2016 is kicking off with a lot of buzz about AI. Mark Zuckerberg’s 2016 personal goal is to build an Iron Man-style AI. Eric Schmidt has been musing about a future with “Eric” and “Not-Eric,” in which Eric is himself and Not-Eric is “this digital thing that helps me.” Phil Libin has been calling AI bots the most important tech trend of the year.
I’m pretty excited about AI and machine learning too. But it’s hard to square this enthusiasm with the present-day performance of machine intelligence. The current crop of services that attempt to interpret natural language — Siri, Amazon Alexa, Google Now, Cortana, etc. and a long list of domain-specific bots — are intended to make our interactions with technology more human, but they fall far short of that promise.
Below I have a few ideas about how design might help bridge the gap between the AI dream and the present-day reality. And also some thoughts about how Slack might give Google / Apple / Facebook / Microsoft a run for their money.
It’s no wonder people are excited about the potential of artificial intelligence. Computer interactions are prescriptive and brittle; what if machines could converse like humans instead of humans flexing to machines? And as devices and services are able to act more intelligently, could AI take more off our plate?
It’s also clear why Google, Facebook, and Microsoft are interested in building a layer that sits on top of the app-centric way we use devices today. This new layer could dynamically route your tasks and questions to the best services and data, without requiring you to curate a collection of apps on your home screen.
But even the people most enthusiastic about AI recognize that human-level machine intelligence is probably still a long way off. A recent poll of experts suggested that there’s a decent chance we’ll achieve human-like artificial intelligence by around 2040, which is actually quite soon (and probably overly optimistic), but still 25 years from now. (Apple, for example, has been thinking about AI since 1987, and 30 years later we’re still a long way away from their vision.)
And in the meantime, there’s no guarantee that conversational interfaces are an improvement over graphical interfaces. Consider this interaction from a recently launched bot:
Responding with “3” to this sort of message doesn’t feel particularly efficient or natural. Most people don’t use a command line to navigate their computer or phone for good reason — the UI hints and the graphical interface are helpful.
Recognizing that the state-of-the-art natural language processing used by Siri, Google Now, etc. leaves much to be desired, some companies are experimenting with “Wizard of Oz” artificial intelligence. Facebook M, along with startups like Fin, Magic, and Operator are blending human intelligence with artificial intelligence — you might think you’re chatting with a bot, but actually there’s a person on the other end. It’s a promising direction (and will generate a lot of good training data), but it’s expensive and slow (Magic recently announced they’d be charging $100/hr for their premium service).
And so in the meantime, while we’re waiting for more sophisticated artificial intelligence, here are 3 ways design can help bridge the gap:
1. Syntax Hints
The open-ended “blank canvas” nature of bots is both a strength and a weakness. There are few if any cues directing you towards questions that the service is likely able to answer or tasks the service is likely able to perform.
What if a bot could steer you towards the syntax it understands? Here are a few rough mockups of how that might work:
You probably would have guessed that “yes” and “no” are valid responses, but in this example the “Snooze: 10 minutes” and the “Snooze: remind me when I get to work” options are also exposed.
Or imagine a bot that would help you order lunch:
Or maybe the best input interface isn’t a text box at all. A question might be best answered with a photo, or a map, or a slider, etc. Selecting a time slot in the example below would be much easier when the context of your calendar is displayed inline:
In some cases there’s a benefit to putting multiple form elements on the screen at once. For example, at Abacus we’ve written our own internal status update bot. It‘s a great tool, but it has one shortcoming: it’s difficult to put an update in a single-line text input. It would be helpful if we could edit the ‘last week’ update and the ‘this week’ update at the same time.
Another complementary approach would be dynamic autocomplete suggestions (Slack’s slash commands are a step in this direction, but they’re difficult to discover).
Enabling bot developers to suggest word completions, common responses, and contextually relevant inputs seems worth exploring. At its best this approach would combine the advantages of conversational messaging with the advantages of app-like graphical interfaces.
2. Be As Smart As A Puppy
Matt Jones coined this phrase ‘BASAAP’ 5 or 6 years ago. His point was that we should try to design things “that don’t try to be too smart and fail, and indeed, by design, make endearing failures in their attempts to learn and improve. Like puppies.”
Given that we’re still a long way away from bots with human-level intelligence, maybe we should set our sights lower. It’s disconcerting that you interact with Siri or Alexa as though you’re talking to a human (and then are inevitably disappointed when they don’t *get* you).
What, if any, personality should bots have? And how much should bots try to infer sentiment? (Are you happy, angry, busy).
Or at the other extreme, maybe we should resist the temptation to anthropomorphize bots entirely. Would we be better served by treating these things as what they are: collections of scripted interactions supported by machine learning?
Whatever the approach, the language and tone we apply to these bots is an opportunity to signal the degree of comprehension people should expect.
3. Domain Selection / Cut Me Some Slack!
Siri and its competitors stumble in part due to the massive domain they address. Ostensibly you can ask Siri anything. You may want to send a text message, look up directions, or select a movie showtime (all likely to work). But you are unlikely to have much success if you want to “make a reservation at that place with the blue awning on Eldridge St.” or find out “how much warmer it will be in Chicago next week vs. NYC” or “buy a bag of those pretzels with peanut butter inside and ship it to my apartment.”
Contrast the domain Siri addresses with the domain of “work.” In the near term, our work lives might be a better candidate than our personal lives to benefit from AI, because our work lives have more repetition.
Here are a few things I do regularly at work:
- Schedule a meeting
- Figure out where to go for lunch
- Meet so-and-so in the lobby
- Check performance metrics
- Submit business expenses for reimbursement
How much of your workday is made up of these sorts of interactions? It’s telling that many people have assistants at work, but very few have assistants at home. Our personal lives are harder to automate.
It’s for this reason that Slack may be uniquely well positioned to build a successful bot (or provide the platform for others to do so).
In the past, AI has been successful within narrow domains. Artificial intelligence has conquered chess, and is on the verge of mastering driving. But true general, human-level machine intelligence is still a ways off. And it seems like “work” might be a promising domain to make some headway.
Wrapping things up —
Past advancements in AI have been almost solely the output of R&D labs. But this next wave of AI development is going to require a lot of work on the front end as well as the back end. Though machine intelligence is still a long way from human intelligence, the progress that engineers have made thus far along with the evolution of platforms (first iOS and Android, now Slack?), are enabling big, fundamental changes to the way we interact with technology. Designers will play a big role in making the most of this opportunity. And with a bit of luck, 2016 might live up to the bot hype.
