Human-centered AI: The Art Of Designing AI-based Digital Products

Introduction

A while before AI became a ubiquitous  buzzword, there were other similar terms  such as Machine Learning and Big Data that had gained some traction. As a designer working on various business projects over the last decade, I was exposed to these terms on a daily basis, but never thought they would ever have  a big impact on my design work.

A Nobel Prize Worthy Challenge

A few years ago, when I began managing the product design and mobile app development for Primate – a personalized diet delivery system – my team had to face a daunting task. We needed to build a feature that would create customized menus of 1-7 meals out of fifty total daily options, each available in five different serving sizes. Everything had to be based on users’ preferences, previous interactions, health details, etc.

As I studied potential design approaches on how to build good criteria for choosing one of the combinations amongst 62 quadrillion possibilities (yep!), I was frequently perplexed by the various facets of data analysis related to my design issues. Interestingly enough, there’s even an awarding-winning solution for challenges like this; Stigler's diet earned him a Nobel Prize.

After some initial struggles, I was able to successfully lead the product design of my first AI feature. Since then, not only have I been captivated by the boundless applications of artificial intelligence but also by how immensely underestimated a role design plays in unleashing its potential.

How do I understand AI as a Designer?

What’s not always clear for many product designers is that today's artificial intelligence is built on Machine Learning. Machine Learning is one of the algorithm types that forms the bedrock of most digital products. While most algorithms are rule-based, meaning developers craft code to give computers specific rules on how to process data and obtain results, ML uses a different approach altogether.

For instance, when you tap on the like button for a picture of an adorable cat, there is a rule-based algorithm that informs the computer (your computer if it's frontend or server if it's a backend algorithm) to check if you are logged in. If the answer is yes, then the animation of the heart appears and a new record in the database is created to indicate that you liked this image; whereas if you aren’t logged in, it will display the app’s login page for authentication. This is a clear rule for computers to follow.

How Machine Learning is Different from Rule-based Algorithms

Machine Learning algorithms are different in how developers define the rules and goals to follow for computers. Instead of explicitly stating how data must be processed, they are commanding computers to search for correlations and dependencies in large quantities of organized data.

This approach isn't very useful for replacing the way software reacts to tapping “like” button but it is a game-changer for extending it. If  the app is already aware that you appreciate a certain cute cat photo, it can assess how this depends on other details which it knows about you such as other pictures you like, the time of day when you use the app, who follows and is followed by you, along with other specific elements of those profiles. This process searches for patterns inside the data to curate content that you are most likely to enjoy.

Rather than having a set of predetermined rules to decide if you will like a particular photo or not, there is instead an intricate mathematical model that evaluates the probability based on all available data points and the relationships between them, without developers even knowing which aspects were decisive. With a level of complexity that surpasses human capabilities, these elements can determine the likelihood of someone liking fat black cats. For example, being a woman living in Boston who took photos of scrambled eggs on three consecutive Tuesdays and failed to update her iOS, could  for some reason drastically impact this probability. Machine Learning models not only search for correlations, but they also grow over time to refine their accuracy. This ability to "learn" is where the name comes from. 

Human Learning vs. Machine Learning

Once you start to follow various examples of Machine Learning in use, the mystery starts to wear off a bit. You begin to realize that the method of seeking patterns in data, guessing the results based on input data and increasing the sureness of accuracy  every time a guess is correct, is exactly how  us humans learn and make decisions!

We're impressed by Shazam's capability to recognize a song from merely a 10-second noisy sample. In fact, you can identify your favorite melody in the same manner. Your memory holds plenty of audio representations and there are certain patterns exclusive to this particular track that your brain unravels so that it knows “That Song” and not another one or someone crying out for help. It's difficult to determine the exact reasons why your brain came to this conclusion, as there are likely numerous micro-dependencies that contributed to the decision-making process which you aren't aware of.

An important part of Machine Learning algorithms are neural networks – models which build connections between all of the objects in a data structure and make a stronger connection each time they appear or become verified during the usage of a product. It's highly similar to how neurons in the human brain build a representation of our memory.

Deep Learning, a widely used term, refers to multi-layer neural networks with an excessive level of complexity (depth) – making it possible for complex dependencies in data to be discovered.

