Facebook's M blends AI assists with human help

M is unique in the hotly competitive field of AI. It's meant to provide information, like the best nearby hiking spots, or the best burger joint. But it's also designed to complete tasks, like help someone order flowers for a parent's birthday, recommend which baby shoes to buy, or book travel arrangements.

"It's powered by artificial intelligence that's trained and supervised by people," said David Marcus, heading of messaging at Facebook, in announcing the initial test launch of the service on Wednesday.

M exists as a feature inside Facebook's Messenger app, currently only available to some Messenger users in the San Francisco Bay Area.

For those who have access to it, answers to some questions will be provided purely based on AI, while others, like those related to tasks, will be handled more by human supervisors paid by Facebook.

This distinguishes M from other digital assistants like Apple's Siri or Microsoft's Cortana, which are focused on providing information but, if stymied, can't hand questions off to humans.

The distinction could also be M's Achilles heel.

For it to succeed, Facebook must strike the right balance between using AI and using humans for people's queries, both to make M cost effective, and to ensure that people's questions are answered appropriately.

In some cases, M might have to make trade-offs, perhaps having the AI provide a less than ideal answer rather than passing someone's question on to a more expensive human, said Mark Riedl, associate professor at the School of Interactive Computing at Georgia Tech. Riedl's research focuses on the intersection of AI, virtual worlds and storytelling.

"It's an economic question," he said.

The concept behind Facebook M is not entirely new. An emerging field of research known as "human computation" seeks to study the effectiveness of blending human interactions with AI.

The research is built on the premise that if an algorithm is viewed as a set of processes, some of those processes could be performed by a computer and some could be performed by a human.

In a way, the research bodes well for Facebook. Some findings have shown that humans plus algorithms can generally do better than humans or algorithms alone, Riedl said.

Still, some of the biggest recurring questions, he said, have revolved around figuring out the interface between humans and computers. When should humans do the work, and when should algorithms Should the algorithms be in charge of the humans, or vice versa

Facebook is likely to gradually make M more broadly available, depending on the success of its initial tests ... and how well the company itself can answer those sorts of questions.