Machine-Learning Declassification Review — The U.S. State Department's AI Triage of Diplomatic Cables

Executive Order 13526 requires U.S. government agencies to review classified records for automatic declassification once they turn 25 years old, and the volume of State Department diplomatic cables has grown roughly six-fold since the 1990s, making full manual review increasingly unsustainable.

From October 2022, the Declassification Division, the Bureau of Information Resource Management and the Center for Analytics set out to train a supervised machine-learning model on past human declassification decisions to triage cables into 'confidently declassify', 'confidently exempt' and 'needs human review' bands, reducing the volume requiring full manual review.

The model was trained on human declassification decisions from 1995-1997 cables. After a successful pilot, it was redesignated a full Program in 2023 and applied to the 1998 cohort of 121,536 cables, producing a confidence score and band per cable.

The model confidently decided over 72,000 cables (about 60% of the cohort) in 20-30 minutes of computing time each, versus roughly a year of manual review previously. The State Department reports 97-99% agreement with human reviewers on the validation set; an independent account by the American Historical Association gives more granular figures of 99.29% accuracy on declassification calls versus 81.43% on withholding calls, and a development cost of about $400,000. About 39% of cables (47,218 of 121,536) still required manual human review because the model could not confidently classify them.

The Department describes the model as deliberately 'overprotective', erring toward withholding when uncertain, and cautions against generalising results to other record types without retraining; separate pilots for AI-assisted FOIA search remain in early testing, not full deployment.