Can AI Drive Like Humans? Spoiler: Not Quite Yet.

Ok wait because this is actually insane. Imagine AI models that could drive like humans without the endless tweaks. Sounds like a dream, right? But here's the tea. Two big-name language models, OpenAI's o3 and Google's Gemini 2.5 Pro, were tested as wannabe human drivers. Spoiler: they're not quite ready to hit the roads.

The Experiment

So, researchers took these AI models and threw them into a simple one-dimensional driving challenge. The idea was to see if they could mimic human driving behavior. They had the models take on a merging scenario. No big deal for us humans, but for AI? A big mood.

Both models managed to pull off some human-like moves. They could intermittently control operations and respond to spatial cues just like a real driver. But, and it's a big but, they flopped on reacting to dynamic velocity changes. No cap, they just couldn't keep up.

Safety First? Maybe Not.

Hold up, because safety is where it gets sketchy. The performances of the two models were all over the place. Diverged sharply is putting it mildly. This is like expecting a golden retriever to suddenly meow. Just not happening.

What's really wild is a deeper dive showed that how you prompt these models acts like a bias. It's model-specific and doesn't transfer. Basically, each AI has its own vibe, and changing the prompts is like speaking a whole new language.

Why You Should Care

No but seriously. If you're thinking AI is going to take over driving anytime soon, slow your roll. These findings are a wake-up call. Sure, LLMs have potential. They could become the backbone of AV evaluation pipelines, but right now? They're still toddlers learning to walk.

So, what are we walking away with here, bestie? It's not time to hand over your keys to AI just yet. These models need a glow-up before they're ready to fully mimic human behavior behind the wheel. And until then, maybe keep your eyes on the road and hands on the wheel.