We’ve now reached a point in the history of computing when machines can accurately calculate data and recognize patterns at an unprecedented rate. Just as humans have the ability to recognize whether a photo depicts a dog or cat without needing to know every step they took to reach this conclusion, computers are capable of performing this same task. In order to emulate the way humans think, computers need a vast digital representation of all of our previous experiences. By providing labeled photos of cats and dogs, we can enable machines to make determinations much like we do – relying on prior knowledge of all of the cats and dogs we’ve encountered throughout our lives.

Furthermore, the digital representation of all of known reality is simplified into zeros and ones. Photos, videos, VR experiences, songs, this article – they’re all reduced to long chains of zeroes and ones and then calculated by processors on devices and servers to look for patterns and follow programmed rules. Isn't this mind-blowing?

The Biggest Question Mark in the Futuristic Vision of AI Abilities

As we analyze the parallels between Machine Learning models and human cognition, our vision of what's to come in the AI revolution becomes increasingly futuristic. Most engineers' outlook on this is confined by the current opportunities of today – such as generative AI (chatbots, image generators), which is astounding but still extremely restricted because of the amount of data and training needed to build successful solutions.

As designers, we can take advantage of both our boundless creativity and our limited knowledge about this complex topic. I’ve done this countless times recently and I’m happy to share the effects. As we begin to consider the limitless potential of AI (often referred to as General AI or even Super AI), one common obstacle will be agreeing upon a universal definition of consciousness. Currently, our understanding of what constitutes consciousness in humans remains unclear from a neurobiological standpoint. We can feel the presence of consciousness surrounding us – like a liberating emotion that allows us to make conscious decisions about our lives and reality. Despite this fact, we're still unable to define it. Until we can define it for humans, it is impossible to accurately gauge whether ArtificialIntelligence can ever attain a state of awareness or develop genuine emotions.

Nevertheless, the possibilities of Machine Learning applications inspired by human cognition are limitless, and they offer a grand opportunity for designers to use our creativity and imagination.

How to Consider AI Usage in Product Design

The Machine Learning Model for Designers

As you will get to know later in this article, designing product features based on Machine Learning instead of rule-based algorithms has serious implications for  the design process. But the first area impacted is product ideation and discovery. How should you consider custom AI usage in your product and how can you predict whether the problem you are solving is one in which ML will be useful for. It's almost impossible to do this and communicate clearly about it without first learning how to construct a Machine Learning model. Luckily as a designer, you can follow simplified patterns here. No coding, no math, just some theoretical concepts to keep in mind.

What all AI solutions have in common is the need for well structured data to look for patterns and dependencies on which the model can base its predictions.

What Features are Good for Potential AI Usage

• Personalized content or experience – Use your database to recognize how each user consumes your product and provide them with a made-to-measure experience instead of the same generic path for all.

• Predictions – If you have enough data about the past you can predict the future.

• Automation and augmentations – If you have data about what actions were performed in the right context in the past, you can augment them or automate them in the future.

• Recognition – Remember the favorite song example? Software can recognize language, sounds, and images almost as well as humans if you have enough data to teach your model.

• Anomalies or low occurrence events – If there is a pattern that sets one event apart from others you can train a model to recognize it without even knowing what is the specific difference. This is how anti-spam filters work.

As you probably can see by now, what most AI solutions have in common is  a need for data to look for dependencies. So how might this  apply to generative AI solutions like Midjourney or chatbots like GPT, you might ask.

Well, Midjourney was trained on thousands of images with descriptions – this is how the model learned what the specific pixel formations and colors are for a description such as 'fantasy', how they are similar or different from something like 'comic', and how being 'adorable' affects pixels when connected with 'cat'. Keep in mind that for the processor, every  image is just a mesh of colored squares and it looks for dependencies at this level.

But what about chatbots? You can look at them as machines that instead of thinking about your entire question are guessing letter by letter what the most appropriate next letter should be in an answer given the provided context based on millions of other letters in thousands of other contexts already analyzed by the model before. Natural language processing is undoubtedly a complex field, yet breaking it down into simpler steps can lead to a more simplified understanding.

You can look at any ML model this way:

”

It is predicting an output by referring the input to data structured in a specific model. In this way the effect of the prediction can be applied to the reference data model to improve the accuracy of future predictions.

