Predicting the Unpredictable: How AI Foretold Taliban Attacks

 

The success of Raven Sentry in predicting Taliban attacks underscored the potential of AI to enhance military intelligence, even as it highlighted the ongoing cat-and-mouse game between technology and human adaptation.

In the summer of 2020, American intelligence analysts in Afghanistan received a critical warning from a tool called "Raven Sentry." This AI tool, operational for only a few months, predicted a high probability of a violent attack in Jalalabad, the capital of Nangarhar province, at the beginning of July. The prediction came true, albeit a bit late, on August 2nd, when Islamic State militants struck the city's prison, resulting in 29 casualties.

The inception of Raven Sentry traces back to October 2019, a time when American forces in Afghanistan were grappling with diminishing resources. Troop numbers were dropping, bases were closing, and intelligence resources were being diverted elsewhere, all while violence was escalating. The last quarter of 2019 witnessed the highest level of Taliban attacks in a decade, necessitating a novel approach to counter the rising violence. This need for innovation led to the adoption of AI.

Political violence, as it turns out, is not as random as it might seem. Research by Andrew Shaver and Alexander Bollfrass, published in the journal International Organisation in 2023, showed that high temperatures correlated with increased violence in both Afghanistan and Iraq. When temperatures rose from 16°C to over 38°C, the likelihood of an Iraqi male expressing support for violence against multinational forces increased significantly. Raven Sentry took this understanding further by using historical data to predict future attacks.

A specialized team of American intelligence officers, colloquially known as the "nerd locker," was embedded within a special forces unit. This environment was conducive to aggressive experimentation, and the team began by studying patterns in insurgent attacks dating back to the Soviet occupation of Afghanistan in the 1980s. Contractors from Silicon Valley assisted in training a neural network to analyze correlations between historical violence data and various open-source information, such as weather data, social media posts, news reports, and commercial satellite images. The resulting model could identify when district or provincial centers were at higher risk of attack and estimate potential fatalities.

Initially, America's intelligence agencies and the Pentagon were skeptical of Raven Sentry. However, the tool's performance quickly proved its worth. By October 2020, the model achieved 70% accuracy, meaning that if it predicted an attack with the highest probability (80-90%), an attack occurred 70% of the time. This level of accuracy was comparable to human analysts, but Raven Sentry operated at a much faster rate.

Anshu Roy, CEO of Rhombus Power, one of the firms involved in Raven Sentry, found the AI's effectiveness intriguing. Optical satellites observed towns becoming darker at night before attacks, while areas associated with past enemy activity became brighter. Synthetic aperture radar (SAR) satellites detected heightened vehicle activity, and other satellites recorded increased carbon dioxide levels, though the reason for this last observation remains unclear. The AI also found that attacks were more likely when temperatures were above 4°C, lunar illumination was below 30%, and it was not raining.

Colonel Thomas Spahr, who documented the experiment in the journal Parameters, noted that modern attacks often mirrored those from the 1980s in terms of location, insurgent composition, and weapons used. Raven Sentry's ability to "learn on its own" meant it continuously improved until its shutdown in August 2021, following America's withdrawal from Afghanistan. Despite its termination, Raven Sentry had imparted valuable lessons to its human operators, who used its predictions to direct classified systems for more detailed investigations.

Since Raven Sentry's shutdown, armed forces and intelligence agencies have significantly invested in AI for predicting attacks, known as "indicators and warnings." These models have matured rapidly, with advancements such as SAR images now offering ten-metre resolution, capable of identifying objects smaller than a metre. A similar model, trained on data from Ukraine's front lines, would quickly become highly effective, according to experts.

Colonel Spahr cautioned that the process is not straightforward. Just as insurgents in Iraq and Vietnam adapted their tactics to counter American technology, adversaries would eventually learn to deceive AI systems. The Taliban's ultimate victory against the United States and NATO in Afghanistan underscores the potential for such adaptation.

America's development of Raven Sentry highlights both the potential and limitations of AI in modern warfare. While the tool demonstrated remarkable accuracy and speed, it also required careful human oversight to interpret its predictions correctly and understand its limitations. The evolving nature of warfare means that AI tools like Raven Sentry will need to continually adapt to stay ahead of adversaries who are also learning and evolving.

In a world where technology often promises more than it can deliver, one might wonder if, someday, we will need an AI tool to predict the predictions of our AI tools.