Tesla, Inc. stands at a critical juncture in the evolution of vehicular automation, where the promise of unsupervised operation collides with the hard realities of sensor physics, software reliability, and regulatory fragmentation. This analysis examines the company's Full Self-Driving (FSD) system through the lens of systemic risk management, drawing on a broad set of claims from user reports, expert critiques, and regulatory actions. The evidence reveals persistent phantom braking and version regressions that erode the driver experience, a sensor strategy that may lack the redundancy necessary for higher levels of automation, and a legal landscape that is far from settled. While Tesla's vision-only approach has shown capability under some conditions 5, the recurring failures in edge cases—glare, heavy rain, shadows—raise fundamental questions about the system's readiness for unsupervised deployment 2,4,31. As with the railroad safety crises of the 19th century, the industry must now balance innovation with rigorous, performance-based certification to prevent the equivalent of a catastrophic derailment. This report synthesizes the current state of affairs and identifies the material risks that investors and policymakers should monitor.
The State of Full Self-Driving: Regressions and the Phantom Braking Problem
The user experience with Tesla's FSD remains profoundly inconsistent, with performance varying not only by software version but also by environmental conditions. After a significant code merge that unified highway and city driving stacks, many users reported a degradation of highway smoothness 34,35. One noted that the prior version 12.7 had felt more refined on urban streets 35. City driving now exhibits hesitation, slow lane changes, and a conservative decision-making character that some describe as "twitchy" at speed 14,35. This suggests that the end-to-end AI architecture, while theoretically elegant, introduces behavioral instabilities that are difficult to predict and validate.
Perhaps the most persistent and safety-critical complaint is phantom braking—unexpected, often harsh deceleration in the absence of a genuine obstacle. Reports cite triggering by shadows 34,35, even in clear weather 36, and describe highway performance as "really bad" in this regard 35. Such events are not merely a nuisance; they undermine the fundamental trust required for driver monitoring and adoption. The irony is stark: a system meant to enhance safety can create a new class of hazards through false-positive activations. As any safety engineer knows, an unreliable safety device is often worse than none at all. Yet, not all experiences are negative—some users report "perfectly smooth" highway operation 35 and daily commutes with only parking maneuvers proving subpar 35. This variance points to a system whose performance envelope is neither fully characterized nor consistently bounded—a situation that recalls the early days of signalization when a signal's meaning could depend on the local weather.
The Sensor Suite Debate: Vision-Only Versus Redundancy
At the heart of Tesla's approach is a strategic bet: that cameras alone, coupled with advanced neural networks, can achieve the perception reliability needed for unsupervised driving. The company has championed this vision-only philosophy 5, but the engineering community remains skeptical. The principle of defense-in-depth, a cornerstone of safety-critical system design, argues for sensor diversity. Cameras are essential for reading signs and lane markings 2, but they are optically vulnerable—to glare, heavy rain, and fog 2,4. Radar, in contrast, excels in fog and provides direct velocity measurements 2, while lidar offers precise 3D mapping that can resolve ambiguous reflective surfaces where stereo vision fails 1,2,25, though it too has limitations over water at shallow angles 27. The argument is not that any single sensor is perfect, but that their failure modes are largely uncorrelated. A recent legislative push in New Jersey (S1677) to mandate cameras plus at least two additional sensor types for driverless commercial vehicles 2 underscores a growing regulatory conviction that redundancy is non-negotiable at higher levels of automation.
The hardware generation further constrains the path forward. It is broadly claimed that the Hardware 3 (HW3) platform will not support unsupervised FSD 29,33, while HW4 might possess the necessary compute headroom 13, though chip-architecture differences complicate the transfer of learned behaviors 6. This creates a potential liability: a large installed base of vehicles that were sold with the promise of future full autonomy may require costly retrofits or be deemed obsolete. If the camera-only suite is sufficient for Level 2 supervised driving but inadequate for the redundancy required at Level 4 16, then the business model of incremental software upgrades may hit a hard physical ceiling. The proof, as always, will be in the performance at the edge cases—not in marketing promises.
Regulatory Fragmentation: A Patchwork of Standards
The regulatory environment for autonomous vehicles resembles the pre-standardization era of railroad gauges: a fractured landscape where market access depends on jurisdiction. The NHTSA's Standing General Order requires manufacturers to report crashes when Advanced Driver Assistance Systems (ADAS) were engaged within 30 seconds of impact and the crash meets severity criteria 3. The agency has also sought extensive documentation regarding Tesla's decision to remove radar 30 and has investigated reduced-visibility scenarios 30, though preliminary findings suggested that reported conditions did not alter lateral positioning or cause significant loss of following distance 7,21. At the state level, California currently does not approve fully unattended driverless operations 17, while New Jersey's proposed bill would effectively ban camera-only autonomous vehicles without human controls 2.
