From Mars rovers that pick their own laser targets to satellites that filter their own imagery, aerospace has run AI for a decade in every place a human in the loop is physically impossible — and deliberately kept it out of the cockpit everywhere else.
When NASA's Curiosity rover spots an interesting rock on Mars, it doesn't wait for instructions from Pasadena. Since May 2016, a system called AEGIS (Autonomous Exploration for Gathering Increased Science) has been picking its own laser targets — scanning navigation-camera images, scoring rock outcrops against criteria scientists set weeks earlier, and firing the ChemCam spectrometer before the images even finish transmitting to Earth. The same software now runs on Perseverance, aimed at the SuperCam instrument. It's one of the oldest continuously-operating autonomous AI systems in any industry, and almost nobody outside planetary science has heard of it.
That obscurity is the pattern for AI in aerospace generally. The industry adopted autonomy decades before "AI" was a boardroom word, for a reason that has nothing to do with hype: physics. A round-trip radio signal to Mars takes 8 to 24 minutes depending on orbital position. You cannot teleoperate a rover in real time across that gap, so the rover has to decide things for itself. That same logic — decide locally because the network can't be trusted to arrive in time, or at all — turns out to explain most of where AI has actually landed in this industry, and where it very deliberately hasn't.
Deep space is the extreme case, but the underlying constraint shows up everywhere in aerospace: satellites that are only in contact with a ground station for minutes a day, aircraft engines generating more sensor data than any downlink can carry, and a Low Earth Orbit environment that is now too crowded for human analysts to track object-by-object. In each case, the choice isn't "AI vs. a human in the loop" — it's "AI vs. nothing," because there isn't enough bandwidth or time for a human to be in that loop. That's a very different adoption story than, say, customer support or marketing, where AI is competing with a human who was already doing the job in real time.
The flip side is flight control itself — the actual surfaces and systems that keep an aircraft or spacecraft from crashing. There, the industry has gone the opposite direction: slower than almost anywhere else, for reasons worth taking seriously rather than dismissing as bureaucratic drag.
Beyond AEGIS, the Ingenuity helicopter that flew on Mars from 2021 to 2024 relied on visual odometry — a vision-based navigation system that tracked terrain features frame-to-frame to estimate its own position, because GPS doesn't exist on Mars and ground control couldn't fly it by joystick across a 10+ minute delay. JPL has published its own roadmap for extending this further: autonomous driving (AutoNav, already in use since the Opportunity/Spirit era for hazard avoidance) and increasingly autonomous sample selection and caching on Perseverance's Mars Sample Return campaign. This is the most mature corner of aerospace AI, in the specific sense that it's been in continuous operational use the longest — nearly a decade for AEGIS — though it's also the application with the smallest deployed footprint, since there are only a handful of rovers and probes to run it on.
Commercial Earth-observation operators like Planet Labs (whose Dove/SuperDove constellation images most of Earth's landmass daily) and Capella Space (synthetic aperture radar, which sees through clouds and darkness) generate far more raw imagery than their downlink budgets can move to the ground unprocessed. Cloud masking, change detection, and increasingly ship or vehicle detection run as machine learning inference either onboard the satellite or immediately on ingest, so customers get a filtered, tagged product rather than raw frames. This is a mature, commercially load-bearing use of AI — it's the actual product, not a feature bolted onto one — but it's also the application area with the least dramatic failure mode: a missed ship detection or a noisy change-detection alert is an inconvenience, not a safety incident, which is part of why it moved fast.
GE Aerospace and Pratt & Whitney have run engine health-monitoring programs for years, streaming sensor telemetry — temperatures, vibration, pressure ratios — off in-service engines and using the trend data to flag parts likely to fail before they do, scheduling maintenance around predicted wear rather than fixed intervals or in-flight failure. This is advisory, not control: the model tells a maintenance crew what to inspect, it doesn't touch a throttle or actuator. That's precisely why it's one of the more mature applications in commercial aviation — it sits entirely outside the safety-critical flight path, so it never had to clear the certification bar that flight-control software does.
