Getting machines to support human experiences, desires, and goals requires imbuing them with representations of human situations that are beyond what can be captured through context features like APIs (e.g., Yelp, Google Awareness) and component detectors (e.g., time, location). To this end, we scaffold the creation of concept expressions: human-created translations of high-level situations like "Places to have a private conversation" into collections of context-features that machines can act on.

Creating concept expressions, however, is no easy task. Designers can fixate on early ideas, and generally struggle to comprehensively conceptualize a variety of ways that a situation can be defined. For instance, to construct a concept expression for “places to have a private conversation,” a designer might fixate on quiet places such as empty offices or quiet rooms at home, and miss entirely the alternative conception that private conversations can also happen in public spaces. They can also miss hindrances, for instance that a restaurant may be good public place for a private conversation, but not if it is too quiet, as others can easily overhear the conversation. 

To support designers reflecting on a situation at large, we propose to use AI models to support designers creating concept expressions by surfacing diverse conceptions or examples that correspond to different conceptions. While such models are hardly perfect and do contain inaccuracies, they often capture a wider span of the various ways to conceptualize a situation, and of possible hinderances that a human designer can struggle to recall. This can help designers find relevant concepts to add to an expression, either as alternative conceptions of a situation (e.g., public spaces for having a private conversation), or of hindrances that need to be considered (e.g., heard by others).  Within AI supported conception, an immediate focus is understanding the process of augmentation. While AI language models are powerful tools for suggesting ideas, creating a concept expression requires a translation of suggested examples, characteristics, and hindrances into sub-concepts and context features with clearly defined relationships to one another. It's important to gain clarity on how a suggestion is incorporated, and to support better ways of doing so in an interface.

An additional focus is on supporting designers in extending their ideas to recall further concepts and details from AI suggestions; in addition to the task of incorporation, we envision the possibility for a fruitful back-and-forth process where AI can help guide a designer's thinking in the same way that a designer would intelligently query the AI.