Design for AI

Episode 6

Why creating a personality for your AI is important, be it recommendation system or AGI. We cover the steps needed to evaluate your system and come up with the best personality for your users..

Music: The Pirate And The Dancer by Rolemusic

Background research links

https://blog.prototypr.io/a-guide-to-developing-bot-personalities-c6eba213d77b?gi=88dd055fd56b https://chatbotsmagazine.com/designing-a-chatbots-personality-52dcf1f4df7d https://medium.com/the-charming-device/how-to-design-intelligence-3-ways-to-make-human-centered-bots-76c5ff7524df Transcripts Today’s episode is about personality, So I thought it best to start with a scenario: For example, you are in the market for finding a lawyer, and like most people looking for a lawyer you need to watch your money. you’ve heard good things about some companies providing virtual lawyer services. You download one since it was the top rated since it was so  friendly. You get started telling it about the background and back and forth is full of jokes from the lawyer. But the jokes just seem off. Then you need to find some more info and take the device down into the basement the virtual lawyer says it lost its network connection and just starts laughing maniacally. Maybe somebody finds this funny, but if they messed up this bad on the humor, you have no confidence that they got the legal part right. Delete that one, obviously friendly was not the way to go. You download the next one rated totally professional. You start the process but it is taking forever. You have to go through one question at a time. This thing feels like it is reading war and peace off of a DMV form. You find yourself getting lost in the monotony and realize you skipped over the most important nuance. This isn’t professional, this is fingernails slowly scraping a blackboard. Ugh, there is no way you’ll make it through the process and remember everything. Another failure, money wasted, and you still need to talk to a lawyer. Lets make sure this doesn’t happen. Today we are covering personalities for AI This is design for AI a podcast to help define the space where Machine learning intersects with UX. Where we talk to experts and discuss topics around designing a better AI. music is The Pirate And The Dancer by Rolemusic Im your host Mark Bailey Lets get started Today we are discussing how to design your AI personality. We will cover the process step by step for what is important and what to avoid. Some people associate finding the right personality with something hippy or new age. This is not that. If you want the book answer, the personality is the distinctive tone, manner and style in which your app will communicate. It is defined by a set of attributes that shape how it will look, sound and feel. The right language, and tone that embodies your app and differentiates it from the competitors. Look, there is a good chance your app and company already have a personality. Your current web or app design already defines the personality of the company. Color choice, type choices, UI layout, documentation, errors all make up the brand. Basically, it’s the company personality that dictates the brand. So the next step is to use that personality, that up to now has been used for the brand, and to translate that over to training the AI. There are some companies that don’t have a personality right now. The reason being is a lot of companies might not have defined a personality up to this point is because of they’ve used a template for their site or app. There are a lot of templates for websites or using default frameworks for building the widgets for apps. There just isn’t a template for this yet in AI. So going to the trouble of creating a personality has to be done on a case by case basis still. Because the world does not need another Clippy. It was an avatar that tried to keep it light by telling jokes along with the help it gave. The problem was the brand for Microsoft Word is much more corporate which created anger at the unexpected behavior. Jokes or wacky interface quirks can only increase user’s interest or desire to explore in the application if it what they are expecting. Personality sells though, so it will pay for itself if you get it right. People can tell when a company has enthusiasm and passion for what they’re doing. The tide will turn soon enough where the AI will stick out like a sore thumb when it is bland. To follow best practices will earn you a spot in the middle of the pack. The problem being most user are not happy with their app as “not terrible”. If it’s important when hiring employees, why wouldn’t it be important when creating an AI personality? AI centered companies are already working in this. Google is hiring creatives to bring humor and storytelling to human-to-machine interactions, and Microsoft Cortana’s writing team includes a poet, a novelist, a playwright, and a former tv writer. Skills to build a personality can come from writers, designers, actors, comedians, playwrights, psychologists and novelists.  Not the normal job descriptions you would expect for tech companies. The integration of these skills into tech roles have sprung terms such as conversation designer, persona developer, and AI interaction designer. So now that we have established a the need lets talk about the creation process for a personality. If you want some long term planning here is some predictions. At some point in the future, companies will probably have many personalities to let people choose their preferred voice or body depending on the AI UI. Different personalities will become popular similar to material design from Google or metro from Microsoft. Which will lead to templatizing a personality similar to how wordpress templates exist now, and it’s only a matter of time before one company sues another for copying their AI personality similar to brand infringements today. Personally I am waiting for the days when enough UX research has been done, that we know which custom AI personality works best for interaction modality. So while a it might sound silly that the best way to get legal information from someone is if they are talking to a salty sailor, there is no way to know which personality can become associated with interaction modalities without creating one for your use case first. What should you not do The biggest question to avoid is usually, shouldn’t I just use my own personality? or the founders personality? There are a couple of problems with this. In true UX fashion remember, you are not your customer. The ability to create a company does not always translate over to good customer interaction for a variety of reasons. Another reason why that usually does not work is because you can’t measure your own personality. Most people only associate with their positive traits, unfortunately there are usually blind traits that can go along with them. So how do we find what personality would work best for the users? Well, we ask them. Poll people to select descriptive words. Use standardized list, like from Microsoft’s word association test work best, because it takes time to balance positive and negative words, and making sure all the areas are covered. I think the word associations are the easiest and fastest to move forward with but if that doesn’t sound good for you I have also heard of other people that have successfully used Myers-Briggs to describe character traits. I know it has been debunked because it over-simplies personality types. But that actually helps since it is simplifying choices that need to be made. Another way to gather the information is something called Spectrum. It was Five Factor Model created by Ari Zilnik It defines personality as a combination of: openness to experience, conscientiousness, extraversion, agreeableness, and neuroticism Be aware that answers usually skew towards positive so pay more attention to the negative feedback. Basically you are asking users to choose the words they associate with your brand, company, and app. The 5 different areas to measure are

