Facebook's AI failed to detect live video of last Friday's New Zealand mosque attack because the system simply isn't smart enough to recognise a mass shooting filmed by the attacker, the social network says.

In a Wednesday blog post, Facebook offered a detailed rundown of how it responded to video of the terrorist attack, which was live streamed on the platform.

Facebook has been using AI-powered algorithms to automatically take down videos and images containing pornography and terrorist-sponsored content. The algorithms can do this because they've been fed a large number of real-world examples of the offending content.

Last Friday's shooting slipped past the system because Facebook doesn't possess training data depicting live-streamed mass shootings. The events are thankfully rare, wrote Facebook VP Guy Rosen.

Complicating the matter is the prevalence of videos on Facebook depicting video game footage from first-person shooters. "Another challenge is to automatically discern this content from visually similar, innocuous content—for example if thousands of videos from live-streamed video games are flagged by our systems, our reviewers could miss the important real-world videos where we could alert first responders to get help on the ground," Rosen added.

As a result, the company had to depend on users to learn about the shooting footage. However, during the live broadcast, the video was only viewed 200 times, none of which generated a single user report. It was only 12 minutes after the live stream ended when a Facebook user finally reported the content.

Before the original video was taken down on Facebook, it had been viewed 4,000 times. But eventually more users began to share and post new copies of the video on Facebook.

Rosen blamed the video's circulation on three groups: bad actors who were deliberately trying to keep the footage up; media outlets covering the incident; and regular users curious about the event. Not helping the matter was how new copies of the videos were being shared in different formats, making it harder for Facebook's automated systems to detect and pull it down.

"In total, we found and blocked over 800 visually distinct variants of the video that were circulating," Rosen said. Facebook's systems eventually became smart enough to automatically remove more than 1.2 million videos of the attack during the upload process. However, about 300,000 additional copies were only taken down after they were posted.

So how might Facebook prevent another shooting from going viral? The company didn't offer any clear solutions. "Some have asked whether we should add a time delay to Facebook Live, similar to the broadcast delay sometimes used by TV stations. [But] there are millions of Live broadcasts daily, which means a delay would not help address the problem due to the sheer number of videos," Rosen said.

However, the company does plan on improving its AI systems to quickly detect and prevent problematic videos from going viral. One experiment involves using audio-based technology to identify different formats of the same video.