Design for AI

Episode 12 In this episode we look at what makes people feel creeped out and how that translates over into AI software. We cover good design, interaction, lesson from horror houses, and the uncanny valley.

Music: The Pirate And The Dancer by Rolemusic

Transcripts Has this ever happened to you? You have created an enterprise app for and the main selling point is the AI integration. You’ve worked hard to make the software personalized based on the data you have, but then the deal starts to fall apart. They tried it out with their employees and they are getting nervous using it. Apparently they are calling the Avatar creepy and there is just something they can’t put their finger on about how the software makes you feel, but they don’t like it. Let’s make sure this doesn’t happen. This podcast is called design for AI It is here 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 by Rolemusic Im your host Mark Bailey Lets get started music Today we will be diving into what makes AI creepy for people and how to design it out of your product. Creepy is defiantly different than the fear. We have covered fear of AI in previous episodes and if your users are getting creeped out it is different. While it sounds similar, the way to tell the difference is if there is a fear of something you know to run away or not use it. But if something is creepy… it might be dangerous but you’re not sure it is… there’s an ambivalence. Basically, someone is telling the users brain that the software is outside the accepted social norms. If it was a person it would be standing too close, or staring, say – we become suspicious of their intentions. Someone can be completely familiar with machine learning, use it every day and still get creeped out by badly designed AI. For this episode We will be looking into current research into the psychology, feedback from previous software and even what causes horror houses to be creepy. The pinnacle psychology paper on creepiness is called “On the nature of creepiness” (https://www.sciencedirect.com/science/article/abs/pii/S0732118X16300320?via%3Dihub) It says there are 4 driving factors of creepiness

• They make us fearful or anxious
• Discern if there is in fact something to fear or not from the person in question.
• Creepiness is seen as part of the personality of the individual rather than just their behavior
• We think they may have a sexual interest in us.

