Tesla autopilot

A Utah resident’s six-year experience with Tesla’s “Full Self-Driving” (FSD) capability offers a detailed glance at the evolution – and current limitations – of the technology, as the company expands its rollout and faces ongoing scrutiny. Arthur Finch, of Salt Lake City, took delivery of a Long Range, Dual Motor Model 3 in October 2019, having paid $6,000 for what was then called Full Self-Driving.

Finch, who has since driven over 169,474 miles in the vehicle, including multiple cross-country trips, details how the system has improved over time. Initially, FSD offered automated highway driving and basic traffic-aware cruise control. It could steer on roads with clear lane markings but struggled with more complex scenarios. “Back in 2019, FSD would automatically navigate my car on interstate highways from on-ramp to off-ramp,” Finch wrote in a personal account of his experience. “It would too give me smart cruise…that would unhurried down for slower moving cars and stop my car at a stop sign or stoplight if a car was in front of me in my lane.”

Today’s iteration, “FSD Supervised,” is capable of navigating city streets and handling more challenging maneuvers like roundabouts and unprotected left turns. Finch notes the system can often complete entire trips with minimal intervention. However, he also points out persistent shortcomings. The current version, v12.6.4, struggles with tasks like exiting his garage, finding parking spaces, and consistently reacting appropriately to school zone flashing lights or sudden dips in the road. He frequently finds himself overriding the system, even when not strictly necessary, due to preference for a different driving style.

Tesla has been criticized for the naming of its driver-assistance features, particularly “Full Self-Driving,” which implies a level of autonomy the system does not yet achieve. The company has responded by renaming it “Full Self-Driving Supervised.” The debate over terminology mirrors broader discussions surrounding artificial intelligence, with systems like ChatGPT gaining prominence. Finch observes that FSD Supervised is one of the few AI systems he interacts with regularly.

A key development is the introduction of Hardware 4 (HW4) in mid-2024, featuring a more powerful computer and higher-resolution cameras. The latest FSD version, v14, is exclusive to HW4 vehicles and has received positive reviews for requiring fewer interventions and handling the beginning and end stages of trips, including parking and unparking. Tesla is currently piloting a “Robotaxi” service in Austin, Texas, utilizing FSD v14 with a supervisor present but no driver actively controlling the vehicle.

Tesla leverages data collected from its entire fleet to train and improve FSD. Every time a driver intervenes, the system prompts for feedback, which is then used to refine the AI. Finch recently test-drove a new Model 3 with FSD v14.2 from a local Tesla Service Center. He reported improvements in highway merging and navigating intersections, and the car successfully parked itself upon his return. According to Edmunds, as of February 14, 2026, over 1,071 used Tesla Model 3s are for sale in Salt Lake City, Utah, with prices ranging up to $5,833 off MSRP. Ken Garff Nissan Salt Lake City currently lists a 2025 Model 3 Long Range for $37,883. (Cars.com reports 67 Model 3s available in the area, while TrueCar lists over 382).

Finch’s brother, who recently purchased a 2024 Model Y with HW4, has also experienced significant improvements with FSD v14. Despite the advancements, Finch emphasizes that achieving true driverless capability requires a level of perfection exceeding that of human drivers. He questions whether Tesla is intentionally withholding v14 features from older vehicles to incentivize upgrades.

To illustrate the challenges of AI, Finch shared an example of using ChatGPT to colorize a historical black-and-white photograph. While the AI produced a visually striking result, it introduced inaccuracies – misrepresenting water ski equipment and altering facial features. He suggests this highlights the difficulty of training AI systems to understand nuanced details and maintain accuracy, even with vast datasets. “AI is not Always Incorrect. A better substitution for the acronym would be Often Incorrect,” Finch wrote.