Intelligence on Tap:
The prospect of using Machine Learning or Artificial Intelligence as a “design material” is an exciting prospect, one with potential that we likely can’t even anticipate fully. However, I think the challenges that the author brings up (and still others he did not) need to be addressed before ML/AI can/should be “on tap.” These challenges are: Designing for transparency, Designing for opacity, Designing for unpredictability, Designing for learning, Designing for evolution, and Designing for shared control. I think, especially that it is critical that users should always be informed of the use of these systems and the unpredictable nature they embody. Also, I think it’s worth mentioning that no matter how “intelligent” these systems may be, they should, on principle, not be able to override the authority of its user. This would make the creator of the system liable for any and all havoc caused by the machine.
Challenges for Working with Machine Learning as a Design Material:
“We did not see research investigating issues such as the impact of false positive and false negative responses from agents, or the need to collect ground truth labels, which might negatively impact UX.” This is something that I had been thinking about when reading the previous article: It sounds so grand when people talk about the power and efficiency provided by machine learning associated with big data powerhouses like Amazon and Google, but what about when it is accidentally or intentionally trained poorly? I think the chatbot experiments proved why this is a huge problem. People may lie to the machine, which has no way of distinguishing true from false statements provided by its user. It’s possible that there may be a way to train the machines to account for this in the future, but I assume that would mean teaching it to lie itself, to understand lying, which would be incredibly problematic and would destroy any trust in the machine that users could build up over time.
Machines Learning Culture:
For this article, I wanted to comment on the part regarding “normalcy.” The Turing Normalizing Machine is designed to somehow identify what makes people “normal,” and the creators hope to decode the mystery of “what society deems ‘normal.'” But here is the problem with this goal: It will never be truly solved. Each person has their own beliefs and biases of what normal is, which in turn affects what they consider to be normal appearances. In consequence, the machine will be getting flawed and contradictory data. Sure, they can construct and image of what the machine thinks is the most “normal” appearance, but even if it could sample the data of every person alive, it couldn’t choose a form that satisfied everyone’s “normal.” In other words, people may see the machine’s aggregate person and think it looks weird, and they would be neither right nor wrong, as what is normal to them is different to people of different locations, beliefs, age, gender, etc. All this to say: It’s a fun little project, but will never “decode the mystery,” and has no real, practical value other than ironically pointing out all the differences in perspectives about “normal” that prevent it from being ever universally accepted.