Authorship obfuscation, rewriting a text to intentionally obscure the identity of the author, is an important but challenging task.Current methods using large language models (LLMs) lack interpretability and controllability, often ignoring author-specific stylistic features, resulting in less robust performance overall.To address this, we develop STYLEREMIX, an adaptive and interpretable obfuscation method that perturbs specific, fine-grained style elements of the original input text.STYLEREMIX uses pre-trained Low Rank Adaptation (LoRA) modules to rewrite an input specifically along various stylistic axes (e.g., formality and length) while maintaining low computational cost.STYLEREMIX outperforms state-of-theart baselines and much larger LLMs in a variety of domains as assessed by both automatic and human evaluation.Additionally, we release AUTHORMIX, a large set of 30K high-quality, long-form texts from a diverse set of 14 authors and 4 domains, and DISC, a parallel corpus of 1,500 texts spanning seven style axes in 16 unique directions 1 .

With increased power and prevalence of AI systems, it is ever more critical that AI systems are designed to serve all, i.e., people with diverse values and perspectives. However, aligning models to serve pluralistic human values remains an open research question. In this piece, we propose a roadmap to pluralistic alignment, specifically using large language models as a test bed. We identify and formalize three possible ways to define and operationalize pluralism in AI systems: 1) Overton pluralistic models that present a spectrum of reasonable responses; 2) Steerably pluralistic models that can steer to reflect certain perspectives; and 3) Distributionally pluralistic models that are well-calibrated to a given population in distribution. We also formalize and discuss three possible classes of pluralistic benchmarks: 1) Multi-objective benchmarks, 2) Trade-off steerable benchmarks that incentivize models to steer to arbitrary trade-offs, and 3) Jury-pluralistic benchmarks that explicitly model diverse human ratings. We use this framework to argue that current alignment techniques may be fundamentally limited for pluralistic AI; indeed, we highlight empirical evidence, both from our own experiments and from other work, that standard alignment procedures might reduce distributional pluralism in models, motivating the need for further research on pluralistic alignment.

How can online communities execute a focused vision for their space? Curation offers one approach, where community leaders manually select content to share with the community. Curation enables leaders to shape a space that matches their taste, norms, and values, but the practice is often intractable at social media scale: curators cannot realistically sift through hundreds or thousands of submissions daily. In this paper, we contribute algorithmic and interface foundations enabling curation at scale, and manifest these foundations in a system called Cura. Our approach draws on the observation that, while curators' attention is limited, other community members' upvotes are plentiful and informative of curators' likely opinions. We thus contribute a transformer-based curation model that predicts whether each curator will upvote a post based on previous community upvotes. Cura applies this curation model to create a feed of content that it predicts the curator would want in the community. Evaluations demonstrate that the curation model accurately estimates opinions of diverse curators, that changing curators for a community results in clearly recognizable shifts in the community's content, and that, consequently, curation can reduce anti-social behavior by half without extra moderation effort. By sampling different types of curators, Cura lowers the threshold to genres of curated social media ranging from editorial groups to stakeholder roundtables to democracies.

Whose labels should a machine learning (ML) algorithm learn to emulate? For ML tasks ranging from online comment toxicity to misinformation detection to medical diagnosis, different groups in society may have irreconcilable disagreements about ground truth labels. Supervised ML today resolves these label disagreements implicitly using majority vote, which overrides minority groups’ labels. We introduce jury learning, a supervised ML approach that resolves these disagreements explicitly through the metaphor of a jury: defining which people or groups, in what proportion, determine the classifier’s prediction. For example, a jury learning model for online toxicity might centrally feature women and Black jurors, who are commonly targets of online harassment. To enable jury learning, we contribute a deep learning architecture that models every annotator in a dataset, samples from annotators’ models to populate the jury, then runs inference to classify. Our architecture enables juries that dynamically adapt their composition, explore counterfactuals, and visualize dissent. A field evaluation finds that practitioners construct diverse juries that alter 14% of classification outcomes.

Because algorithm audits are conducted by technical experts, audits are necessarily limited to the hypotheses that experts think to test. End users hold the promise to expand this purview, as they inhabit spaces and witness algorithmic impacts that auditors do not. In pursuit of this goal, we propose end-user audits-system-scale audits led by non-technical users-and present an approach that scaffolds end users in hypothesis generation, evidence identification, and results communication. Today, performing a system-scale audit requires substantial user effort to label thousands of system outputs, so we introduce a collaborative filtering technique that leverages the algorithmic system's own disaggregated training data to project from a small number of end user labels onto the full test set. Our end-user auditing tool, IndieLabel, employs these predicted labels so that users can rapidly explore where their opinions diverge from the algorithmic system's outputs. By highlighting topic areas where the system is under-performing for the user and surfacing sets of likely error cases, the tool guides the user in authoring an audit report. In an evaluation of end-user audits on a popular comment toxicity model with 17 non-technical participants, participants both replicated issues that formal audits had previously identified and also raised previously underreported issues such as under-flagging on veiled forms of hate that perpetuate stigma and over-flagging of slurs that have been reclaimed by marginalized communities.

