Saleema Amershi's research while affiliated with Microsoft and other places

... Last, we randomize question, block, and interface order to control for biases due to showing interfaces or questions first. Metrics Following previous work on evaluating human and ML coordination and trust, we assess several metrics to evaluate user experiences [15,20,22]. We evaluate the following statements along 1-7 Likert scale at the end of the survey: ...

... Secondly, testing different ML errors in advance would allow designers to create user interfaces that respond or adapt to ML errors, making the overall UX more pleasant or understandable. Previous studies have shown that ML errors can greatly impact the UX [1,10,11], the mental models users form to interact with the system, and their trust in the ML model [17]. Different error types could differently influence the UX, e.g. ...

... Therefore, analogical search engines should help to reduce the cognitive effort required in the process, for example by proactively retrieving results that are 'usefully' misaligned such that searchers can better recognize (as opposed to having to recall) salient constraints and refine their problem representation. This process is deeply exploratory [93,115,118] in nature, and suggest the importance of both providing end-users a sense of progress over time [103] as well as adequate feedback mechanisms for the machine to adjust according to the changing end-user search intent [72,95,96]. For example, while the machine may 'correctly' recognize a significant anaogical relevance at a higher level of purpose representation and recommend macro-scale mechanisms to a scientist who studies nano-scale phenomena (P1 Study 1 ) or solid and semiconductor-based cooling mechanisms to a scientist in liquid and evaporative cooling systems (P3 Study 1 ), these analogs may be critically misaligned on the specific constraints of the problem (i.e. the scale or materialistic phase) and thus considered by end-users as useless and even harmful. ...

... Such high user-trust in AI systems played out in several ways in our study: ready acceptances of terms, conditions, and loan decisions, often to the extent of users reevaluating their own competencies and abilities. However, design and research in user-centered AI often assumes low trust in AI, and begins with questions of 'how might we design for increased user trust in AI'? Instead, designs must plan for appropriate failures assuming high-user trust in AI systems [14]. Research must address questions such as decreasing user trust or increasing user distrust in AI systems. ...

... To complicate things further, there is often some ambiguity about whether the change the control offers is transient (e.g., for the current session) or for the longer term. One proposed design is to highlight items that will be added and removed from the user's feed when a control is changed (Schnabel et al., 2020). ...

... But the absence of any necessary condition explains that organizations cannot successfully leverage the ML value creation mechanism. While previous studies recognize and mention many resources and capabilities necessary for successful implementation of ML applications (e.g., Akkiraju et al., 2020;Amershi et al., 2019;Fountaine et al., 2019;Mikalef & Gupta, 2021;Tarafdar et al., 2019), these resources and capabilities together with environmental factors have not yet been linked to specific ML value creation mechanisms. Without this integrated view, previous studies attribute the high failure rate to organizational implementation and restructuring time lags (Brynjolfsson et al., 2017). ...

... HCML has captured the interest of ML and other computing disciplines. This interest has been reflected in a series of workshops and activities on human-centered perspectives at major HCI venues, such as ACM SIGCHI in 2016 [22], and two in 2019 [26,38]. There has been attention in ML too, as shown through the rapid ascent of the FAT ML workshop to its own independent ACM FAccT conference. ...

... In a personalized context, Buschek et al. [8] examined the potential pitfalls for achieving user interests in the co-creation context. The limitations on the machine side, identified as lack of machine creativity [41], and usability [37], and thus, highlight a biased AI with trained system bias but lacks discussions on the source of bias and the mismatch between individual expectations and system abilities. In terms of mismatched expectations, Eiband et al. [15] reported that users might intentionally provide flawed inputs when a system fails to achieve their satisfaction in everyday intelligent applications. ...

... At the software development stage, it has been recommended to include experts in user interface design 8,18 . Designing a good user interface and interaction requires careful consideration of the cognitive load of the end users 10,18,68,99,100 , by showing only relevant information in the right context, and by allowing adjustment of its behavior by end users 99 . ...

... The automation aspect of NLP was a successful conflict element that evoked a range of reactions (both individual and collaborative [96]). In that way, we used design fiction to co-speculate with the participants (implication, ideation), to learn about their values expectations, neither the underlying details of model workings [144], nor solutions [20,21]. ...