Anyone who's ever worked with any kind of artificial intelligence system knows that automation is difficult. Theoretically, an AI system works like this: the system recognizes as much as it can of the query asked of it then matches those terms against its database to come up with the most logical response - the bigger the database, the higher the accuracy. And that is how they generally work, but the problem is computer systems, no matter how sophisticated, are only able to calculate based on core logic. They can't (at least at this stage) think through the query and come up with nuanced conclusions. And the disconnect here is humans are inherently nuanced - the only way to create a truly responsive and accurate AI system is to have human input, to train the system by highlighting errors and offering alternatives.
This is why AI assistants like Apple's Siri are always a little bit off. Because there's no human input in that process - you have to know the right questions to ask, the right keywords to input, and Siri will then respond to your query based on previous responses and data it's learned from other users. But it never gets enough feedback to really improve accuracy - regional language variation, even individual colloquialisms can throw the system off, and when that happens, people lose faith in the product.
But what if you reverse engineered it? What if, instead of relying totally on AI, you went in with an AI system that was overseen, and assisted, by a team of human trainers who ensured the responses were accurate? More than that, what if those human trainers were also able to do human things on your behalf, like place orders and act intuitively, based on your personal data? And what if, in doing so, your AI systems were being trained with feedback at every stage, every step, accelerating your their learning with every interaction.
This is the approach Facebook is taking with their new personal assistant for Messenger, called 'M'.
Set to take on Siri, as well as Google's 'Google Now' and Microsoft's 'Cortana', M is built into the Messenger experience, avoiding the sometimes awkward process of talking at your phone and (repeatedly stating "yes" to confirm actions). Taking a different approach to their rivals, Facebook has baked M into functionality that's already popular amongst their user base - Messenger is one of the most popular apps in the world, boasting 700 million monthly active users. Working with what people use and are already familiar with, M will enable users to ask questions via message.
But the real differentiating factor with M is how it works. M not only uses AI data matching - it's processes are also overseen by real people. Facebook Messenger lead, and former PayPal CEO, David Marcus outlined the new functionality in a Facebook post:
Facebook's hope with M is to utilize the popularity of Messenger, combined with the data resources of Facebook, to create the ultimate utility for mobile discovery, stepping in on Google's long-held turf. While Google is still the dominant player in search, Facebook's hoping that M might be able to keep users within Facebook's walls longer - if you're already on Messenger and you want to search for answers, M will enable you to do so without clicking out and switching to another app.
"We start capturing all of your intent for the things you want to do," Marcus told Wired. "Intent often leads to buying something, or to a transaction, and that's an opportunity for us to [make money] over time."
If M can capitalize on the masses of data Facebook has in it's databanks, as well as the added input of human trainers refining and updating the system in real-time, it may just be the ultimate assistant service. But how do you scale such a process?
The M process works like this:
- Users will tap a button at the bottom of the Messenger screen to send a message to M
- M's AI software will analyze and assess the message to determine the nature of the query, then ask follow-up questions based on its algorithm and data matches
- Once the query is clarified, M will go about completing the query - whether that's a data-based response or an actual physical task - and will send an update when complete
In this process, users won't know whether a human has intervened, or whether the process was fully automated - the task is completed by M, that's all you need to know. In terms of physical tasks, M's trainers have customer service backgrounds and are able to make judgement calls on more intricate tasks, where the software might struggle to comprehend. But as they do, they're also able to input that data into the M system to better educate it on human response and signal interpretation. The more people use the service, the more they improve the AI, which, eventually, might be able to take over the whole process.
But not yet.
M has a dedicated team of trainers on contract roles, and Marcus anticipates that as the service expands, they may employ thousands of them, at a high cost. Given that, how can Facebook expect M to become a profitable offering?
Business, with a Capital 'M'
The success and failure of M, of course, is reliant on its accuracy. The more accurate and functional the system is, the more it'll be used, so getting it right early on is crucial. But if Facebook can get it right, if they can provide a functional assistance service via Messenger, then the potential revenue opportunities are significant.
"If, for instance, you have a lot of calls that have to be placed by people to cable companies, that's a pretty good signal that their customers would actually like a better way to interact with the company and maybe they should have a presence inside of Messenger directly," - David Marcus to Wired.
This is just one example of how M data could be used to power enhanced customer service opportunities via the Messenger platform, an area of focus for the next evolution of the service. So far, Facebook hasn't pushed to monetize Messenger, ever wary of impacting user experience (with lessons learned from previous changes to Facebook proper, most notably to its infamous News Feed algorithm). But that's soon set to change - they've already announced new eCommerce options for Messenger, including improved customer service and payment options for the platform. With M, Facebook aims to build a vastly improved data set on how people are using the platform and what queries people want responses to - imagine if, when setting up your next Facebook ad, Facebook's system could tell you what questions people were most commonly asking about your niche, even your business specifically? The bigger the data set, the more Facebook will be able to assist brands to better utilize the platform for marketing and advertising purposes - if successful, M could change the way businesses use Facebook for outreach purposes.
But then, of course, that's a big 'if'.
Sink or Swim
Essentially, Facebook is banking on their ability to deliver with M. The program is starting small, with Messenger users in the Bay Area being the first to get access, and a slow roll-out planned from there on in. This will enable Facebook to scale their AI system in a more accurate and regionally-focussed way - it's likely that each location will need its own, separate M servers and systems to help it account for localized variations and dialects, and local knowledge from the M trainers will also, no doubt, play a part. Taking a location-by-location approach also makes it easier for Facebook to drop the project if it's not successful.
The crucial element in the initial stages of the M project will be accuracy and speed - if Facebook can provide a great service, that will increase user adoption and enable them to grow the functionality of the system. Given the breadth of Facebook's data, and the focus they're putting on training and refining their AI, M is possibly best-placed among the assistance services to see success and to be of most use. But it depends on how Messenger users see it, whether it's seen as a cool new function or an intrusion on what's considered a private space.
It's one of the many challenges the project needs to confront, but the logic behind M is sound, the approach - a different take on the AI process - is solid. Now we just have to wait and see if M can win over the masses.