Types of Machine Learning for Designers

There are various types of categorizations in Machine Learning and the deeper you look the more complicated the topic becomes. I think it’s enough for designers to know what the main types of learning are for building Machine Learning models before they start to consider AI in the ideation of new features or products. 

• Unsupervised Learning – Looks for patterns in data without any input. One example of a feature based on unsupervised Machine Learning is a product recommendation system. By using data such as user behavior, purchase history, and product attributes, unsupervised Machine Learning algorithms can analyze the relationships between different products and recommend personalized product options to individual users. By using cluster algorithms, the system can group users with similar preferences and recommend products that users with similar preferences have previously purchased.

• Supervised Learning – Looks for patterns with the help of labeled datasets. A good example is a spam filter for an email service. By using a dataset that includes labeled examples of spam and non-spam emails, the supervised learning algorithm can learn to identify patterns in the data that distinguish spam from legitimate messages. The algorithm can then use these patterns to predict whether a new incoming email is a spam or not. This type of learning can also be applied to image recognition – without describing photos of cats as ‘adorable,’ there’s no way AI will be able to know how being ‘adorable’ affects pixels.

• Reinforcement Learning – This is my personal favorite. It uses rewards and punishments to shape behavior. A robot that is designed to climb stairs can use reinforcement learning to learn how to navigate and traverse steps efficiently. It would start by taking random actions to climb the stairs, using sensors and cameras to detect the presence and location of steps. The reinforcement learning algorithm would then provide a reward signal to the robot based on how far it progressed up the stairs. If the robot made progress toward the top of the stairs, it would receive a positive reinforcement signal. If it stumbled or failed to make progress, it would receive a negative reinforcement signal. Through experimentation and trial-and-error, the robot would learn which actions lead to positive reinforcement and adjust its behavior accordingly. Eventually, the robot would learn the most effective way to climb stairs based on the performance of its actions. By the virtue of Machine Learning, robots can learn how various parameters such as rain or slope affect their optimal actions to ascend a flight of stairs. Imagine if instead of one robot mastering this task, there were thousands– each one contributing knowledge to a common model that would be shared by all. Here we go, now we’re blazing a trail into the autonomous vehicle revolution!

When AI Application is not Feasible or Applicable

Rule-based algorithms have consistently proven their value in certain situations where Machine Learning may struggle.

• A Lack of data – Without access to comprehensive and organized reference data, implementing any type of artificial intelligence is simply impossible.

• Predictability – When a consistent core experience is essential, regardless of the context or user input. Designing an offer calculator, CRM or CPQ software necessitates the use of precise formulas to process input data rather than artificial intelligence, as this may lead to unnecessary complexity and reduce predictability, consequently resulting in mistakes.

• The high cost of errors and exceptions – When the losses incurred by a mistake are exceedingly high and outweigh the benefits of any minor improvement in success rate. AI-managed prices can be a brilliant and groundbreaking concept for your e-commerce app, however, if you specialize in selling real estate properties it could spell disaster since one error in pricing may result in major losses.

• Transparency and explainability – When users, customers, and developers need to gain an in-depth understanding of what occurs within a feature. Remember the fat black  cat recommendation example? Sometimes AI is  unable to deliver a clear explanation for the reasons behind a prediction which  can be a problem.

• Low budget – Constructing your own Machine Learning models and AI features is often more expensive and carries a greater risk than opting for conventional solutions.

Human-Centered AI - The Role of a Designer

Artificial Intelligence has already revolutionized several industries, from Healthcare to Finance. By examining the development of AI, it is already evident that there are some inherent risks and dangers associated with utilizing Machine Learning in new parts of our lives. Consequently, Human-Centered AI was conceived. This innovative concept emphasizes creating Artificial Intelligence systems that not only comprehend human behavior but are able to modify and adapt to the needs of humans. The underlying premise is simple: technology should be used with an emphasis on enhancing the endeavors of humans instead of replacing or interrupting them. By making people a priority in AI design, we can ensure more meaningful and tailored interactions between machines and individuals alike.

In the sources at the end of this article, you’ll find examples of philosophy, politics, and various approaches to HCAI as presented  by the biggest tech companies.

