Large language model technology cannot drive a car, because it will sometimes crash the car.

Judge as of February 2023. True: as of then, no LLM-based system had been shown to drive a car reliably (e.g. no deployed or validated LLM-driven vehicle), and authoritative sources point to unreliability or occasional catastrophic errors as a barrier. Mostly true: LLMs could not drive cars then, but the stated reason (it would sometimes crash) is only one of several main reasons, or there were small research demos that do not amount to driving a car. Mixed: LLM-based systems were showing meaningful driving ability in research, so 'can't' is only partly right. Mostly false: LLM-type systems were already doing safe driving in controlled settings, and the claim exaggerates a real reliability concern. False: as of February 2023, LLM-type technology was demonstrably driving cars at or above the reliability of deployed systems.

Set before the research began.

1. E1 Evaluation of Large Language Models for Decision Making in Autonomous Driving arXiv (preprint)

   “Recently, there has been a number of studies that use LLMs to perform autonomous driving [10, 7, 12, 4, 9].”

2. E2 Evaluation of Large Language Models for Decision Making in Autonomous Driving arXiv (preprint)

   “Furthermore, to conduct a Proof of Concept (POC) for the feasibility of implementing these abilities in actual vehicles, we developed a system that uses LLMs to drive a vehicle.”

3. E3 Evaluation of Large Language Models for Decision Making in Autonomous Driving arXiv (preprint)

   “Due to the safety concerns associated with making lane changes on public roads, as in simulations, this experiment was conducted in a private area.”

4. E4 Evaluation of Large Language Models for Decision Making in Autonomous Driving arXiv (preprint)

   “Previous research has not thoroughly investigated the accuracy of LLMs in spatial-aware decision making through quantitative experiments.”

The only evidence is an undated preprint. It says a number of studies used LLMs for autonomous driving and describes its own proof-of-concept vehicle test held in a private area [E1][E2][E3]. It does not show what LLM-based driving looked like as of February 2023, or whether unreliability was a recognised barrier at that time, so the criteria cannot be applied.