operations in Phoenix, San Francisco, Los Angeles, Austin, and Atlanta. Each new city means new school districts, new bus routes, and new potential for similar incidents. Miami-Dade County has one of the largest school bus fleets in the United States.
This expansion-while-under-investigation posture raises questions that go beyond Waymo as a company. The regulatory framework for autonomous vehicles in the United States does not currently give NHTSA clear authority to halt an AV company's operations pending the outcome of a safety investigation, the way the FAA can ground aircraft. NHTSA can pressure voluntary recalls and, eventually, mandate them — but the pace of enforcement is measured in months and years, while Waymo's expansion moves in weeks.
Liability questions are equally unresolved. If a Waymo robotaxi strikes a child crossing in front of a school bus — an event that has not occurred but remains a statistical possibility given the incident count — who bears responsibility? Waymo, as the operator of the autonomous system? The remote assistance contractor who misclassified the bus? Alphabet/Google, as Waymo's parent company? The regulatory framework for answering these questions is still being written, which creates enormous legal uncertainty for everyone involved.
This regulatory vacuum is not unique to Waymo. As we noted in our coverage of David Sacks' exit from the Trump administration's AI policy role, the federal government's approach to AI and autonomous systems regulation has been characterized more by deference to industry than by proactive safety standards — a posture that may need to change as incidents like this accumulate.
The school bus incidents have surfaced several specific failure modes that Waymo and the broader AV industry need to address — and that regulators need to compel, not merely request.
On perception: School buses need to be treated as a priority safety classification, not a general object category. Waymo's fifth-generation system should have specific, hardened logic for recognizing deployed stop arms, flashing red lights, and crossing control arms under all lighting and weather conditions. If the system cannot reliably classify these signals, it should default to stopping — not to proceeding.
On remote assistance: The January 12 incident demonstrates that remote operators need far more rigorous training and accountability structures for safety-critical classifications. A remote agent who misidentifies a stopped school bus should trigger an immediate review, not just a data point in a post-incident report. Waymo should also consider whether the system's design — which allows a remote operator's "no" answer to override a vehicle that had correctly stopped — needs to be rethought for high-stakes scenarios involving children.
On recalls: A software recall should not be filed until there is high confidence the fix addresses the root cause and has been validated under realistic conditions. Filing a recall and then continuing to see violations undermines both the recall process and public trust in Waymo's safety engineering.
On operations: AISD's request — halt service near schools during bus hours until the problem is verified fixed — was not unreasonable. Refusing it in the context of an active federal investigation and ongoing violations was a significant strategic and ethical misstep. Waymo should adopt clear operational protocols that pause service in specific zones when a documented safety issue is unresolved.
On regulation: NHTSA needs clearer authority to impose operational conditions on AV companies during active safety investigations. The current framework, which relies primarily on voluntary recalls and negotiated timelines, is not adequate for a technology being deployed at scale in communities with children.
There is a reason autonomous vehicle companies spend enormous resources on edge cases like construction zones, unusual road markings, and atypical pedestrian behavior. Urban driving is full of scenarios that require careful handling. But school buses do not belong in the edge case category. They are large, distinctively colored, equipped with multiple redundant warning systems, and operate on predictable schedules in predictable locations. Failing to stop for them is not an exotic failure mode — it is a foundational safety gap.
Waymo has built a genuine track record in autonomous driving. The company's overall safety statistics, compared to human drivers, are impressive across many metrics. That record matters and should not be dismissed. But the school bus incidents reveal something important about the limits of statistical safety arguments: a system can be better than average while still having specific, serious failure modes that are unacceptable in deployment.
A software recall that doesn't stop the violations. A training event organized by a school district that didn't translate to improvement. A human remote operator who told the car the school bus wasn't a school bus. And an expansion strategy that continues to add cities while the investigation is open. Each of these, individually, might be explainable. Together, they describe an organization that has not yet fully resolved the gap between its safety marketing and its operational reality.
Children are crossing the street. The law says stop. Getting this right is not a stretch goal — it's the minimum.
Pennsylvania's Senate passes the SAFECHAT Act (SB 1090) with near-unanimous bipartisan support — establishing age-appropriate AI chatbot standards, self-harm safeguards, and $10,000 civil penalties for violations.
Washington passes HB 2225, HB 1170, and SB 5395 before adjournment, setting precedent for chatbot child safety, AI disclosure, and health insurance AI regulation.
A coalition of AI companies and Silicon Valley investors is pumping $100 million or more into the 2026 midterms, funding pro-AI candidates and targeting legislators who support safety guardrails — the biggest tech lobbying play in American electoral history.