NASA just let an LLM plan a Mars rover's route — and that should terrify every space engineer who's ever written a mission-critical function.
The Summary
- NASA's JPL used Anthropic's Claude to plan two Mars drives for Perseverance in December 2025, with humans checking the route before upload
- Space engineers are testing whether non-deterministic AI can give spacecraft the flexibility to interpret surroundings and plan tasks autonomously
- The shift is driven by necessity: missions are becoming more complex, distant, and numerous than human-in-the-loop control can handle
The Signal
Space engineering has a determinism problem. For 60 years, the unwritten rule was simple: if you can't predict exactly what your spacecraft will do in every scenario, you don't launch it. This wasn't paranoia. It was survival math. When your rover is 140 million miles away and a software bug could end a billion-dollar mission, you write code that does exactly what you tell it to do, every single time.
Generative AI breaks that model completely. Feed Claude the same Mars terrain data twice, and it might suggest two different routes. Same input, different output. Robert Ambrose, NASA's former chief of Software, Robotics and Simulation, puts it bluntly: engineers hate non-deterministic systems because "how you got into a certain situation changes the behavior."
"If you come into the same situation but from different paths, the outcome could be different."
But here's what changed: the old model doesn't scale. NASA and private companies are planning missions to the outer solar system where light-delay makes real-time control impossible. A command to a spacecraft near Neptune takes 4 hours one way. You can't drive a rover through that lag. You also can't hire enough mission planners to manage dozens of satellites, rovers, and probes simultaneously. The economics and physics both point the same direction — toward machines that think for themselves.
The experiments happening now are careful. When JPL used Claude for Mars route planning, humans still checked every decision before upload. When NASA and IBM put a compressed AI model on the ISS in May to identify floods and clouds from orbit, it was in observation mode, not control. Astronauts tested an LLM for maintenance questions in July — essentially a better manual, not a decision-maker.
But the progression is obvious:
- Observation: AI identifies what it sees
- Planning: AI suggests what to do next
- Execution: AI does it without asking
We're between steps one and two right now. The jump to three is where it gets interesting, and where every space engineer's training starts screaming. Because once you hand execution authority to a non-deterministic system, you've accepted that you can't fully predict what happens next. You've traded control for capability.
The Implication
The question isn't whether spacecraft will run on generative AI. Light-delay and mission complexity make that inevitable. The question is how space engineers will verify systems they can't fully predict. Expect a wave of research into AI alignment and formal verification methods for neural networks — not from AI safety philosophers, but from aerospace engineers who need to sleep at night.
Watch for the first mission where an AI makes a consequential decision without human approval. That's the real threshold. When it happens, it won't be announced with fanfare. It'll be buried in a mission log somewhere, the day a rover chose its own path and nobody stopped it.