Hands-on with Facebook M: the virtual assistant with a (real) human touch

A smarter child

There’s something eerie about realizing you live in a time people once considered a sci-fi fantasy. In Spike Jonze’s Her, Joaquin Phoenix's character uses an artificial intelligence-powered operating system voiced by Scarlett Johansson. He can ask her to oversee his calendar, play music to match his mood, and essentially manage his life. Of course, he eventually falls in love with her (it?).

Facebook M, the social network’s text-based virtual assistant, is only in its infancy, but I couldn’t help feeling like Her’s Theodore Twombly when I had it order me a sausage breakfast burrito. I didn’t even have to open a new browser tab, but I did have to go pick it up down the street — Amazon is trying to solve that problem via drones, I’m told. But it was still a good example of how Facebook, through a natural evolution of its software, has begun intertwining with our basic wants and needs.

Last week, I received access to the beta version of M, which lives inside Facebook Messenger, through happenstance. (Beta testers can invite other people so long as they live in California.) The service first launched in August to a small subset of users; Facebook says it has a "few thousand" across California. It’s been advertised as an easy way to find a good local restaurant, get factoids you’d rather not have Google search for, and setting reminders. M can’t yet access other apps on your phone the way Siri or Google Now can, but it’s designed to be more powerful than those services.

That’s because Facebook has a growing customer service-style team of M "trainers" at its Menlo Park office who oversee the software. The few dozen contractors take the steering wheel whenever the query is beyond M’s capabilities, though Facebook says humans monitor communication from start to finish. That’s how I got my breakfast burrito. A real person saw the request come through M and made the phone call to Good Eats Cafe on Pine Street here in San Francisco, putting it in under my name. Had the cafe had its own online ordering system, like many pizza places do, M could have used my credit card on file with Facebook to pay for the order automatically.

using artificial intelligence to order a breakfast burrito

Placing an online food order is a neat example of lifestyle automation, but M felt revelatory when I had it call Amazon’s customer service line for me. I told M I wanted to check on a refund I was promised for an item that never arrived because it was damaged in transit. I had been on the phone once before with Amazon regarding the item, and they told me to wait a few days to see if the orders page processed the change. When it didn’t, I asked M to investigate.

The AI assistant only asked for the email address that I placed the order with and the product name. About 35 minutes later, I suddenly had an email from Amazon telling me my refund had been issued. M chimed in a few minutes later, telling me, "OK - Amazon has informed me that your refund is being processed and the amount will be reflected in your account in 2 to 3 business days!" Welcome to the future, I thought.

Eliminating the need to sit on hold with Comcast, Ikea, or United Airlines is a glorious gift I hadn’t quite considered software capable of solving. But how can something this powerful, with real humans doing real work unseen in the background, scale to the more than 700 million Facebook Messenger users around the globe?

The team behind M, which includes the 10-person startup Wit.ai that Facebook acquired back in January, says it’s approaching AI in this fashion because it doesn’t want to constrain the service. "You have lots of AIs — like Siri, Google Now, or Cortana — whose scope is quite limited," Wit.ai founder Alex Lebrun told Wired in August. "We wanted to start with something more ambitious, to really give people what they’re asking for." The goal is to have M learn from the trainers over time and become smart enough to take on new tasks. The team plans to increase the number of trainers gradually, but Lebrun says it won’t be easy.

Even then, there’s something vaguely dystopian about M. Eventually, M could be the personal online butler of 1.5 billion people, helping us all lead more blissful lives as we tap out instructions to our software companion every day. Yet all the while, a sprawling network of contractors overseas will be there in the shadows to perform the more complicated tasks, and for a pitiful wage — not unlike the backbone of low-paid workers who power corporate customer service today.

Despite the humans behind the scenes, M can be awkward and stunted. The software is less personable than your standard chatbot, and more like a 1-800 call service automaton than you’d sometimes like. After all, according to Lebrun’s team, M’s AI is based on two old-school algorithms that form the statistical and probabilistic backbone of many automation services today.

Facebook steers clear of controversy

M won’t go anywhere near partisan politics or touchy subjects. Ask it what’s going on in Syria, and it will link you a BBC timeline of the war-torn country’s ongoing conflict. Ask it about climate change, and M will reply that it doesn’t have an opinion on that. The same goes for vaccinations. What about if jet fuel can melt steel beams? "Unfortunately I didn't find any conclusive answer to your question. The topic is heavily debated," M replies. Accurate, for sure, but only mildly more interesting than not having an opinion at all.

Facebook’s take on the personal assistant provides an interesting contrast to Siri and Cortana. Those services’ usefulness, which is low, usually takes a back seat to the stupid jokes, silly responses, and back-and-forth rabbit holes you can go down when conversing with your smartphone. Siri can’t order you food, and definitely can’t call customer service lines, but it does have an opinion on Steve Jobs.

M can’t talk to you, and its recommendations for restaurants, movies, and video games tend to rely on Yelp ratings and your favorite genres and food types (which you’ll have to input yourself). It recommended Halo 5: Guardians because I told it I owned an Xbox One and like shooting games, and it told me some titles off the New York Times bestseller list when I asked what book from 2015 I should check out. But it doesn’t have any taste of its own, and it seems unlikely the human beings behind M will be able to imbue it with those qualities.

So ultimately M ends up feeling more human than Siri or Cortana in its ability to get things done, but less capable in displaying the human-like qualities we expect out of our futuristic AI. It’s also an interesting step sideways from the personality-devoid and data-based approach of Google Now, which relies on its massive data hoard to predict what you’d want to know.

Still, there's hope for M. After asking it if I could send an invite to a friend, M kindly reminded me that I have five total invites and now have four remaining. Confused, I asked whether it had sent out a previous invite I requested a few days earlier. M corrected itself, "Yes I did! I’m sorry, you’ll now have 3 more friends invites!"

Was that M talking, or a human that input that response? I have no way of knowing — and that seems to be the point.


Verge Video: 360-degree video is coming to Facebook

The best of Verge Video