Designing Interfaces for Human-Computer Communication: An On-Going Collection of Considerations
share
While we do not always use words, communicating what we want to an AI is a conversation – with ourselves as well as with it, a recurring loop with optional steps depending on the complexity of the situation and our request. Any given conversation of this type may include: (a) the human forming an intent, (b) the human expressing that intent as a command or utterance, (c) the AI performing one or more rounds of inference on that command to resolve ambiguities and/or requesting clarifications from the human, (d) the AI showing the inferred meaning of the command and/or its execution on current and future situations or data, (e) the human hopefully correctly recognizing whether the AI's interpretation actually aligns with their intent. In the process, they may (f) update their model of the AI's capabilities and characteristics, (g) update their model of the situations in which the AI is executing its interpretation of their intent, (h) confirm or refine their intent, and (i) revise their expression of their intent to the AI, where the loop repeats until the human is satisfied. With these critical cognitive and computational steps within this back-and-forth laid out as a framework, it is easier to anticipate where communication can fail, and design algorithms and interfaces that ameliorate those failure points.
READ FULL TEXT