In the past when UX designers wanted to test the new interfaces they would either test it on a paper mock-up, a paper replication of the upcoming interface with no interactive part, or with the method called “Wizard-of-Oz”, where a UX designer is the one acting what the interface will do. But there is a big problem when we start making interfaces with AI. AIs do not to behave the way we want them to, or the way we expect them to. It is difficult to simulate an AI as a human because we don’t quite understand how it works under the hood. Paper prototyping is completely out of the question as the interaction is not straightforward. And “Wizard-of-Oz” is a strange approach since it would require a human to act the AI, technically programmed to act as a human.
This problem the topic of the CHI 2023 article “Designing Responsible AI: Adaptations of UX Practice to Meet Responsible AI Challenges”. The paper follows a big Google team of both UI designers and AI researchers in the process of developing an AI application. The research team worked with the UI and AI team for 6 months to get the best overview of the process of designing for AI applications. This as a user study conducted as a “fly on the wall” way where the testers follow the test subjects doing their regular business and occasionally asks questions. Along the way they also did some interviews to understand their process.
The designers found it difficult to incorporate AI in the process since they had to find good ways to test it. Their goal is to anticipate suffering and harms that AI technology could bring to end-users, society, and the public. But this is a rather vague goal that is hard to test. And with AI it is hard to predict.
The most important thing when making new technology is that it, under no circumstances, must harm the user. The harm can be untimely errors without a way to recover or go back. But when we work with AI, this can be more difficult since we cannot test the edge cases like we would do in normal UX design. And users and designers have a tendency to overestimate the power of AI. Modern AI tools can be a quite phenomenal and therefore people tend to look to it as some kind of divine answer to all their questions.
The researchers followed the designers as they worked on designing a new AI tool. There were also a couple of Responsible AI researchers in the group. The difference between Responsible AI researchers and UI designers is mainly how much they know about AI and how to incorporate it appropriately into a tool. They could also feel that the other developers would ask the Responsible AI researchers more for advice than the UI designers on AI. This was not so surprising as this is of course more their domain. However, the article states that it is supposed to be different.
AIs are new and upcoming and there are specific researchers who work on Responsible AI. However, the papers message is that all UX designers should be concerned with this. UI and AI are not supposed to be two different things. They need to be incorporated from the beginning and all the way into the user studies. There is a lot of hidden work in the process of designing new tools. We need to make sure that the users can see what is going on — it needs to be transparent. Users need to feel safe in what is going on. And when the AI fails, it needs to fail gracefully. But this is not something that is added on as an extra feature when the rest is done. It needs to be a part of the planning and the first testing.
In the end, this means that when creating Human-Centered AI it needs to be, well, human-centered. But this can only be done if we reevaluate the entire design process. One of the important parts are also that the UI designers gain knowledge and confidence within the AI regime. It is supposed to be their domain.
Paper citation:
- Qiaosi Wang, Michael Madaip, Shaun Kane, Shivani Kapania, Michael Terry, Lauren Wilcox.Designing Responsible AI: Adaptations of UX Practice to Meet Responsible AI Challenges. In Proceedings of the ACM Conference on Human Factors in Computing Systems, pp. 249:1–249:16. https://doi.org/10.1145/3544548.3581278
