Conveying the Boundaries of Artificial Intelligence

When confronted with AI, somehow all we want to do is break it.

When I first ran into SmarterChild over a decade ago I remember how my friends and I all tried to trip it up and test the limits. What if you ask it’s favorite color? What if I ask questions about myself? It inevitably failed and we laughed. Silly computer, intelligence is for people. My teenage friends and I were not alone in this instinct — can you remember the first time you tried to test the limits on Google? Or have you seen all the blogs and articles devoted to just the funny or wrong things Siri says? It is human nature to push these limits.

While the ‘Wow’ moments may be quick to impress, we’re actively seeking AI’s failures. Unfortunately, we haven’t standardized on a way to set expectations about limits. So, how do various products go about conveying limitations and barriers to their abilities to a deliberately antagonistic audience? Here’s a quick look:

Type: No restrictions, anything goes.

When to use it: An answer is better than no answer and you’re able to provide multiple alternatives.

Examples: Your favorite search engines

Okay so there’s no quick tips or instant search box but it is better than nothing because sometimes we all need help with these things.

A great example is Google Web Search — sometimes you really have to stop to think about how wonderful Google is. To not get an answer you practically have to type something totally convoluted or extremely long. For over a decade, people have been Google Whacking (i.e finding two english words that don’t return a result) in an attempt to find its limits. Of course, it’s a limited practice because as soon as you found one there is inevitably a search result to cover it.

To be able to attempt this type of AI you need a Google-like breadth of responses — this means almost perceptually infinite pieces of data to train on.

Type: Unseen barriers

When to use it: You can provide answers to a fairly wide variety of questions but are limited to single pre-programmed answers.

Examples: Slackbot, Siri, Alexa, Cortana, Google Now

Y U no understand me slackbot???

If you’re using something with unseen or conversational interface you’re pretty limited. As a product, you basically get one chance to get your answer right so if you don’t have a magic one-box or pre-programmed answer you have to present the user with boundaries that are unknowable until encountered. You can convey this with a simple ‘no results found’ but more often companies have opted to soften this.

This ‘softening’ is accomplished two ways: humor and redirection. Adding in humor takes the bite out of rejection — you didn’t hit a wall, you found a new inside joke. Redirection on the other hand, e.g. ‘You could google it’ ,means that while the service isn’t able to provide an answer they’re passing you off on the hopes that someone else will. This isn’t ideal but prevents the user from hitting a dead end.

Type: Severely restricted

When to use: when your ability to answer questions is severely limited.

Examples: Your average chat bot

Playing around with bot-maker chatfuel.com If all you can answer is “To be” or “Not to be,” Don’t let your users input something else.

These are primarily popular because 1) chatbots and AI are cool! So everyone wants to hop on the bandwagon 2) in theory it provides users a more natural way to interface with your product. Already in Messenger? Of course you should be able to check your stocks or get the weather without leaving! These extremely limited option sets try to clearly indicate just how restricted the bot is in order to better set expectations. In practice, however, they can easily becoming a clunkier interface of branching options. Imagine a shopping bot. You want to shop shoes? What kind of shoes? What size? The seemingly endless decision trees are quickly exhausting to navigate for anything but the simplest operations.

That said, these are comparatively easy to create and whole platforms, like Chatfuel, exist on the basis of enabling these often simple transactional interactions.

Type: Hidden Magic

When To Use: Enabling micro-AI on an existing platform

Examples:

One of my favorite Hangouts features. I’ve started actively missing this feature on my laptop or when using other chat services.

This is by far my favorite example of clever AI — it’s giving me things I didn’t even know to ask for. While you could argue Google snippets in search are just as clever in parsing, those only occur when I’m already asking a question. This is the type of magic I never even realized I wanted.

Why are these so good? They happen silently. When Aparna Chennapradaga spoke about AI at Products That Count, she talked about balancing ‘WOW’ to ‘WTH’ moments. The delight of these hidden magic moments is that particularly when you’re cautious about triggering them, you almost always skip the ‘What The Hell’ and only get the ‘WOW.’ When my friend asks how much movie tickets were in Messenger, I don’t say “$10” because I mean to start a transaction, it just anticipates my needs and gives me the option to.

Because users almost never consciously evoke this, they don’t approach with a ‘let’s break this’ attitude nor do they face disappointment when they encounter subtle barriers.

So what does the future of AI look like?

As always, our expectations will rise and we’ll quickly expect more and more out of AI systems. The technology that currently wows us will be expected minimal functionality and we’ll have little patience for ‘severely restricted’ interactions or bots with too many unseen barriers. While personal assistants may seem exciting now, users will want more and more hidden AI where they don’t even have to request the action. One-off bots may experience short-term wins but long-term, the winners will be platforms. Platforms that can subtly integrate AI feature sets without perceived bloating or complexity of interaction will find themselves delighting more and more users. Until then, we can look forward to proliferation of bots with basic functionality and.. Well. lots of boundaries