AI: designing for unexpected
It’s well known and documented that generative AI can at times produce hallucinations, inconsistent results and misleading information. AI is probabilistic, it cannot provide 100% accurate outputs. As designers how do we design AI-infused products with the knowledge of inconsistencies will happen.
Predictable design
If you imagine a blog post on a web page, as designers we know that the blog post will most likely contain a title, headings, paragraphs, graphics and some meta data about the post. We may not know the exact content that goes into those areas, but we could design something that is predictable, repeatable and meets user’s expectations of what and how they view a blog post.
In generative AI, we lose this control over this predictable UI. We won’t know what users will see, or what options they will have.
The value of AI
When we talk about AI agents and generative AI, the value of this technology is perhaps not in it’s consistent, meticulous approach to a task. Instead in it’s ability to handle tasks autonomously and adapt to a situation over time. As designers we need to understand two things:
- what are the limitations?
- what are the minimum requirements of acceptability?
A third question, and one that I’ve faced recently, is in which context and under which circumstances is it ok to have an imperfect outcome?
Examples
Let’s take a look at an example of where it may be ok for an imperfect outcome and when it’s not.
A common example of narrow AI is personalised recommendations for music and films, imagine your favourite streaming platform where it suggests something you may like. At best it suggests something that is relevant and that you want to engage with, at worse it’s slightly annoying as it doesn’t suggest something relevant and it takes slightly longer to find the thing you’re looking for. But overall no disasters or long term impacts from that experience.
Another example in which AI is making significant progress in is healthcare. If an AI tool could make an inaccurate diagnosis about a test result, and without the proper checks in place, this could result in serious consequences.
Two different situations, where you’d want to consider measures that account for errors from the beginning of the design process.
A new focus
Jakob Nielsen summarised it well that:
Instead of designing the specific UI, you will be designing rules and heuristics for the AI to employ when it decides what to show to the user.
We have this interesting separation between the designer and the user, in that new void AI sits between. We are telling AI about the rules and heuristics to follow, giving direction, rather than certainty of the final output.
Nielsen goes on to say that this is not much different to when responsive design was introduced into web design. We moved away from a fixed, one layout, design to something that was fluid and flexible where sizing of item changes, things hide and show and layouts change. We are just expanding on that further with the introduction of AI
Relearning the fundamentals
Tobias van Schneider recently posted that he believes that we will see a growth in the teaching of fundamental design principles again. With a focus on the why of design decisions, such as the psychology, history, cognitive perception and intentional problem solving.
His post wasn’t specifically aimed at designing AI products, more on the point that as the barrier to produce design artefacts reduces, there will be more people who can “design” but not understand the underlying complexity or understanding of what makes a good design.
But I feel that these skills around psychology, problem solving, history of design will also be relevant to how we design the rules and heruistics for the AI to employ.
