On the evening of June 20, a Tesla Model 3 left its lane on a quiet residential street in Katy, a suburb west of Houston, and drove straight through the brick wall of a family home. Martha Avila Mantilla, 76, was standing in the front room. She was airlifted to hospital and died shortly after.
The driver, Michael Butler, told Harris County police he was using the car's self-driving feature at the time. Tesla moved quickly to dispute that framing. Its head of AI, Ashok Elluswamy, posted publicly that vehicle data showed Butler had pressed the accelerator pedal to 100 percent, reaching 73 mph in a residential area, and that the pedal was still pressed even after the car had gone through the wall. Tesla's CEO Elon Musk backed that up on the same platform, arguing the system drives slowly in neighborhoods and a high-speed crash makes no sense if FSD was in control.
That version may be correct. Pressing the accelerator hard enough does override the system. But there is a reason investigators are not simply taking Tesla's word for it.
In a 2019 Florida crash, Tesla told lawyers and investigators that crash data did not exist. An independent researcher later recovered it directly from Tesla's servers. That data showed the system had detected the pedestrian who was struck. A federal jury last year found Tesla 33 percent responsible for that crash, awarding $243 million in damages, even though the driver had admitted taking his eyes off the road. The same question that arose there arises here: did the system do something that contributed to a panicked driver response, or was this purely a pedal error by someone who had mentally stepped out of the driving task?
That second possibility is itself a problem Tesla has not fully answered. When a driver hands control to an automated system for a stretch of road, their brain naturally disengages. Reaction time slows. When something unexpected happens, the wrong pedal is a common mistake. The system being "good enough" to make a driver comfortable is not the same as the system being safe enough to trust without full attention.
The regulatory backdrop makes this worse for Tesla. NHTSA is already running an Engineering Analysis covering 3.2 million Tesla vehicles, the step just before the agency can force a recall. A separate probe covering 2.88 million vehicles was opened after 80 documented cases of the system running red lights or crossing into oncoming lanes. Tesla was given multiple deadline extensions to produce data for that second probe. The agency has also raised concerns that Tesla may be under-reporting crashes because its own internal systems cannot consistently identify when the driver-assistance feature was active at the time of impact.
Tesla dropped the "Autopilot" name for its standard driving-assistance tier in January 2026, following a California court ruling that the name misled consumers about what the technology actually does. The premium tier, called Full Self-Driving (Supervised), costs $99 a month and still requires the driver to watch the road and be ready to take over at any moment.
That last part, the supervision requirement, is the detail that matters most for anyone who owns or manages a fleet with these vehicles. The system is not autonomous. When something goes wrong, the question of who bears responsibility, the driver, the manufacturer, or both, is being fought out in courts and federal agencies right now, crash by crash. The Texas case will add to that record. The Harris County district attorney's office will decide whether to pursue charges against Butler once investigators finish their review. A lawsuit on behalf of Avila's family has already been announced.