Summary
Heer’s focus in this paper is on refocusing AI and machine learning into everyday interactions that assist users in their work rather than trying to replace users. He reiterates many times in the introduction that humans should remain in control while the AI assists them in completing the task, and even brings up the recent Boeing automation mishaps as an example of why human-in-the-loop is so essential to future developments. The author then describes several tools in data formatting, data visualization, and natural language translation that use AI to assist the user by suggesting actions based on their interactions with data, as well as domain-specific languages (DSLs) that can quickly perform actions through code. The review of his work shows that users want more control, not less, and that these tools increase productivity while allowing the user to ultimately make all of the decisions.
Personal Reflection
I enjoyed this paper as an exploration of various ways people can employ semantic interaction in interfaces to boost productivity. Furthermore, the explorations in how users can do this without giving up control was remarkable. I hadn’t realized that the basic idea behind autocorrect/autocomplete could apply in so many different ways in these domains. However, I did notice that the author mentioned that in certain cases there were too many options for what to do next. I wonder how much ethnographic research needs to go into determining each action that’s likely (or even possible) in each case and what overhead the AI puts on the system.
I also wonder how these interfaces will shape work in the future. Will humans adapt to these interfaces and essentially create new routines and processes in their work? As autocomplete/correct often creates errors, will we have to adapt to new kinds of errors in these interfaces? At what point does this kind of interaction become a hindrance? I know that, despite the number of times I have to correct it, I wouldn’t give up autocomplete in today’s world.
Questions
- What are some everyday interactions that you interact with in specialized programs and applications? I.E., beyond just autocorrect. Do you always utilize these features?
- The author took three fairly everyday activities and created new user interfaces with accompanying AI with which to create better tools for human-AI collaboration. What other everyday activities can you think of that you could create a similar tool for?
- How would you gather data to create these programs with AI suggestions? What would you do to infer possible routes?
- The author mentions expanding these interfaces to (human/human) collaboration. What would have to change in order to support this? Would anything?
- DSLs seem to be a somewhat complicated addition to these tools. Why would you want to use these and is it worth learning about the DSL?
- Is ceding control to AI always a bad idea? What areas do you think users should cede more control or should gain back more control?
When it comes to scale and speed, AI certainly scores over humans. I feel like I would cede control to AI when it comes to applications that require sifting through a lot of data. Depending on the accuracy of the AI, I would still like to have control over how the AI sifts through data (i.e. user-provided keywords, thresholds, etc.).
Hi,
your mentioning of learning how many options to show reminded me of one of the previous readings where people had the option to choose more results to display vs show higher precision results.
I think rather than ethnographic research to pick a definite number, I think it would serve well to allow the users to pick how many they want, especially since one shoe can’t fit all feet.
Regarding your question 1, I use Grammarly for writing documents online. I always keep it on, and using it is very intuitive and natural.
I interact with Microsoft Outlook on a daily bases and everyday morning I receive a follow up email asking me to schedule reply to an email or schedule an appointment if I have previously stated that. I can either take the proposed action or dismiss the recommendation.