AI product design: Identifying skills gaps and how to close them


TLDV: It’s not uncommon for AI product designers to encounter challenges in their work, especially when it comes to understanding the complex technologies behind AI products. This can make it difficult for them to communicate effectively with technical experts, leading to further complications. Additionally, keeping up with the ever-evolving design patterns for human-AI interaction can be challenging. In this article, I’ve compiled a list of resources that can enhance designers’ knowledge and skills in these areas so they can tackle their work with confidence and ease.

Rethinking design approaches

Advances in Artificial Intelligence have produced exciting opportunities for human-computer interaction. From identifying your cat in the photos to enabling autonomous driving, AI offers numerous promising new possibilities in user experiences. It facilitates forms of interaction that were previously unimaginable.

Despite the vast potential of AI, incorporating its capabilities into design methodologies is not a straightforward task. According to recent research, designers are struggling with the complexities of envisioning and prototyping AI systems.

Why is that? Traditional UX and HCI design techniques, such as sketching and prototyping, might not be sufficient for addressing the unintended consequences of AI in product design.

“HCI professionals cannot easily sketch the numerous of ways an AI system might adapt to different users in different contexts” was stated by Graham Dove and his colleagues during the conference on Human Factors in Computing Systems. “…. And nor can they easily prototype the types of inference errors a not-yet developed AI system might make” could I add, citing Philip van Allen.

Problems: Mapping AI design challenges in design process

Qian Yang and her team at Carnegie Mellon University studied the challenges HCI researchers and professionals face while working with AI. They structured these challenges and related research papers within the widely recognized double diamond design process model.

This image outlines challenges in AI system design within the Double Diamond model, divided into four phases: Discover, Define, Develop, and Deliver. Each phase highlights specific issues such as articulating AI capabilities, prototyping behaviour, designing interactions, and managing AI performance. Challenges also include foreseeing AI effects, avoiding the Uncanny Valley, and accountability for AI errors, emphasizing the complex process from initial problem identification to final solution
Diagram created by the author

Here are my takeaways from this paper, broken down by the design process steps:

  1. Discover (divergent thinking stage)
This image concentrates on the “Discover” phase from a larger framework detailing challenges in AI system design. It highlights key difficulties such as: Articulating what AI can and cannot do. The technical feasibility of a design idea being highly dependent on data. Lacking knowledge on how to purposefully use AI in specific design problems. Sketching divergent AI interactions
Diagram created by the author
  • Designers struggle to understand the limitations and capabilities of AI, which hinders their brainstorming and sketching processes.
  • The technical feasibility of a design idea depends on having access to sufficient, diverse, and high-quality data for effectively training the AI models.
  • Even when understanding how AI works, designers find it challenging to ideate many possible new interactions and novel experiences with much fluidity.
  • Choosing the right AI technique for a design problem requires a deep understanding of AI technologies, which can be challenging.

2. Define (convergent thinking stage)

This image focuses on the “Define” phase in the design of AI systems, as presented in a larger framework. It highlights two specific challenges: The difficulty in fast prototyping AI system behavior. The challenge of foreseeing the potential effects of AI.
Diagram created by the author
  • UX design involves rapid prototyping to evaluate human impact and make improvements. Rapid prototyping of AI products presents challenges in predicting the user experience.
  • Scott Klemmer suggests creating Wizard of Oz systems or rule-based simulators as an early-stage interactive AI prototype. Josh Lovejoy and Jess Holbrook explore it further in their article. While a valid option, this approach fails to address UX issues resulting from AI inference errors.
  • The second approach Qian Yang and colleagues proposed is to deploy a functioning AI system among real users to understand its intended and unintended consequence fully. However, teams cannot realize the value of rapid and iterative prototyping since the process takes too much time and does not allow for early failures.

Develop (divergent thinking stage)

This image concentrates on the “Develop” phase of AI system design, detailing several key challenges: Designing fuzzy, open-ended interactions. Explaining AI behaviors to users. Designing shared control between the AI and users. Designing interactions that constantly improve AI performance. Communicating AI system evolution over time to users.
Diagram created by the author
  • AI technical experts are a valuable but often scarce resource for many UX design teams. Some designers find it challenging to work effectively with AI engineers due to a lack of shared workflow, boundary objects, or a common language to facilitate collaboration.
  • Fuzzy, open-ended interactions are complicated to design. They introduce a high level of complexity because they allow users to express themselves in various ways.
  • Understanding the concepts and terminology associated with AI can be challenging, making it difficult to effectively communicate how the AI system works and why it behaves the way it does.
  • How do we design shared control between the AI and the users if three in five (61 percent) are wary about trusting AI systems? Users may not trust the AI to make decisions on their behalf or may be skeptical of relinquishing control. Building trust in the AI system’s abilities and ensuring transparency in its decision-making process is essential but can be difficult.
  • Designing interactions for AI systems is challenging as their environment can change rapidly, requiring adaptable interactions.
  • AI systems are dynamic and constantly evolving entities. Communicating this ongoing evolution to users in a timely, relevant, and understandable way can be complicated.