The situation in Canada is particularly convoluted: Quebec enacted a ban on Level 3 autonomous vehicles in 2024 but included an exception for vehicles "whose sale is allowed in Canada" 32—a loophole that is effectively meaningless because no federal authorization process exists 32. Ontario permits consumer Level 3 32, while British Columbia has banned them entirely 32. In contrast, China is moving decisively with binding Level 3 and Level 4 driving standards 15, and a new GB/T standard for Level 2 driver assistance will take effect in January 2027 28. This fragmentation means that any rollout of unsupervised FSD will encounter jurisdiction-specific delays. Tesla may be forced to adapt its hardware, especially where sensor redundancy is mandated. The lesson from history is clear: without common standards, interoperability fails and public trust erodes. The industry needs a federal certification framework as much as the 19th-century railroads needed a uniform track gauge.
Beyond Autonomy: Braking Dynamics and Cold-Weather Range
Safety does not end with perception. The vehicle's fundamental dynamic behavior—braking and energy management—remains a competitive differentiator. Tesla's one-pedal driving strategy relies heavily on regenerative braking, but unlike some competitors, the system does not blend friction braking seamlessly when regenerative capacity is reduced 22. When the battery is full 22 or cold temperatures limit regen 22,24, the deceleration feel changes markedly, and the driver must compensate with the friction brakes. This inconsistency can increase reaction times in unexpected braking situations 22. Polestar, by contrast, blends regen and friction braking invisibly 22, and Nissan's e-Step provides a transparent one-pedal mode with brake light activation 20. These are not mere convenience features; they are human-factors issues that affect safety margins.
Cold-weather range remains a critical adoption barrier. LFP battery chemistries can lose 30–40% of their range in cold conditions 19,26, though some users in Norway report satisfactory winter performance 18. Sodium-ion batteries promise better low-temperature performance 9,12, representing a potential future advantage. Meanwhile, navigation systems from Lucid offer conservative and accurate charging stop planning that inspires driver confidence 23—a feature that Tesla must match to maintain its lead in real-world usability.
Legal and Competitive Risks
Legal and reporting risks add further layers of uncertainty. A lawsuit alleges that certain Tesla vehicles cannot physically deliver the advertised FSD functionality 10, and the legal fine print characterizes FSD as only supervised 33. Yet the SAE J3016 standard clarifies that design-intended Level 4 features should not be reclassified as Level 2 merely because a test driver supervises them 27. This tension between marketing and engineering reality recalls a time when railroad companies claimed safety improvements while resisting the air brake. The IIHS estimates that only 22% of reported crashes involving automation are police-reportable 11, suggesting that comparisons of crash rates may be systematically biased. In addition, remote teleoperation interventions carry a 100% collision rate in one sample 17, exposing the limits of fallback systems. On the competitive front, the emergence of specialized perception models like LingBot-Depth 1,25 and Chinese AI chips powering advanced driver-assist 8 indicate that Tesla's technological lead is under pressure from multiple directions.
Key Takeaways and Recommendations
The synthesis of these claims yields several material conclusions for stakeholders:
- FSD Inconsistency Undermines Value: The persistent variability in behavior, especially phantom braking and regression after updates, erodes the user experience necessary to justify premium pricing and mass adoption 34,35,36.
- Regulatory Headwinds Favor Sensor Redundancy: The patchwork of state and international rules, particularly the push for sensor diversity in New Jersey and the evolving Chinese standards, creates significant deployment risk. A vision-only architecture may prove unviable for unsupervised operation in key markets 2,3,16,31.
- Hardware Roadmap Faces Obsolescence: The inability of HW3 to support unsupervised FSD could lead to costly fleet retrofits, while HW4's architectural differences complicate software porting. This is a material financial and engineering risk 4,29,33.
- Braking and Battery Remain Competitive Factors: Cold-weather range loss and regenerative braking behaviors directly affect real-world usability and safety. Continuous improvement in thermal management and battery chemistry is essential to sustaining market share 22,23,26.
The path forward calls for transparent performance metrics, rigorous validation of edge-case handling, and a cooperative approach to setting safety standards. Just as the Westinghouse air brake transformed rail safety by making it impossible to ignore a failure, the autonomous vehicle industry must design for failure transparency. The time for aspirational promises is past; the imperatives are measurable reliability and certifiable redundancy. Only then can we ensure that this new mode of transportation does not repeat the preventable tragedies of the past.