In 2009, Google introduced Gmail Autopilot.
“As more and more everyday communications take place over email, lots of people have complained about how hard it is to read and respond to every message. This is because they actually read and respond to all their messages.”
Autopilot replied to email as if it were actually you. It included an authenticity control panel to adjust for “tone, typo propensity, and preferred punctuation.”
Autopilot was launched as an April Fools’ prank. It was one of several programs tied to Google CADIE (cognitive autoheuristic distributed intelligence entity). Google provided several examples that demonstrated how helpful autopilot could be, including this one:
Should two Google users have Gmail Autopilot activated, the automated systems will chat with each other for up to three messages. Beyond that, the blog instructions noted, “our experiments have shown a significant decline in the quality ranking of Autopilot’s responses and further messages may commit you to dinner parties or baby namings in which you have no interest.”
This ridiculous prank wasn’t too far off. In the fall of 2015, Google introduced an AI-like tool called Smart Reply for Google Inbox. Google recently shared that as of Feb. 1, 2016, Smart Reply was in use for more than 10 percent of mobile Inbox replies.
Google Inbox is an alternative and compatible program to Gmail. It’s Google’s experimental, futuristic take on email. While the tech landscape has changed dramatically over the past decade, email has changed very little. Outlook was designed first as a fat client and later adapted to the web. Gmail was born in the web but was built for the market-leading Internet Explorer browser. Google created Inbox to see if it could change the game without forcing all of its Gmail users on to the adventure.
The AI potential of Inbox’s Smart Reply
Inbox, introduced in 2014. offered a “completely different type of inbox, designed to focus on what really matters.” For most people, organizing and finding information in email is time consuming. Critical and irrelevant messages arrive with equal fanfare. It’s up to the user to manually sort and organize content so that appropriate actions can be taken or the content can be found later if/when needed. Inbox offers assistance with features such as automatic bundling and highlighting, as well as improved user-controllable notifications.
However, Smart Reply is the killer feature of Inbox. It’s still very limited today, but it has strong potential. Unlike Inbox, the Smart Reply feature doesn’t have much of a learning curve. Google automatically scans incoming messages to determine if a reply is appropriate (newsletters, confirmations, etc. do not generate responses) and then uses its neural network to generate three appropriate, short responses. Selecting a response opens a pre-populated reply that can be sent or further edited. I would not be surprised if Smart Reply finds its way into the old-school Gmail client.
It is important to note key differences between fictional and humorous Autopilot and the useful Smart Reply. Smart Reply doesn’t actually send a response—it offers three options as a starting point. The magic of Smart Reply is its ability to determine the type of reply that’s appropriate, and it does so impressively well. It doesn’t know whether a yes or no is the preferred response, but it does understand the sender is seeking a yes or no response.
Smart Reply is used by more than 10 percent of mobile Inbox replies. That’s high enough to prove the concept, but it’s still too low to be truly considered useful. In my opinion, the issue is Smart Reply isn’t personal enough. Smart Reply gets smarter with usage as Google tweaks its algorithms; however, they are crowd-sourced optimizations. For example, I often reply to meeting request inquiries with a link to my online calendar tool. That’s too user specific for Smart Reply.
Another big limitation of Smart Reply is it’s not integrated across other applications. Currently, if someone asks me when we might be able to have lunch, Smart Reply offers a guess such as Tuesday. It would be nice if Smart Reply could look in my Google Calendar to see what days are open for lunch. Even better, it could note locations to factor in commute time. These integrations seem trivial (and inevitable) compared to what it already does.
In many ways, the hard aspects of artificial intelligence are being solved before the useful parts. Google Now (on Android) is similar. Both are just a few steps away from being truly useful. Google Now tells me when it’s time to leave for the airport. It determines this by comparing my current location, drive time and flight departure information obtained from my calendar. That’s impressive, but it’s wrong. I usually take the bus to the airport and, in fact, query Google to determine the bus schedule. Google Now should tell me the time to leave for the Park and Ride.
AI in the near term
The best AI solutions, certainly in the near term, won’t make decisions for us but will foster our decision-making processes. Smart Reply takes the right approach by offering responses rather than responding. AI is coming at us like the proverbial freight train—Google, Facebook, IBM, Microsoft and even GM are investing heavily in AI.
Fully automated AI, such as self-driving cars, are useful only in situations that can be well defined. Fully automated AI is powerful, but takes a lot of time to perfect. The shorter-term opportunities for AI will be more interactive, such as AI that asks how we intend to get to the airport, so it can recommend a departure time. I still want to determine who I have lunch with, but a suggestion of Tuesday is only useful if I am open on Tuesday.
Google Autopilot was a joke. How can a computer know how to respond to all of my emails? I hear smart people say things like AI will soon determine what I want to eat—unlikely at best. The big near term AI opportunity is to be helpful, not magical.