Federal defense and intelligence agencies are struggling to deploy superintelligence reliably, with weaknesses in software infrastructure and data management undermining advanced AI systems. The challenge is shifting attention from buying commercial models to building the foundations needed for dependable intelligence operations.
According to SC Media, which cites reporting first published by DefenseScoop, agencies encounter AI systems that alternate between exceptional performance and completely unusable answers. Experts argue that the supporting software environment is frequently responsible for these failures.
Infrastructure creates problems Large language models (LLMs), AI systems that generate responses using patterns learned from data, produce probabilistic rather than guaranteed outputs. They therefore need infrastructure capable of managing context, memory and governance.
A major weakness involves poorly structured datasets. Missing metadata can make it difficult to distinguish operational intelligence, such as tactical information, from permanent reference records.
Building more reliable systems Experts recommend that agencies develop their own orchestration frameworks instead of depending entirely on proprietary platforms controlled by individual vendors.
They also identify neuro-symbolic AI as an important component. This approach combines deep learning with rule-based reasoning, helping make AI decisions more traceable and auditable in sensitive environments.