As the U.S. Department of Defense pours investment into tools to detect fake content, experts warn that the real danger isn’t the doctored image, but the „poisoning” of databases. Washington’s adversaries are no longer content with creating lies; they are rewriting the source code of the reality that AI learns.
The Detection Trap: A Right Solution for the Wrong Problem
The Pentagon seems to have decided that the fight against disinformation is a simple scanning problem. Programs like DARPA’s Semantic Forensics or partnerships with Silicon Valley firms are designed to identify artificially generated videos or audio. However, this approach ignores a far more insidious threat located upstream: the compromise of the AI models that military analysts and policymakers rely on to discern truth from falsehood.
Researchers warn that hostile state actors are infiltrating propaganda directly into the training data of large language models. A recent study shows that as few as 250 malicious documents are enough to compromise a massive model. Once false information is „baked” into the AI’s internal structure, the only fix is a full retrain—an extremely expensive and often ignored process.
From Tehran to the Kremlin: How Global Perception is Shaped
Examples of manipulation are already documented. During protests in Iran, the regime’s internet blackout allowed state-controlled information flows to dominate what AI algorithms „learn” about the region. In parallel, Russian and Chinese propaganda networks have flooded public web archives, such as Common Crawl, with millions of pieces of disinformation that now appear almost verbatim in AI-generated responses.
The problem is not limited to adversaries. Even allies, by restricting media access to conflict zones or limiting satellite imagery, contribute to creating blind spots in the „memory” of algorithms. The result? When a military analyst requests a synthesis of tensions in the South China Sea or the Middle East, the answer provided may already be shaped by a foreign influence operation, without any „deepfake” detector being able to spot the anomaly.
Data Security: The New Frontier of National Defense
To close this vulnerability, experts propose shifting to strict data provenance standards. If the origin of the information used to train a model cannot be verified, it should not be trusted for critical intelligence or defense missions. Furthermore, „red-teaming” exercises must evolve: it is no longer enough to check if an AI produces harmful answers; we must test whether it has already been influenced to favor certain political narratives.
The stakes are no longer just about identifying a fake video clip. The true information war is being fought for control over what the machines we have come to blindly trust are being secretly taught to believe.
