Large language models (LLMs) are trained on broad corpora and then used in communities with specialized norms. Is providing LLMs with community rules enough for models to follow these norms? We evaluate LLMs' capacity to detect (Task 1) and correct (Task 2) biased Wikipedia edits according to Wikipedia's Neutral Point of View (NPOV) policy. LLMs struggled with bias detection, achieving only 64% accuracy on a balanced dataset. Models exhibited contrasting biases (some under- and others over-predicted bias), suggesting distinct priors about neutrality. LLMs performed better at generation, removing 79% of words removed by Wikipedia editors. However, LLMs made additional changes beyond Wikipedia editors' simpler neutralizations, resulting in high-recall but low-precision editing. Interestingly, crowdworkers rated AI rewrites as more neutral (70%) and fluent (61%) than Wikipedia-editor rewrites. Qualitative analysis found LLMs sometimes applied NPOV more comprehensively than Wikipedia editors but often made extraneous non-NPOV-related changes (such as grammar). LLMs may apply rules in ways that resonate with the public but diverge from community experts. While potentially effective for generation, LLMs may reduce editor agency and increase moderation workload (e.g., verifying additions). Even when rules are easy to articulate, having LLMs apply them like community members may still be difficult.

As Large Language Models (LLMs) are nondeterministic, the same input can generate different outputs, some of which may be incorrect or hallucinated. If run again, the LLM may correct itself and produce the correct answer. Unfortunately, most LLM-powered systems resort to single results which, correct or not, users accept. Having the LLM produce multiple outputs may help identify disagreements or alternatives. However, it is not obvious how the user will interpret conflicts or inconsistencies. To this end, we investigate how users perceive the AI model and comprehend the generated information when they receive multiple, potentially inconsistent, outputs. Through a preliminary study, we identified five types of output inconsistencies. Based on these categories, we conducted a study (N=252) in which participants were given one or more LLM-generated passages to an information-seeking question. We found that inconsistency within multiple LLM-generated outputs lowered the participants' perceived AI capacity, while also increasing their comprehension of the given information. Specifically, we observed that this positive effect of inconsistencies was most significant for participants who read two passages, compared to those who read three. Based on these findings, we present design implications that, instead of regarding LLM output inconsistencies as a drawback, we can reveal the potential inconsistencies to transparently indicate the limitations of these models and promote critical LLM usage.

Generative AI has the potential to create a new form of interactive media: AI-bridged creative language arts (CLA), which bridge the author and audience by personalizing the author's vision to the audience's context and taste at scale. However, it is unclear what the authors' values and attitudes would be regarding AI-bridged CLA. To identify these values and attitudes, we conducted an interview study with 18 authors across eight genres (e.g., poetry, comics) by presenting speculative but realistic AI-bridged CLA scenarios. We identified three benefits derived from the dynamics between author, artifact, and audience: those that 1) authors get from the process, 2) audiences get from the artifact, and 3) authors get from the audience. We found how AI-bridged CLA would either promote or reduce these benefits, along with authors' concerns. We hope our investigation hints at how AI can provide intriguing experiences to CLA audiences while promoting authors' values.

While diffusion-based text-to-image (T2I) models provide a simple and powerful way to generate images, guiding this generation remains a challenge. For concepts that are difficult to describe through language, users may struggle to create prompts. Moreover, many of these models are built as end-to-end systems, lacking support for iterative shaping of the image. In response, we introduce PromptPaint, which combines T2I generation with interactions that model how we use colored paints. PromptPaint allows users to go beyond language to mix prompts that express challenging concepts. Just as we iteratively tune colors through layered placements of paint on a physical canvas, PromptPaint similarly allows users to apply different prompts to different canvas areas and times of the generative process. Through a set of studies, we characterize different approaches for mixing prompts, design trade-offs, and socio-technical challenges for generative models. With PromptPaint we provide insight into future steerable generative tools.

Thinking of technology as a design material is appealing. It encourages designers to explore the material's properties to understand its capabilities and limitations, a prerequisite to generative design thinking. However, as a material, AI resists this approach because its properties emerge as part of the design process itself. Therefore, designers and AI engineers must collaborate in new ways to create both the material and its application experience. We investigate the co-creation process through a design study with 10 pairs of designers and engineers. We find that design 'probes' with user data are a useful tool in defining AI materials. Through data probes, designers construct designerly representations of the envisioned AI experience (AIX) to identify desirable AI characteristics. Data probes facilitate divergent thinking, material testing, and design validation. Based on our findings, we propose a process model for co-creating AIX and offer design considerations for incorporating data probes in design tools.

Tuning machine learning models, particularly deep learning architectures, is a complex process. Automated hyperparameter tuning algorithms often depend on specific optimization metrics. However, in many situations, a developer trades one metric against another: accuracy versus overfitting, precision versus recall, smaller models and accuracy, etc. With deep learning, not only are the model's representations opaque, the model's behavior when parameters "knobs" are changed may also be unpredictable. Thus, picking the "best" model often requires time-consuming model comparison. In this work, we introduce LAMVI-2, a visual analytics system to support a developer in comparing hyperparameter settings and outcomes. By focusing on word-embedding models ("deep learning for text") we integrate views to compare both high-level statistics as well as internal model behaviors (e.g., comparing word 'distances'). We demonstrate how developers can work with LAMVI-2 to more quickly and accurately narrow down an appropriate and effective application-specific model.

Across all Wikipedia language editions, millions of images augment text in critical ways. This visual encyclopedic knowledge is an important form of wikiwork for editors, a critical part of reader experience, an emerging resource for machine learning, and a lens into cultural differences. However, Wikipedia research--and cross-language edition Wikipedia research in particular--has thus far been limited to text. In this paper, we assess the diversity of visual encyclopedic knowledge across 25 language editions and compare our findings to those reported for textual content. Unlike text, translation in images is largely unnecessary. Additionally, the Wikimedia Foundation, through the Wikipedia Commons, has taken steps to simplify cross-language image sharing. While we may expect that these factors would reduce image diversity, we find that cross-language image diversity rivals, and often exceeds, that found in text. We find that diversity varies between language pairs and content types, but that many images are unique to different language editions. Our findings have implications for readers (in what imagery they see), for editors (in deciding what images to use), for researchers (who study cultural variations), and for machine learning developers (who use Wikipedia for training models).

In this work, we study the use of Twitter by House, Senate and gubernatorial candidates during the midterm (2010) elections in the U.S. Our data includes almost 700 candidates and over 690k documents that they produced and cited in the 3.5 years leading to the elections. We utilize graph and text mining techniques to analyze differences between Democrats, Republicans and Tea Party candidates, and suggest a novel use of language modeling for estimating content cohesiveness. Our findings show significant differences in the usage patterns of social media, and suggest conservative candidates used this medium more effectively, conveying a coherent message and maintaining a dense graph of connections. Despite the lack of party leadership, we find Tea Party members display both structural and language-based cohesiveness. Finally, we investigate the relation between network structure, content and election results by creating a proof-of-concept model that predicts candidate victory with an accuracy of 88.0%.

Social media is playing an increasingly vital role in information dissemination. But with dissemination being more distributed, content often makes multiple hops, and consequently has opportunity to change. In this paper we focus on content that should be changing the least, namely quoted text. We find changes to be frequent, with their likelihood depending on the authority of the copied source and the type of site that is copying. We uncover patterns in the rate of appearance of new variants, their length, and popularity, and develop a simple model that is able to capture them. These patterns are distinct from ones produced when all copies are made from the same source, suggesting that information is evolving as it is being processed collectively in online social media.