Now before writing off the last item as something that doesn’t cross over from psychology to software design: know that in surveys, men are seen as creepy more than women, so using a female voice for say Siri, Cortana, Alexa, and just about every AI voice system out there can lower the creepiness level for women using your product. For all the other driving factors, it makes sense that AI, can put people on edge. This is something that is new, people are not familiar with. The machine learning model is trying to do something independently of the user and they are not sure if they want to trust it or not. To cover all the things that are reported to add to a creepy feeling, I’m going to break it up into two different groups: Good design, and good interaction. Good design Visual design When people are creeped out by a person, contributing features to that person are that they could have greasy hair, a peculiar smile, bulging eyes, long fingers,  unkempt hair, very pale skin, bags under their eyes, dressed oddly, wearing dirty clothes. In computer terms if ever there was evidence that you need a well designed homepage, landing page, or first impression for your product this is it. Research shows (https://people.ok.ubc.ca/stporter/Publications_files/Is%20the%20Face%20a%20Window%20-%20final.pdf) that your brain will make judgements about trustworthiness within 39 milliseconds of seeing a face. As soon as something captures our attention as being abnormal, we start to deconstruct the face, then from there, then we deconstruct the person.The same is true for using your AI software. If your software is well designed then you have not blown your initial impression. This is important because of the “halo effect,”. Basically Attractive people were deemed to be trustworthy, whether they were Nobel Laureates or criminals. So what do I mean by good design? Well I’ll give the counter example. If the app was created and just throws all the data up on the screen without a thought to how it would look, what are the chances that when you use it, it might raise your suspicions that they could have forgotten something in their haste and your data could be vulnerable. Basically details are important. Pay attention to them. The more details you can make sure are right for the user for the initial impression the better. Sound usage The next area for good design is language use. People were looking for signs of kindness or aggressiveness in the faces of those they were evaluating. The same is true of software. Aggressive or abrupt language will put people on edge. So even more so than normal software, AI already has a reputation from movies so it needs to explain what is happening without being abrupt or using euphemisms as much as possible. Another thing that creeps people out is a person who stands too close to your friend, or uses overly friendly language. The perfect example is the creepy salesperson who will instantly act like your friend. If you product includes any sales make sure the language in the product is toned down, more professional. This goes against the current trend of having the language be overtly friendly so this could be one of the causes of where AI creepiness comes from. In the same category as language, pay attention to the nonverbal cues. Don’t mimic non-word voice nonverbal cues, such as hand movements or body language. While they are fundamental to smooth human interaction; this is starting to get into the uncanny valley which we will cover soon but is so easy to get wrong. Even if you get it right, it makes people suspicious and literally get the chills (http://pss.sagepub.com/content/early/2012/05/18/0956797611434535) A good example of this was duplex, a speech system from Google that was demoed at 2018 IO conference(https://www.wired.com/story/google-duplex-phone-calls-ai-future/). It convinced unknowing people it was human because of the amazing accuracy of the speech engine including pauses, sighs, and uhms. But, in all the reviews of the system, people said it was creepy because the reasons people were saying it was creepy was because it sighed and said hmm. I’ve mentioned this in previously, the technology can get ahead of technological understanding, and when it does, things look like non-understandable magic. And people are afraid of things they don’t understand. Another non-verbal cue that creeps people out is the person laughing at unpredictable times or displaying inappropriate emotion. I have covered working in humor into your AI personality, and this is a perfect example. If it is done wrong it comes across as creepy. The inclusion of emotion in machine learning is fraught with obstacles so tread slowly and verify changes. Basically user test after each addition and regression test for compatibility of each task to make sure the emotional tone stays consistent through the entire user journey. The last thing that crosses over from the research into design is that the most frequently mentioned creepy hobbies involved collecting things, and most likely your product is collecting a lot of data about them. So, my recommendation, is don’t talk about all the info you are collecting. Now, I am not saying to lie, or hide the fact that you are collecting data, because that would be worse. There just doesn’t need to be a feature that your app beeps every time it learns another factoid about the user. I’ve seen this from very technical products. They want to sell the technology, but this is not the way to do it. Good interaction Next we will be covering good interaction. There are things that make people creepy when someone interacts with them. The biggest creepy warning that was reported that comes from interaction with a person is when they make it nearly impossible to leave the conversation without appearing rude; or that the person relentlessly steered the conversation towards one topic. So make sure your product has a way to talk to a human. Too many people see AI as a way to totally remove call centers from your product. You will still need user support even with a perfect machine learning model. In this same vein, don’t have conversational as only interaction interface. It doesn’t work in office or in the public, so forcing users to interact in that way will put them on edge. This also means that you don’t want your app to keep circling around to the same script until the user agrees. This point affects sales apps the most, If your product is only for sales then make sure to say goodbye when they drop out of the sales funnel. If your product has other functionality then return to where the user was previously in the interface so they don’t stay in a sales funnel. The next most creepy thing is asking for details of about personal life. Too many questions are a problem. Too personal of questions are a problem. Just think about the flashlight app that on your phone that is requesting access to your address book. You will need to review the questions you are asking the user. Is it something you already know from somewhere else? Is it something that could get from inference from putting the answer from two other questions you asked? And the last interaction faux pax is talking about personal life too much. This affects startups a lot. You are proud of your product and company. But don’t force your story on your users. If they ask, or look for it that is fine but keep personal details to its own area. Interaction lessons from horror houses Next let’s take some tips from horror houses. They increase creepiness through introducing stillness before large movements. To avoid this for your product pace the level of interaction. Try to keep questions and actions balanced through the entire the user journey. If you are using a scripted conversational UI, each block of text should be as similar of size as possible to keep the pacing the same. The next tip from horror houses is that they use sound for distraction to cover motion. Basically if you hear a scream from one direction then you won’t notice movement from another direction so after the sound is over, suddenly there is a zombie next to you. The lesson to avoid this problem is to make sure sound and motion line up so the user knows what movements and sounds are associated with which actions. Create a hierarchy of user actions so if there are multiple actions happening at the same time you can prioritize sound and movements so the user is not overloaded. Another lesson from horror houses are to use sudden unexpected changes, movements to increase creepiness. Basically something that you are not expecting to move will move. For an example think of a lot of mannequins scattered around a room so as you enter the room the ones close to you are obviously mannequins but one farther back is a person disguised as a mannequin that can jump up as soon as your back is turned. The lesson to avoid this is to make sure to use a consistent design language. For conversational UI’s this means consistent trigger words that follow whatever platform you are using, or creating your own if you are creating your own platform. For consistent visual design, a design system will help. Using consistent colors, language, shapes, and flow will all lower the users suspicion levels. Uncanny valley The uncanny valley refers to the idea that human react favorably to humanoid figures and interaction until a breaking point where they become too human. At that point, the small differences between the human and the inhuman – maybe an awkward gait, an inability to use appropriate eye contact or speech patterns – become more noticeable because everything else is right to the point of causing discomfort and creepiness. The idea originated with Japanese roboticist Masahiro Mori’s 1970 essay anticipating the challenges robot-makers would face. This is basically the reason for for cartoon avatars, cartoon looking robots, and why video game errors seem so creepy. The most recent example of this is what I have talked about earlier, with Google’s duplex. It was close, but not 100% human voice so it was squarely in the the uncanny valley. Of course the way to avoid this is to not go into the valley. Like most others products you can use cartoony levels of detail, animals instead of people, or not use avatars at all. It is also important to let the user know the limitations. If the user is expecting perfection the uncanny valley is wider then if they know where the limits are. Conclusion Putting it together To put this all together, think of it like this: creepiness is your brains way of detecting danger. Anything out of the ordinary, or unexpected your brain is going to warn you about. We evolved to err on the side of detecting threats in such ambiguous situations. So use human centered design to avoid problems. Starting with the need. A lot of creepiness comes from trying to solve your need instead of the customers need, so make sure that you are trying to solve the customers problem. Just using the machine learning model to sell them stuff or collecting all their data is going to trigger warning bells in the user’s head. Next map out the journey for how the user expects to solve their problem. Test the user journey to make sure. This is to keep down the unexpected turns in the journey. It is just good design to keep the user updated on where they are in the journey, where else they can go, and what they can do at the point in the journey. In normal software if you don’t do this with normal software it will create confusing software. If the user knows there is machine learning involved with the software they will think the the AI is in the drivers seat. The lack of control leads to the creepy feeling. To keep the user in control let them decide on the info you collect. Ask them if they want better interaction through data collection. Ask if they want personalized ads. If your product requires data to be collected to function correctly it is better to tell the customer that the product can’t continue than to try to collect data without warning the customer. User testing We have covered a lot of things to check for to remove creepiness from your product. How do you know it works? Of course do user testing. Besides the types of user testing I’ve mentioned already there is a very quantitative way to measure it- people feel cold when creeped out. (https://www.ncbi.nlm.nih.gov/pubmed/18947346) So if nothing else work just get a baseline of what your users think the temperature is. Then measure the new version for improvements. and we are going to need to end on that note, Unfortunately, that’s all the time we have for this episode, but I would love to hear back from you on how you were able to avoid creepiness in your products. Use your phone to record a voice memo, then email it to podcast@designforai.com The question I want to know the answer to is what ways you have heard people say AI or machine learning was creepy? That is also an awesome way to let me know what you like and would like to hear more of, or If you have questions or comments record a message for those too. 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 Thank you