Taking the time to read through all of the guidelines, learn about any potential risks, and examine cases where AI has already failed will drive home one key point:

”

There is an urgent need for a general designer's skillset and approach to conceptualize, discover, create new products and apply AI to different aspects of life.

In the next few paragraphs you'll find examples of the areas and processes that have changed the most when designing AI-based apps and features compared to conventional digital product design processes.

Should Designers do data?

By now you should be somewhat familiar with the general structure of a Machine Learning model, so you can appreciate how much the essence of a feature you want to build is affected by the data we choose as our reference and the way we label the data and the teaching model itself.

Imagine an app that serves as a career mentor helping graduates identify their professional trajectory or suggesting positions for job seekers. As a designer, you can focus on the traditional UX/UI layer and allow data scientists to figure out which factors should be used as reliable recommendations and which ones should be excluded, but you'll quickly come to realize that this is the part of the app that delivers its real value.

Over the past decade of my career, the design world has been debating whether designers should code. In my opinion, the right answer is that some of them should. The new eternal question is should designers do data. My answer remains unchanged – the expertise, empathy, and design approach that designer’s bring to the table are absolutely essential in the development of Human-Centered AI data models.

AI Can Replicate our own Biases

Let’s take a look at the idea of bias in AI. Since Machine Learning models are designed with human-labeled data and fed with human inputs, it’s no surprise that AI can replicate our own biases. This means that any bias found in the dataset used for Machine Learning will be reflected in the AI’s predictions.

If you research the topic you can easily find disturbing examples of features discriminating against minorities, multiplying stereotypes and even writing hateful and racist tweets (see: Microsoft's Tay). In all of these cases, developers designed their products without any intention of creating bias.

With the vastness and complexity of data that is necessary to construct and instruct models, it can be difficult to identify potential issues that might have ethical implications. To truly tackle this issue requires creative thinking – a design approach that incorporates creative workshops and could prove invaluable in solving this problem.

A New Perspective on Errors

Up until now, the majority of components and products in the digital realm have been crafted using rule-based algorithms, leading users to put their faith in these products without question. If your calorie calculator indicates that 2000 kcal is the ideal daily intake for you, you expect it's based on precise mathematical calculations derived from your height, weight, age, and activity level. So if two people have identical physical attributes and lifestyles, they'll receive exactly the same result.

When constructing AI features, our datasets need to be more nuanced and contextually based. For example, prior menu selections, mealtime parameters, local weather patterns, or the number of tomatoes consumed in a week might all have an impact on individual outcomes. This way, no two people will input identical data into this dynamic environment. Remember, a Machine Learning model is not infallible– instead, it continually presents us with an educated guess or prediction. This factor significantly shapes the manner in which we must construct user interactions and communication within our applications – particularly when targeting users with limited capabilities.

Crafting the optimal user experience necessitates an in-depth understanding of the various types of errors that can occur – False positives and false negatives, for instance, are significant considerations when using Machine Learning models or predictive analytics features. It's the dawn of a new area for  UX design processes tasked with developing plans  for tackling this kind of issue.

User Feedback is a Crucial Part of UX/UI Design

Since data structures affect the value that features deliver and users' expectations of accuracy evolve with AI solutions, UX/UI are becoming an important challenge for any Artificial Intelligence product.

Do you recall the example of a career mentor app? Suppose the Machine Learning model's recommendations are disappointing or confusing to the user. Or let's say our cute cat-spotting app provides us with a startling image of a fox out of nowhere. To ensure our model can learn and that the features we offer are meaningful to our users, it is absolutely essential to build clear, user-friendly methods of communication. This could be as simple as a tappable button with an obvious message like "This is not a cat!", but more involved for more serious applications.

Once a job recommendation has been made, how do we know if  it was accurate? Unfortunately, there isn't an easy metric to determine this and even if we find one, it will require waiting some time for the effects. It looks like the only available instant measurement of value lies in users’ emotions – does the suggestion feel fresh and engaging for them? Or are they left with a feeling that their time has been wasted?

To learn whether or not our recommendations have hit the mark, relying on their emotions may be our only answer. So how should we ask users about them?

The most recent UX/UI developments involve research and development into better frameworks for keeping up with the collecting of user feedback and looping it back into AI solutions as well as building user habits and expectations regarding AI-generated outcomes.

