This is part two of my story about the ontological dissonance with artificial intelligence. You can catch up on part one over here.

TL;DR: I proposed that we should stop trying to build artificial intelligence into our image and allow it to develop its own identity. Otherwise its growth will be hampered and we will end up building a fundamentally flawed technology.

The misguided founding premise of AI development contributes to its heated and emotional public perception: what is it, what it can do and what we should expect of it.

The scourge of failed expectations

If you tell an average person that they have a virtual assistant with artificial intelligence on their iPhone, there is a good chance they will immediately ask it when it was born or what gender it is. Such abstract and anthropomorphic questions blow Siri’s mind. It will do whatever it can to appear useful — come back with a witty remark or do a quick Google search and rather mindlessly show the results. Not very helpful, is it?

Reality hits the user like a brick: despite its clearly human traits, Siri is not human. It doesn’t think and communicate like us, it is not as capable as us. It comes across as kind of dumb every now and then. Even though Apple says “Siri has answers to all kinds of questions” — it really doesn't. In fact it has very, very few (which it gets right more often than not) and for the rest, user is asked to train it manually with

Specialization seems to be the most viable chatbot design strategy today. Chatbots which focus on a narrow expertise are more accurate and faster than generalists. The problem is that people already associate “AI” or “virtual assistants” with amazing and limitless intelligence in a package which is both fun and professional in a very familiar, human way. General public doesn’t accept the notion that AI is in very early stages of development and unless it specializes in one thing at a time, it quickly becomes awkward and unusable.

How do we let AI breathe?

The following train of thought is very common:

1. AI is inhuman therefore it doesn’t care about us.
2. AI is incredibly smart. We don’t know when it will become omniscient and we won’t be able to stop it.
3. AI will leapfrog and kill us once it realizes we pose a threat to its survival.

I love how the first point contradicts the entire argument: AI is supposed to be inhuman, so let’s ascribe it all the humanist attributes we can: feelings, desire to learn, survive and destroy. We really have no idea what motives will drive it, we don’t even know if it will have any motives at all.

That is the predicament we find ourselves in, so the question is: what can be done to help AI develop its own identity and in turn allow it to serve us better?

Educate

Calm, measured and persistent education on AI’s identity and capabilities will take years. Marketing hype is not helping. It won’t go away immediately, but if you can — don’t let your marketing department write copy for your chatbot. Customers will hate yet another sales pitch. Your chatbot is not a vessel of your brand, nor it is a marketing brochure. Build it to solve actual problems, not drive your customer engagement. Formulate its functionality and personality based on user research, not business metrics.

Build it right

As designers and engineers, we need to take careful steps to create our chatbots so that they leave no doubt as to what they are and what they can do. Nothing more, nothing less. Make it clear from the start and every single time your chatbot communicates with the user. Consider a unique personality, a voice which doesn’t attempt to mimic human characteristics. It is possible to build trust with the user without resorting to the ‘human-like warmth’ trick. Show the value by making the chatbot responsive, usable and useful. Make it engaging with tasteful and elegant design. Allow your chatbot to build its reputation and rapport with the user through its accomplishments and not by forcing a copied approximation of some cheesy persona.

Your voice is you

Sensory deprivation of one sense strengthens another. If we hear noises and don’t see what is coming, imagination kicks in and paints often unsettling pictures. When we interact with a voice coming out of a phone, or represented by chat bubbles on a screen, we disseminate the voice and build an image of the person we’re talking to. Gender, tone, pacing, vocabulary, accent, slang, punctuation (or lack thereof!) — it all matters. Just like in graphical UI design creators use typography, colors, textures, layouts and animations to give products personalities and tell stories, in conversational design the voice is where it’s at.

In spoken communication, the tone, pacing and delivery of words, sentences and longer passages is what makes or breaks the perception of comfort or awkwardness. Try saying “hey, what’s up?” quickly, slowly or with uneven pauses between words. Try to drag out “up”. Say “hey” with a lot more energy than the rest of the sentence.

in written communication, low caps may come across as lack of confidence.

When you omit punctuation you create an impression of hip carelessness which is often interpreted as forcing it to be cool

YOU KNOW WHAT HAPPENS WHEN YOU USE ALL CAPS.

Changing the typeface has a big impact too.

Long or short sentences, complex or simple structure, vocabulary, slang, spelling variations, deliberate spelling or grammar mistakes. Typeface, font size, spacing, color and styles. All of these attributes participate in the same process as in spoken communication: every different delivery carries a different emotional load, implies a different state of mind of those who communicate and in turn evokes an appropriate, distinct response.

When you decide on what voice your chatbot will have, you give it a personality. By default, it will be human-like because human is all we know and users will associate what they see and hear with what they know. However, be very conscious of who you are creating and don’t put too much color into it. Navigating the line between too human and too robotic is incredibly tough because users will naturally skew towards a humanist interpretation. Just avoid the extremes — stay away from a drinking buddy jokester, stay away from T-800 — and iterate to find a voice that stands on its own. Just don’t spend too much time researching user preferences: they will want human and you already know that is not the right thing to do.

An actual conversation

One of the fundaments of human communication is context maintenance and switching. We are able to maintain several trains of thought at once, switch topics, combine them and so on. Conversation is a free flowing exchange of information and both parties contribute equally.

I think the aspect of context should be the top priority for chatbot developers. The ability to maintain context in a conversation fulfills the fundamental needs of effective communication: paying attention, understanding and ability to contribute in return.

There are chatbots which know how to maintain context. They will answer follow up questions without forcing the user to describe what they’re referring to every single time. We have one of those at Sourcebits and the solution is as simple as it is smart. We are off to a good start, but this is a very complex problem: language nuances or cultural context make it very tough for AI to figure out when to maintain or when to switch. AI cannot infer meanings from body language, references or tone.

Context maintenance proves understanding. I think it is the first step to make chatbots effective communicators.

Use the best practices of UI design

Throughout the last 10 years, we learned a lot about user’s behavior, expectations and mindset in cognitively challenging situations. Mobile devices pose a challenge which optimized the quality of design: we know our users pretty well now.

I think that experience applies in the field of AI as well. We have many tools to optimize conversational interfaces. Consider the following:

• consistent left-right participant placement in chat layout
• message and metadata typography
• chat bubble shape and color palette (allow use of color to help the chatbot to express non-verbal messages?)
• message animations (types, timings, frame rate)
• message grouping
• non-verbal responses (typing indicator)
• emoji, animoji, stickers (aren’t these too human though?)
• instead of forcing the user to type or speak, speed up data input with UI controls (buttons, segment controls, checkboxes, single/multiple choice lists, sliders, media carousels, video and music players etc.)

And last but not least: the language. I consider copywriting very much a part of effective UI design. Consistent and clear copy goes a long way: focused on actions, with concise and practical vocabulary, devoid of any ambiguity or emotional charge. Simple and to the point.