Tesla FSD: Why "Supervised" is Just a Fancy Word for "Driver is Still on the Hook"
Another Tesla FSD crash, another fatality. This time in Katy, Texas, a 2026 Model 3 hit 73 mph, killing a 76-year-old woman. Tesla's Head of AI, Ashok Elluswamy, says the vehicle logs confirm FSD (Supervised) was active, but the driver manually overrode it by flooring the accelerator to 100%. This incident once again brings the capabilities and limitations of Tesla FSD under intense scrutiny. This isn't a "driver error" in the traditional sense; it's a fundamental design flaw in the human-machine interface. We're building systems that allow for dangerous ambiguity, and then we blame the human when the system's limitations become lethal.
The whole "Full Self-Driving (Supervised)" branding is a masterclass in marketing fluff designed to shift liability. It's Tesla's principal autonomous technology, aiming to take over all driving duties and eventually become a robotaxi. Autopilot is getting phased out, so this is it. But here's the thing: if it needs supervision, it's not self-driving. It's an advanced driver-assistance system with a dangerously misleading name.
I've seen systems with better failure modes in industrial robots from 2010. The core problem is how Tesla FSD handles driver input, especially the accelerator.
The Accelerator Trap: A Logic Error, Not a Feature
When FSD (Supervised) is engaged, you have a few ways to disengage it: hit the brake, move the gear stalk, or press the right scroll wheel. You can also yank the steering wheel. Simple enough, right?
Here's the dealbreaker: pressing the accelerator does not disable FSD. The system allows you to use the accelerator while it's engaged, say, to overtake another car. FSD stays active. This is where the Katy crash happened. The driver pressed the accelerator to 100%, and FSD was still running. The system didn't disengage; it just let the driver take over the speed control while still managing steering and other aspects. This isn't a feature; it's a critical logic error that creates a "pedal misapplication" scenario.
Many advanced driver-assistance systems from other manufacturers, when engaged, will immediately disengage or provide clear haptic/auditory feedback if the driver applies significant accelerator input that conflicts with the system's intent. This design philosophy prioritizes safety by ensuring the human driver is unequivocally in control or the system is unequivocally off. Tesla FSD's approach, however, creates a dangerous grey area where the system allows the driver to accelerate while still claiming to be active, leading to situations like the Katy crash where the driver's intent and the system's state are dangerously misaligned. This isn't just a technical oversight; it's a profound misunderstanding of human factors in automation.
The failure happens here:
This isn't a "driver error" in the sense of a distracted driver. This is a system that allows a driver to apply maximum acceleration while still claiming to be "self-driving." It's a recipe for disaster, putting the driver in a cognitive bind where they think the system is helping, but their input can lead to catastrophic outcomes without clear disengagement.
The Vision-Only Monoculture Risk
Tesla's FSD relies on cameras only. That's it. No lidar, no radar (at least not in the same way Waymo uses it). This is a fundamental architectural choice that introduces a monoculture risk. The NHTSA is already investigating a 2023 FSD crash that focused on the system's ability to "detect and respond appropriately to reduced roadway visibility conditions" like sun glare or fog. Cameras struggle with that. It's a known limitation.
The reliance on a single sensor modality for something as critical as autonomous driving is a gamble with high stakes. While cameras offer rich data, they are inherently susceptible to environmental variables beyond just glare or fog. Heavy rain, snow, dust, or even complex lighting scenarios at dawn or dusk can significantly degrade camera performance. This 'monoculture' approach means that if the primary sensor fails or is compromised, there's no robust backup. It's a significant architectural choice for Tesla FSD.
In contrast, sensor fusion, as demonstrated by Waymo, provides layers of redundancy, allowing the system to cross-reference data from different sensor types (lidar, radar, cameras) to build a more complete and resilient perception of the environment. This fundamental difference in architectural philosophy directly impacts the safety envelope of systems like Tesla FSD, especially as they push towards higher levels of autonomy and the promise of robotaxis.
Contrast this with Waymo's Ojai Robotaxis. They're deploying vehicles with Waymo’s sixth-generation self-driving system, packed with 13 cameras, four lidar sensors, six radar units, and an array of external audio receivers. That's a sensor fusion strategy built for redundancy and solid against environmental challenges. Lyft's network policy even states that autonomous vehicles using only one type of sensor won't qualify. That's a direct shot at Tesla's camera-only approach for its Cybercab and robotaxis.
The social sentiment around FSD is polarized, and this sensor difference is a big part of why. Enthusiastic users claim FSD is highly capable, "never gets distracted," and "takes safe driving to the next level." They see opposition as "politics and misinformation." But then you talk to former Tesla AI data labelers. A Reuters investigation found 7 out of 9 wouldn't ride in a Tesla on FSD. They've "seen it fail," routinely observing speeding and failures on camera. That's the chasm between perception and reality. The people who built and trained the system don't trust it. That should tell you everything.
The Path Forward: Clear Disengagement or Nothing
The U.S. Department of Transportation is proposing changes to allow vehicles designed exclusively for automated driving systems to omit brake pedals. That's for truly autonomous vehicles, not "supervised" ones where the driver is still expected to intervene. The current Tesla FSD (Supervised) model is a dangerous hybrid. It gives drivers a false sense of security while retaining the expectation of immediate human intervention, even in scenarios where the system's behavior is ambiguous.
The current regulatory framework for autonomous vehicles is still evolving, often struggling to keep pace with rapid technological advancements. Agencies like the NHTSA are caught between fostering innovation and ensuring public safety. For systems like Tesla FSD, which operate in a grey area between Level 2 and Level 3 autonomy, clear, enforceable standards for human-machine interaction and disengagement protocols are desperately needed. The ethical imperative for designers is to create systems that are not only capable but also transparent and predictable in their limitations. This means designing for failure in a way that prioritizes human safety above all else, removing any ambiguity about who is in control at any given moment. Until then, the 'supervised' label for Tesla FSD will continue to be a source of confusion and potential danger.
The fix is blunt: if a driver applies significant accelerator input, FSD must immediately and unambiguously disengage. No "remaining active" while the driver floors it. That's a logic error that needs to be patched out, not explained away as driver override. The system needs to fail safe, not fail ambiguous. Until then, "Full Self-Driving (Supervised)" remains a liability, not a feature, and it puts every driver in a no-win situation when the system hits its limits.
