In a curious experiment, a developer let the AI Claude pilot a Cessna 172 in the X‑Plane 12 simulator. The model built a Python script on the fly, logged every maneuver, and adjusted controls in real time. Within minutes, Claude executed a clean takeoff and climbed to 4,500 ft as planned before attempting a landing sequence that was still early.

The first attempt revealed a hard‑pitch issue: Claude’s controller lacked rate damping, causing a sudden 125 kt nose‑down and a crash that reset the aircraft to runway 09. After tweaking the loop with slew‑rate limits and asymmetric altitude targets, a second launch achieved stable cruise at 82 kt and a smooth descent toward the final landing attempt.

During the third run, Claude switched to a pure proportional pitch‑attitude controller, eliminating the integral term that had previously over‑corrected. The aircraft climbed to about 1500 MSL, maintained level cruise, and executed three 90° turns of a left circuit. The pattern held, but a 20‑second idle in the script caused a second crash during the final landing.

Claude’s trials expose the limits of current LLM‑powered flight control: while the model can generate working code and adapt to sensor delays, it still lacks robust state estimation and failsafe logic. The experiment demonstrates both the promise and the pitfalls of relying on generative AI for real‑world piloting tasks in the future of automation systems.