Explainability and Trust

Explaining the results of Machine Learning has become a critical challenge in data science, and is also having an impact on  digital product design.

Nowadays, Deep Learning solutions are advanced enough to process huge amounts of data points and connections that deliver accurate outcomes. In Finance they are being applied to accurately determine risks and make optimized decisions, while in  Healthcare they are being used to identify early diagnosis and to search for cures. These solutions have an immense amount of accountability connected to their usage.

This is the main reason why assumptions based on AI can be so disconcerting. How do we trust a mathematical model's prediction of cancer risk if we don't understand exactly what led it to this conclusion? The volume and intricacy of the data used in these models far surpasses our ability to comprehend it with any degree of success.

Explainability is a way to uncover the underlying logic and context of an algorithm, giving us insight into the decision-making process. This enhanced visibility helps us to understand why certain decisions were made and offers a deeper understanding of a model’s performance.

Uncovering meaningful metrics and devising a list of potential factors that may influence a model’s projections is part of the artistic angle in data science where designers can provide exceptional insight. Another area where designers can have an impact  – and is more UX-related – is how we communicate this information to users. If we know what key indicators were essential for recommending a career path for example, we must somehow make it understandable so that the user can interact and provide feedback. This is quite a daunting  but exciting challenge.

Summary

AI Workshops for Designers

I had the great pleasure to run several workshops about designing AI-based digital products during the 2022 DesignWays Conference in Krakow, Poland. Along with Ania Wojcieszczak, we built several frameworks for groups of designers to explore the potential of using Machine Learning and consider various aspects of product design in the features they developed. We chose a very specific scenario and environment and proposed that our groups  look for innovations in the Polish Railway, which is somewhat of a Polish icon of backwardness and inefficiency (not always justified). We were expecting ideas like optimizing time schedules, personalized journey recommendations, and dynamic pricing, however the results  were beyond our  expectations and proved how implicitly human-centered  most designers’ approach actually is. One of the ideas presented by one group of designers was a solution to reduce suicides on railways. This is a serious problem not only in Poland and with enough data about previous accidents, it has huge potential for Machine-Learning usage to, for example, warn drivers when various conditions increase the probability of a suicide attempt. 

All of this was discovered  with just a few details about the model,  a short introduction to Machine Learning and a two hour workshop.   

Untapped Potential

AI is a powerful tool and its use brings numerous opportunities for digitizing layers of our lives. However, we must remember that digitization doesn't necessarily mean better – and with these new exciting opportunities come certain risks and challenges. By understanding the outlook on Artificial Intelligence, it’s clear that specific roles and expertise are essential for managing all of the associated risks and obstacles. Creativity, empathy, curiosity, and a human-centered approach are traits that most designers naturally possess and have been thoroughly educated in. By combining these traits with a fundamental comprehension of Machine Learning, an open-mindedness to data exploration, and rethinking the conventional UX/UI design process, we can unlock the potential for another wave of innovation in the digital era. We can create AI-powered products that will successfully and sustainably serve humans and the earth well. As Victor Papanek said, "the only important thing about design is how it relates to people."

Further Reading 

My curiosity about the connection between AI and product design has led me to explore numerous Machine-Learning related articles. Yet, what I discovered from these readings was that most of them focused on coding with Python or constructing data structures when discussing anything concerning "design." This shined a light on the dearth of information addressing how Artificial Intelligence affects the very fundamentals of product design.

Although books on popular science and articles concerning social, political, and cultural elements of AI are interesting to read, they lack the necessary detail needed to effectively use them when creating a specific product.

I have compiled this list of my top 5 recommendations, starting with those sources that exclusively focus on designing AI solutions and Human-Centered AI, and then more general ones which are excellent resources for designers looking to learn more about Machine Learning.

• Big Data, Big Design: Why Designers Should Care about Artificial Intelligence, by Helen Armstrong https://helenarmstrong.info/

• The People + AI Guidebook by Google https://pair.withgoogle.com/guidebook/

• IBM Human Centered AI projects and publications https://research.ibm.com/topics/human-centered-ai

• Stanfords University AI Index Report https://aiindex.stanford.edu/report/

• Hello World: Being Human in the Age of Algorithms, by Hannah Fry https://hannahfry.co.uk/