Deliver (convergent thinking stage)

This image concentrates on the “Deliver” phase in the design of AI systems, focusing on several critical challenges: Anticipating or mitigating unpredictable AI behaviours. Avoiding the Uncanny Valley to ensure the AI is not perceived as creepy. Determining accountability for AI errors.
Diagram created by the author
  • AI systems can evolve and adapt as they interact with their environment or receive feedback. This evolution can lead to changes in behavior that are difficult to predict or control. Some AI models can be black boxes, meaning it’s challenging to understand how they arrive at their decisions. As a result, it may be challenging to anticipate/mitigate unpredictable AI behaviors
  • The goal is to create realistic and relatable AI without being unsettling or creepy. Designers must navigate this delicate balance when designing AI, especially those with human-like characteristics such as voice, facial expressions, or behaviors. If the AI is too human-like but not quite perfect, it can fall into the Uncanny Valley, which may result in adverse reactions from users.
  • AI errors can happen due to biased or flawed data. It’s unclear who should be held responsible as issues can arise from data collection, preprocessing, model development or implementation.

Taking action: Resolving identified challenges

I would classify all the problems mentioned into two main categories:

  • Difficulties, related to lack of understanding of technologies behind AI systems
  • Difficulties, related to lack of understanding design patterns for AI systems

Difficulties, related to lack of understanding of technologies behind AI systems

Just getting more UXers assigned to projects that use ML won’t be enough. It’ll be essential that they understand certain core ML concepts, unpack preconceptions about AI and its capabilities, and align around best-practices for building and maintaining trust.
Josh Lovejoy

There is still significant debate about which types of AI knowledge are relevant for UX design. However, it is becoming increasingly agreed upon that UX designers require some technical expertise in AI to work effectively with it.

Most AI courses assume prior knowledge of statistics, probability, linear algebra, calculus, and programming. Without this background, understanding many AI concepts can be challenging.

You don’t need an advanced understanding of AI, but comfort with math and computer science is essential. If any of these subjects make you uncomfortable, consider taking one of these highly-rated courses.

Probability: Fat Chance: Probability from the Ground Up from Harvard

Statistics: Fundamentals of Statistics from MIT

Linear Algebra: Linear Algebra 18.06 from MIT

Calculus: Single Variable Calculus and Multivariable Calculus from MIT

Programming: Learn Python from Codecademy, Google or University of Michigan. Personally, I prefer the course “Python for Everybody” by the University of Michigan. While it is longer, it provides more detailed explanations.

Having some familiarity with each topic will provide a great foundation for taking one of these courses:

1) AI For Everyone (6h, 49.99/month) – the best non-technical introduction to AI, taught by Andrew Ng, creator of the renowned Stanford Machine Learning class

2)Professional Certificate in Computer Science for Artificial Intelligence(5 months, 432€) – a two-part professional certificate from edX that tracks Harvard’s CS50 and CS50AI courses, allowing learners without the prerequisite CS knowledge to break into AI.

3) “AI Foundations for Everyone(40h, 49.99/month) a specialization offered by IBM, which is recognized as a revolutionary leader in emerging technologies through Coursera. The specialization includes three courses:

4)Elements of AI (30–60h, free) – a course by the University of Helsinki and MinnaLearn that explains what is possible (and not possible) with AI and how it affects our lives – with no complicated math or programming required.

Would you happen to know of any other helpful resources? If so, please feel free to drop them in the comments.

Difficulties, related to lack of understanding design patterns for AI systems

There is a growing need for design skills specific to human-AI interaction, but relevant courses are limited. Fortunately, there are numerous designers and researchers who are generously sharing their knowledge and expertise. Here are the ones I follow:

If there’s anyone I missed, please don’t hesitate to share their name in the comments along with any helpful resources they created. Let’s work together to expand our knowledge and help each other grow.

Final thoughts

Artificial Intelligence shapes how we think, feel and behave. It drives the decisions that define our future. We have the responsibility to use this potential for humane technology. Building an AI based on our diverse values and needs requires thoughtful design.

Josh Lovejoy and Jess Holbrook

Artificial intelligence is increasingly being integrated into various digital products and services. User interaction with AI will be a critical factor in determining the success and adoption of these products. However, if we fail to define and adopt new interaction patterns and technologies – and instead rely on outdated heuristics and a limited understanding of AI – we risk hindering innovation.

By combining persistence and creativity, designers can tap into AI’s full potential and pave the way for a better future.

Thank you for reading, and please share additional resources in the comments to help us collectively build this new paradigm!

AI product design: Identifying skills gaps and how to close them was originally published in UX Collective on Medium, where people are continuing the conversation by highlighting and responding to this story.






Leave a Reply

Your email address will not be published. Required fields are marked *