• Awareness – How aware is the user of the company, product, needs for the product?
• Consideration – perception of quality, value of the product. misunderstand or can’t find features.
• preferences – How do features differentiate product from competators
• Action – Getting stuff done.
• Loyalty – Will the user want to use your app again?

Depending on the actions that the AI is created to help with sound, haptics, visuals, or AR/VR,  can be aspects of your interface for the user. If anything like this exists for the current interaction get feedback on that as well. Sound Associations can be done by comparisons to companies that have patented their sounds like Porsche or Harley Davidson. When talking to the users to get the word association get a really good sense of your customer’s personality, what are their goals, what stage are they at in their life, and most importantly who do they aspire to be. This will come into play later. The next step is to run the same word association tests with internal people, but from the aspect of aspiration. Where are the decision makers trying to take the product. This is perfect since this is what PMs and stakeholders are thinking about anyway. Now comes the comparison. How you are currently perceived vs how would you want to be seen in the future? Take stock of responses, how do they stack up to expectations. You shouldn’t expect the word associations to be exactly alike but they shouldn’t be too far apart either. If there is too much drift from the customers perception then there either needs to be frank discussion about where your company is for product excellence Or there needs to be a whole lot of work on the fundamentals. The reason you don’t want to reach too far, is that it comes off as untrustworthy. I mean, you are who you are. Also too far of a reach and the chance of getting it wrong starts to skyrocket. If you get it wrong this is a lot of work that is going to waste. Once you have the aspirational view of how you are perceived. Check the goals against what the customer’s goals. These should also be close. A good example is to go into a teen clothing store. The employees tend to be clones of the people from their ads. It’s not a coincidence, stores are choosing to mirror their target aspirational demographic for their customer interactions.

• So from that point of view what would your employees look like?
• How do your customers align with their peers?
• What motivates the people to do what they do?

Areas of personality that need to be defined include

• Professional vs casual
• Don’t want to take all personality out if you are going full professional.
• Also be aware that if you are going casual, informality changes for different groups, cultures so make sure to gather information all your markets.
• Humor level and type
• Do you use dry humor or silly humor?
• The best example I can come up with why this is a difficult question is: I want you to name 2 comedians with the same style humor?
• Generalist or a specialist
• Are you trying to reach a conversion quickly and effectively? Or is the whole bot experience crafted to engage long term as part of a larger creative campaign?
• Brief or long discussions
• Unless the destination is the personality you don’t want to slow down the interaction. Aim for minimal clutter and fuss.
• Understated vs extroverted
• Cautious vs go where others fear to tread
• Individual creativity vs group consensus
• Strong opinions vs easy going

