What space missions have actually handed to AI in flight and on the ground, and why crew-critical judgment remains a different assurance problem.
Mission operations already runs AI in real flight use, but only where the job is bounded: spotting telemetry patterns, ranking science targets, fitting approved activities into constraints, executing preauthorized plans. The public record shows routine operational use; it does not show hazardous or crew-critical judgment migrating to opaque AI, because human-rated systems are certified around fault tolerance, recovery, and the ability of crew or ground to override. The AI that earns a place in the loop is not a copilot with broad permission. It is a bounded operator with a named job, explicit constraints, and a rehearsed way back out.
NASA's human-rating guidance ties crewed operations to fault tolerance, recovery, situational awareness, configuration control, and the ability of crew or ground to control or override critical functions. A strong model score never substitutes for that demonstrated safety case.
JPL's AEGIS has selected rover science targets against scientist-set parameters since 2010. Perseverance's onboard planner became the primary way the rover is operated in October 2023, and by January 2025 it had executed more than 7,800 requested activities across 429 sols.
A Mars Science Laboratory telecom anomaly-detection system ran in daily operations and reported a 90% cut in team workload, with human review and hard safety thresholds still in the path. Space-station consoles got data-driven monitoring alerts the same way. The model flags; a person or a deterministic rule decides.
ESA's TECO scheduler has run as a reliable service since 2013, applying hundreds of constraints to weekly schedules of more than 100 actions. Humans step in when an anomalous conflict cannot be resolved; the tool supports operational authority rather than replacing it.
Hubble moved routine operations from 24x7 staffing to 8x5 in 2011 only after an eight-month shadow period, anomaly simulations, and changes across planning, ground systems, and responsibilities. Perseverance's planner shipped through a requirements-traced verification campaign. A successful demonstration is not yet an operating capability.
For every proposed AI function, specify the input, the decision, permitted and prohibited actions, the escalation owner, the deterministic safeguard, and the reversal path. Separate detection, recommendation, scheduling, and command authority.
Telemetry prioritization, science-data triage, contact and activity scheduling, and nominal health monitoring match the strongest deployed evidence.
Never accept a pilot on nominal accuracy alone. Test false alarms, missing data, configuration drift, conflicting constraints, degraded communications, operator override, and reconstruction after an event.
For functions that can cause, or fail to prevent, a catastrophic hazard, start from the applicable human-rating and safety requirements. A generic AI governance review is not a substitute.
The open question is whether today's crew-critical boundary is a durable design principle or a solvable assurance-engineering problem. The tell will be the first human-rated program that certifies an adaptive AI function for a crew-critical envelope without every model change reopening certification. Nothing in the operational record shows that yet, and whoever solves it changes the economics of flight autonomy.
For each proposed AI function, write down the input, the decision, the permitted action, the prohibited action, the escalation owner, the deterministic safeguard, and the reversal path. Separate detection, recommendation, scheduling, and command authority; most autonomy arguments dissolve once those four are separated.
Telemetry prioritization, science-data triage, contact and activity scheduling, and nominal health monitoring match the strongest deployed evidence: AEGIS target selection, the Perseverance onboard planner, TECO scheduling, and console anomaly alerts. Leave novel, coupled, or hazardous conditions on the human side of the map.
Score the function on false alarms, missing data, configuration drift, conflicting constraints, degraded communications, operator override, and reconstruction after an event. Nominal accuracy alone proves a demo, not an operating capability.
Require requirements traceability, versioned data and model inputs, operations-like simulation, shadow operation, named shift procedures, and a rollback plan before changing authority or staffing. Hubble ran eight months of 24x7 shadow operations and anomaly simulations before cutting to 8x5.
For any function that can cause, or fail to prevent, a catastrophic hazard or an abort, start from the applicable human-rating and safety requirements: fault tolerance, recovery, situational awareness, configuration control, crew override. A generic AI governance review is not a substitute, and a model score is not a safety case.
A named operating owner for each AI function's authority envelope, accountable for the escalation path, the deterministic safeguards, the shadow-operation record, and the decision to suspend or roll back.
A routine-use / demonstration / claim decision brief per capability, so a one-mission demo or a contractor-reported metric never buys the authority that only routine, verified operational use has earned.
The trail from telemetry to alert to decision to authorized action to verified outcome, plus off-nominal test results and shadow-operation records, rather than model accuracy scores.
Start with owner, briefing, and proof for one system and one play. If the gap is material, widen to a readiness look across the operation's monitoring, scheduling, and command loops, and build the operating machinery only when the operator wants it run.
Three boundaries hold after verification. Telemetry-model accuracy does not establish autonomous diagnosis or recovery authority: the KETTY prototype produced too many false alarms on real telemetry, and the strongest recent metrics are contractor-reported rather than independent. Workload, staffing, and savings figures, including the 90% number, are agency program-reported results, not generalizable ROI. And the most detailed human-rating language comes from a superseded document; formal compliance work starts from the current authorities.