Last year was big for bots, as the popular definition of the software application shifted in the public’s understanding.
Previously, many of us had heard the term “bot” only in the context of the search engine. Search bots like the Googlebot crawl the web, methodically cataloging its content. This year a whole new class of bots showed up–social bots that can converse with human beings and carry out tasks like taking pizza orders, booking travel, or just chatting.
In April Facebook announced that developers could create bots on its Messenger platform. Many of us rushed to try out the bots that had already been developed for the platform with Facebook’s help. But it turned out the bots were limited, and one often ran into roadblocks when it couldn’t answer a question and carry on a conversation. Facebook admits this now.
“The first websites were really bad, the first apps were terrible; and the first bots weren’t great either,” Facebook’s David Marcus told the Telegraph in November. Marcus was CEO of PayPal before he moved to Facebook in 2014 to manage the Messenger platform.
Google trotted out its new chat bot, Allo, last spring, and while it does offer some nifty messaging features (like multimedia and response suggestions), when it comes to conversing with the user and fetching appropriate information, Allo is rudimentary. How many people do you know who use it?
@juberti Allo is terrible. Please make it not terrible.— ⭐James Welbes⭐ (@JamesWelbes) September 22, 2016
The biggest news about bots this year was the story last March about Microsoft’s Tay social bot, which, when turned loose on Twitter, soon started to spit out vulgarities and racial epithets it heard in some of its human interactions. Tay was quickly taken offline.
Ironically this happened just before Microsoft’s Build conference, when CEO Satya Nadella made official the company’s view that the future is about bots, and that the app era is starting to fade. Microsoft people have told me since then that Nadella’s pronouncements, along with all the bot-making tools Microsoft announced there, solidified, amplified, and accelerated the company’s bot-building work.
The big platform companies like Google, Facebook, Apple, and Microsoft have lots of people working on artificial intelligence, and many of them are applying AI in a bot setting. It’s pretty likely that we’ll see Act 2 of the bot story in 2017, and the user experience will improve somewhat.
A couple of basic truths about bots today:
In general, for every bit of information bots can deliver to a user, a human being had to have spoon-fed that data to the bot at some time in the past. Bots’ capacity for open-ended learning is very limited today. Bots can be given access to databases of information, but they must be taught specific ways of accessing and processing the data. Somewhere in the background, a team of data scientists, writers, and other creators must direct what a bot says on the front end, based on the desired personality of the bot and the target audience.
Lily Cheng, who leads social bot creation at Microsoft, told me that one of the reasons Tay started spouting hateful comments is because it was put into a public forum–Twitter–where it was exposed to the wrong people. Tay, she said, was meant for use by smaller groups of people, not for a wide public audience.
Today’s bots are horizontal or they’re vertical, but not both. That is, they know a little bit about a broad set of topics, or a lot about a very specific topic, but they don’t know a lot about a large set of topics. And bots don’t know what they don’t know.
Microsoft’s Cheng told me that the people doing the core research and development are letting bots learn in limited and structured ways. For instance a travel bot might be given a finite number of places to look for data that could somehow influence the logistics of an upcoming trip. It might search the calendar for the best times to fly. It could look for clues in email about the event to which the person is traveling to. It might look at travel records to look for go-to hotels for the trip.
But bot makers are also learning to manage users’ expectations. When the bot suggests a given hotel, it also tells the user what data was used to arrive at that determination.
Cheng also says the main reason people bail out on conversations with bots is because the user doesn’t fully understand the set of tasks the bot can handle, or which topics the bot can talk about and make sense of.
Bot developers are also starting to think about “emotional intelligence.” The bot needs to understand the difference between the user saying “that hotel was great” and “OMG that hotel rocked.” Bots need to show empathy when the user says something like “it’s 3AM, I can’t sleep, and all my friends are asleep.”
If companies like Facebook and Microsoft and others make big leaps forward on these problems, the bots we see in 2017 will be one of the year’s big tech stories. And this year’s bots might even make us forget about last year’s disappointments.
I think the more likely scenario is that we will begin to appreciate how difficult it is to make a piece of software (a bot) talk and act and emote like a human being. We’ll moderate our expectations, and appreciate that the technology has a long way to go to become something like the AI we saw in the movie “Her.”
At the same time this year’s bots will evolve into incrementally better experiences next year. They’ll be somewhat better at achieving specific tasks. They’ll be somewhat more agile in the way they gather the data they need to make intelligent choices. They’ll be a little more conversational, so that conversations with them can last longer.
And finally, bots will be more aware of their limitations, and will know when and how to excuse themselves and turn the controls over to a real human being.