Precognition of disasters

It's Tom here, covering Tristan as he takes a two-week sojourn through time and space.

I've been struck by the recent influx of AI systems that promise to predict attacks by enemy combatants. In the last week alone, two eye-catching examples have emerged.

The first was courtesy of the US military. General Glen VanHerck said the Pentagon is developing AI that could predict events "days in advance." Cue the inevitable Minority Report comparisons.

The system was recently tested on a simulated threat to a crucial site (think: the Panama Canal).

"We would take artificial intelligence and use machine learning to take a look and assess, for example, the average number of cars in a parking lot that may be there in a specific location to a competitor or a threat," General VanHerck said during a press conference. "And we monitor that over a period of time."

Days later, a research team unveiled their own attempt to predict attacks by adversaries. The group said they've developed models that predict and explain acts of terrorism.

The researchers were more measured in their claims than General VanHerck. In tests, their algorithms proved effective at predicting attacks in places with a fairly high prevalence of terrorism. But they acknowledged that "black swan events," which occur very sporadically, are almost impossible to predict.

Nonetheless, their study provided an interesting analysis of worldwide terrorism trends. 

They found, for instance, that in regions strongly affected by terrorism, such as Southeast Asia, the risk of a terrorist event peaked roughly 200 weeks after a previous attack. However, in areas with a lower terrorism prevalence, such as Russia and Eastern Europe, the risk reached its apex at around 400 weeks:

This finding highlights that terrorists do indeed tend to target the same locations multiple times. The reasons for this can be manifold, but in general, it seems that these locations reflect the strategic equilibrium between reachability (i.e., cost for attacking for the terrorists) and the strategic value (i.e., benefits of an attack) of a location.

Their moderate claims made me suspicious of the US military's aplomb, particularly as the Pentagon won't disclose any details about the system or the trials. There's always a risk in trusting decisions made by opaque algorithms, especially when they're used in military applications.

However, the AI used in the experiments seems less scary (and interesting) than the headlines make out. The system sounds more like an incremental advance on the type of data analysis that the Pentagon has been doing for years.

It might produce more accurate predictions, but that doesn't mean they will be heeded. From 9/11 to the US Capitol riots, American leaders have shown a tendency to mishandle warnings of deadly events.

If you want to read more about the use of AI in warfare, check out these Neural stories on quantum computers in the military and the Pentagon getting sucked into the AI hype machine.