With roughly tens of thousands of trackable objects in orbit now and mega-constellations like Starlink pushing that number up fast, human analysts can no longer manually screen every close approach. Companies like LeoLabs (which operates its own phased-array radar network) and Slingshot Aerospace (whose Beacon platform coordinates conjunction data and maneuver notifications across operators) use automated screening and risk-scoring to triage which of the thousands of daily predicted close approaches actually warrant a human decision to maneuver. This is a genuinely new application area — it didn't really exist as a commercial product category before LEO congestion made manual screening impossible — and it's maturing fast precisely because the alternative (not screening) is a visible, catastrophic failure mode that regulators and satellite operators alike want to avoid.
| Application area | Maturity | Primary driver | Failure mode if AI is wrong |
|---|---|---|---|
| Deep-space autonomous targeting/navigation | High (9+ years in continuous use) | Light-time delay makes real-time human control impossible | Wasted sol, missed science target — recoverable |
| Satellite imagery processing | High, commercially load-bearing | Downlink bandwidth can't carry raw data volume | Bad classification/missed detection — inconvenience |
| Jet engine predictive maintenance | High, but advisory-only | Sensor data volume exceeds manual review capacity | Missed/false maintenance flag — cost, not safety, since humans verify before action |
| Space traffic conjunction screening | Medium, growing fast | Object count has outpaced manual screening capacity | Missed real conjunction — satellite loss, debris cascade risk |
| Flight control / autonomous flight decisions | Low — advisory tools only, no certified ML in control loop | — | Catastrophic; this is exactly why it's gated |
This is the part that trips people up, because it looks like aerospace is "behind" on AI when it's actually enforcing a distinction the rest of the industry has mostly ignored: advisory AI and control AI are not the same risk category, and aerospace has decades of institutional memory about what happens when software bugs reach flight-critical systems.
Commercial airborne software is certified under DO-178C, a standard built around traceable, deterministic requirements-to-code verification — you can, in principle, prove that a given line of certified flight software does what its specification says and nothing else. A trained neural network doesn't have a specification in that sense; it has weights learned from data, and proving what it will do on an input it hasn't seen is an open research problem, not a certification checklist item. That's not a temporary gap regulators haven't gotten around to closing — it's a structural mismatch between how DO-178C verifies software and how machine learning models actually work.
EASA has been the most concrete about mapping a path through this. Its AI Roadmap defines three levels: Level 1 is AI that assists a human (who remains the decision-maker and can override it), Level 2 is human-AI cooperation or collaboration, and Level 3 is autonomous AI making decisions without a human in the loop. EASA's published guidance so far — its AI Concept Paper, now in its second issue — covers Level 1 and 2 applications; Level 3, the category that would include an AI system with authority over flight control, remains a research and roadmap item rather than something with an approval pathway today. The FAA has taken a similarly cautious public posture, treating machine-learning-based systems as requiring case-by-case special conditions rather than a standard certification route.
What this means in practice: the predictive-maintenance model and the satellite image classifier above are Level 1 or at most Level 2 by that framework — a human always verifies before acting on their output. Nothing currently flying commercially is Level 3 in the sense of an ML model with unsupervised authority over a flight-critical decision. The rover autonomy described above is a partial exception, and it's instructive why: Mars rovers are expendable, uncrewed, and operate in an environment where the "worst case" of a bad autonomous decision is a wasted day, not a loss of life — a risk profile regulators don't apply to passenger aircraft.
Aerospace didn't skip the AI wave and it isn't lagging it — it split the problem cleanly in two long before "AI" was a product category. Where the failure mode is inconvenience (a bad engine-wear prediction a human double-checks, a satellite image misclassification) or where there's literally no alternative because physics rules out a human in the loop (a rover 15 light-minutes away, a debris field growing faster than analysts can screen it), machine learning has been deployed for years and in some cases over a decade. Where the failure mode is catastrophic and a certification regime exists specifically to prevent it, AI remains advisory by design, and the real news to watch isn't a new model release — it's whether EASA or the FAA actually publish an approved pathway for Level 3 autonomy, because that's the signal that the certification wall, not the technology, has started to move.