Summary
In this paper, the authors suggest 18 design guidelines to build for Human – AI infused systems. These guidelines try to resolve the issues with many human interaction systems that either don’t follow guidelines or follow some guidelines that are not tested or evaluated. These issues include producing offensive, unpredictable, or dangerous results and might let users stop using these systems and therefore proper guidelines are necessary. in addition, advances in the AI field introduced new user demands in area like sound recognition, pattern recognition, and translation. These 18 guidelines help users to understand what the AI systems can do and cannot do and how well can do it, show information related to the task on focus with the appropriate timing in a way that can fit the user social and culture context, make sure that the user can request services when needed or ignore unnecessary services, offer explanation why the system do certain things, maintain a memory of the user recent action and try to learn from it, etc. These guidelines went through different phases starting from consolidating different guidelines, heuristics evaluation, user study, and expert evaluation and revisions. The authors hope that these guidelines will help building better AI-infused systems that can scale and work better with the increased number of users and advances AI algorithms and systems.
Reflection
I found the idea of putting together design guidelines very interesting as it make a standardization that can help building human-AI systems, and also help evaluating or testing these systems and can be used as a baseline when building large scale AI infused systems to avoid an well known issues associated with the previous systems.
I also found that the collection of academic and industrial guidelines are interested since it is collected based on over 20 years human interaction which can be regarded as a very valuable and rich information that can be used in different domains and fields.
I agree with the authors that some of AI-infused systems that not follow certain guidelines are confusing and not effective and sometimes counterproductive when the suggestions or recommendations are irrelevant, and that explains why some AI enabled or infused systems were popular on certain times but they couldn’t satisfy the user demands and eventually stopped being used by users.
Questions
- Are these guidelines followed in Amazon Mechanical Turk?
- The authors mention that there is a tradeoff between generality and specializations, what tradeoff factors when need to consider?
