Mission control should treat the AI’s revised recommendation as conditional evidence, not a definitive answer. The fall from 12% to 0.8% is reliable only if it remains stable after correcting or excluding the station with the 2.4-second timing discrepancy.
Reliability should be tested through a rapid, independent challenge process:
- Recalculate the encounter with the suspect data included, corrected, and excluded.
- Compare the AI result with an independent orbit-determination method and analyst propagation.
- Examine the full uncertainty region and possible miss distances—not just the headline probability.
- Test whether small changes in timing, observations, or model assumptions cause the recommendation to reverse again.
- Confirm that the AI is operating within conditions for which it has been validated and check its past calibration on comparable encounters.
If human experts and the AI still disagree, mission control should not resolve it through authority, seniority, voting, or presumed machine objectivity. The disagreement should be converted into explicit, testable differences: Are they trusting different measurements, estimating uncertainty differently, or applying different risk thresholds? Running both conclusions from the same controlled inputs will reveal whether the conflict comes from calculations, assumptions, or acceptable-risk judgments.
A fixed deadline should then be used. If validated data consistently support the lower risk and the result is not highly sensitive to the questionable observation, cancellation is reasonable. If the estimate remains unstable or the collision risk cannot be bounded below the operational threshold, controllers should choose the option that performs acceptably across the plausible scenarios—most likely a smaller, independently screened maneuver that preserves fuel for another correction.
Thus, the AI should influence the decision in proportion to the demonstrated robustness of its recommendation. Unresolved uncertainty should not automatically favor either action or inaction; it should favor the least irreversible option that protects against the most serious credible outcome.