Now none of the word associations need to be shown to the developers. Trying to paraphrase brand guideline, and it always gets reduced down to BS words like innovative and progressive. You will not have the brand guidelines next to you when writing dialogs. Developers won’t have them next to them when writing code. It needs to to be easier. If your organization was a famous person who would it be? Since you already have all the personality traits and aspirations, who does it describe. It can’t be Stephan fry or Barrack Obama. Those are not good choices. It is the equivalent of boiling everything down to the words innovative and progressive. You want to choose a different personality for at least the 5 main touch-points: Awareness, Consideration, preferences, Action, and Loyalty. Your app might be more specialized so your areas might differ depending on your needs. Now that you know who the AI should act like in different situations let’s talk about some of the things to avoid. First lets talk about humaness or better titled as, When to convey an bot is an bot. Most companies have a code of conduct of how employees interact with customers. I am surprised the same companies will many times not put the same thought into the personality of their AI that is the first interaction point with their company. Right now, conversational AIs are good enough to pass off for human. If you need an example, I’ll link to how Google wowed people with their 2018 IO demo of Google Duplex (https://www.theverge.com/2018/5/8/17332070/google-assistant-makes-phone-call-demo-duplex-io-2018) but the next week the articles tone changed quickly saying the voice was trying to trick people (https://www.wired.com/story/google-duplex-phone-calls-ai-future/) . Nothing changed in the demo, just people talking. Current culture was caught off gaurd with the quality of the humaness, and it is human nature to try to the label someone as tricky when you are caught off guard.  Sooner or later, with the current tech your AI is going to fall into the wrong side of the uncanny valley. So it is easier to not try to claim humanness when asked, but also don’t try to deny it. In the principles of Google Assistant’s personality – don’t shutdown conversation by denying humanness. Don’t lie and claim preferences. Using the artful dodge https://youtu.be/6Se8A-GNM2g?t=1470 The next hurdle is to make sure to take into account for Internationalization. Currently for websites type choices and layout carry over cultures so this can catch some people off guard. humor does not cross borders well, or even across regions. I know China has different formats for standup comedy depending which city you in. Cues for informality changes for different groups, cultures. An example of this is the sound your mouth makes when your brain is processing. In the US depending on region it can be “uhm” or “ahh” in China it is “nega” putting the wrong pause word for a region in a search to be more casual can put you on the wrong side of the uncanny valley. The third topic is situational awareness For example, how your AI should act in offline vs online How does your interaction change when it isn’t connected to the network. The level of interaction also depends on how many cognitive abilities the user has to devote to the interaction. If you can detect they are driving, your responses will probably be shorter. There is a lot of complexity and nuance to this It helps that the AI can detect more info What is the emotional context of how the person feels at that point

• how they feel right now
• What can you detect in voice and body language?
• What can you know from context of user journey?
• What do you know from user profile?

The last topic I want to cover is errors. When the server is down, humor makes the problem worse. People do not feel like they are taken seriously. They will lose trust in your AI since it is not acting appropriately. Instead of humor, try to empathize Acknowledging and validating an emotion is often enough to make customers feel understood and release negativity of the bad situation caused by an error. So once you have created and implemented your personality how do you know it is working? Let’s talk about testing personality success. You are trying to find out: Do decision-makers select products and services more or less? Currently there is a lot of counting tweets or Instagram pictures. I would recommend against that. They are hard to quantify because of the high noise. Ways that I would measure

• sentiment analysis with AI
• Measure the brand strength through qualitative and quantitative surveys
• AB testing choices are good to compare against the baseline. This is a good time to pull out the brand values. You can get word associations for the changed personality to see how it affects word choice.
• and of course, keep track of your analytics for unsure answers. Does the personality help the confidence level go up for what the AI is making decisions on from the info gathered from the user?

So that’s all we have for this episode, If you have questions or comments, use your phone to record a voice memo, then email it to podcast@designforai.com If you would like to see what I am up to, you can find me on Twitter at @DesignForAI Thank you again and remember, with how powerful AI is, lets design it to be usable for everyone