The Pentagon is pouring significant resources into “catching” deepfakes, operating under the assumption that the war against disinformation is primarily a detection game. Programs like DARPA’s Semantic Forensics and various defense-contracted tools are laser-focused on identifying synthetic images, audio, and video once they have already entered the information bloodstream. While these tools are impressive feats of engineering, they fundamentally misunderstand the evolving nature of the threat. The Department of Defense is preparing to catch yesterday’s digital forgeries while ignoring a far more sophisticated and dangerous reality: our AI systems are not being tricked by fake media; they are being trained to be biased through the poisoning of their foundational data.
The true vulnerability lies upstream, deep within the massive datasets that feed large language models and analytical platforms. Adversaries have moved beyond simply manufacturing fake videos; they are now actively embedding propaganda, skewed narratives, and doctored historical records into the very information pools that power the artificial intelligence defense analysts and policymakers rely upon. When a model is trained on a steady diet of state-controlled content—such as Iranian propaganda or laundered misinformation from Russian networks like RT—that bias becomes a permanent fixture of its “intelligence.” Once this poisoned data is baked into a model’s weights, the system doesn’t just display a “deepfake”—it begins to perceive, summarize, and prioritize the world through the skewed lens of a hostile power.
This isn’t a theoretical risk; it is a documented reality of the modern information landscape. Researchers from the Atlantic Council’s Digital Forensic Research Lab and institutions like the UK AI Security Institute have verified that propaganda machines—including known Russian and Chinese influence operations—have successfully infiltrated open-web archives like Common Crawl, which serves as a primary fuel source for many foundational AI models. Recent studies have demonstrated that it requires remarkably few malicious inputs—sometimes as few as 250 documents—to fundamentally corrupt the outputs of a large-scale model. Because the cost of retraining these massive AI systems is astronomically high, this poisoned intelligence often remains embedded indefinitely, waiting to be triggered by a defense analyst’s seemingly innocent query.
Crucially, this problem is exacerbated by the way we manage regional information. In conflict zones like the Middle East, government-imposed restrictions on journalists and satellite imagery serve as a natural filter that biases the data AI models consume. When local, state-controlled narratives become the only available record of events, the AI learns to treat those narratives as objective facts. Whether it is Tehran blocking internet access or major powers pressuring corporations to censor satellite views, these actions shape the digital reality for the systems that guide our high-level strategic decisions. We are effectively outsourcing the cognitive development of our analytical tools to the very adversaries trying to deceive us.
To bridge this critical security gap, the Pentagon must shift its strategy from “detection” to “provenance.” Trusting an AI model without knowing the verified history of its training data is a massive liability. We need rigorous standards that demand transparency, ensuring that only models built upon clean, audited, and verifiable information are used in sensitive intelligence contexts. Furthermore, current “red-teaming” efforts, which primarily focus on whether a model can be goaded into saying something offensive, are woefully inadequate. We must start testing for the presence of adversarial influence—actively searching for systemic bias injected into the model’s core logic during its training phase.
The most immediate priority, however, is a widespread cultural shift among defense and intelligence professionals. There is a dangerous lack of awareness regarding the feasibility and impact of data poisoning; many still view AI as a neutral calculator rather than a malleable system that can be quietly re-programmed by a state actor. The fight against disinformation is no longer simply about spotting a fake photograph or a dubbed video. It is about recognizing that the artificial minds we have come to trust are being slowly and methodically taught to lie. If we don’t secure the foundation of our data today, we will find ourselves fighting a war based on an artificial reality that we ourselves helped construct.