Machine learning classifiers for human-facing tasks such as comment toxicity and misinformation often score highly on metrics such as ROC AUC but are received poorly in practice. Why this gap? Today, metrics such as ROC AUC, precision, and recall are used to measure technical performance; however, human-computer interaction observes that evaluation of human-facing systems should account for people’s reactions to the system. In this paper, we introduce a transformation that more closely aligns machine learning classification metrics with the values and methods of user-facing performance measures. The disagreement deconvolution takes in any multi-annotator (e.g., crowdsourced) dataset, disentangles stable opinions from noise by estimating intra-annotator consistency, and compares each test set prediction to the individual stable opinions from each annotator. Applying the disagreement deconvolution to existing social computing datasets, we find that current metrics dramatically overstate the performance of many human-facing machine learning tasks: for example, performance on a comment toxicity task is corrected from .95 to .73 ROC AUC.

We have limited understanding of how older adults use smartphones, how their usage differs from younger users, and the causes for those differences. As a result, researchers and developers may miss promising opportunities to support older adults or offer solutions to unimportant problems. To characterize smartphone usage among older adults, we collected iPhone usage data from 84 healthy older adults over three months. We find that older adults use fewer apps, take longer to complete tasks, and send fewer messages. We use cognitive test results from these same older adults to then show that up to 79% of these differences can be explained by cognitive decline, and that we can predict cognitive test performance from smartphone usage with 83% ROCAUC. While older adults differ from younger adults in app usage behavior, the "cognitively young" older adults use smartphones much like their younger counterparts. Our study suggests that to better support all older adults, researchers and developers should consider the full spectrum of cognitive function.

We study the effects of haptic augmentation on tapping, path following, and drag & drop tasks based on a recent flagship smartphone with refined touch sensing and haptic actuator technologies. Results show actuated haptic confirmation on tapping targets was subjectively appreciated by some users but did not improve tapping speed or accuracy. For drag & drop, a clear performance improvement was measured when haptic feedback is applied to target boundary crossing, particularly when the targets are small. For path following tasks, virtual haptic feedback improved accuracy at a reduced speed in a sitting condition. Stronger results were achieved in a physical haptic mock-up. Overall, we found actuated touchscreen haptic feedback particularly effective when the touched object was visually interfered by the finger. Participants subjective experience of haptic feedback in all tasks tended to be more positive than their time or accuracy performance suggests. We compare and discuss these findings with previous results on early generations of devices. The work provides an empirical foundation to product design and future research of touch input and haptic systems.

Generative models often use human evaluations to measure the perceived quality of their outputs. Automated metrics are noisy indirect proxies, because they rely on heuristics or pretrained embeddings. However, up until now, direct human evaluation strategies have been ad-hoc, neither standardized nor validated. Our work establishes a gold standard human benchmark for generative realism. We construct Human eYe Perceptual Evaluation (HYPE) a human benchmark that is (1) grounded in psychophysics research in perception, (2) reliable across different sets of randomly sampled outputs from a model, (3) able to produce separable model performances, and (4) efficient in cost and time. We introduce two variants: one that measures visual perception under adaptive time constraints to determine the threshold at which a model's outputs appear real (e.g. $250$ms), and the other a less expensive variant that measures human error rate on fake and real images sans time constraints. We test HYPE across six state-of-the-art generative adversarial networks and two sampling techniques on conditional and unconditional image generation using four datasets: CelebA, FFHQ, CIFAR-10, and ImageNet. We find that HYPE can track model improvements across training epochs, and we confirm via bootstrap sampling that HYPE rankings are consistent and replicable.

that goal. Despite its crucial role in how a number of prominent self-tracking apps function, there has been relatively little investigation of the goal-setting and achievement aspects of self-tracking apps. Here we explore this issue, investigating a particular goal setting and achievement process that is extensive, recorded, and crucial for both the app and its users' success: weight loss goals in MyFitnessPal. We present a large-scale study of 1.4 million users and weight loss goals, allowing for an unprecedented detailed view of how people set and achieve their goals. We find that, even for difficult long-term goals, behavior within the first 7 days predicts those who ultimately achieve their goals, that is, those who lose at least as much weight as they set out to, and those who do not. For instance, high amounts of early weight loss, which some researchers have classified as unsustainable, leads to higher goal achievement rates. We also show that early food intake, self-monitoring motivation, and attitude towards the goal are important factors. We then show that we can use our findings to predict goal achievement with an accuracy of 79% ROC AUC just 7 days after a goal is set. Finally, we discuss how our findings could inform steps to improve goal achievement in self-tracking apps.

Crowd-powered systems leverage human intelligence to go beyond the capabilities of automated systems, but also introduce privacy and security concerns because unknown people must view the data that the system processes. While automated approaches cannot robustly filter private information from these datasets, people have the ability to do so if the risk from them viewing the data can be mitigated. We present a crowd-powered approach to masking private content in data by segmenting and distributing smaller segments to crowd workers so that individual workers can identify potentially private content without being able to fully view it themselves. We introduce a novel pyramid workflow for segmentation that uses segments at multiple levels of granularity to overcome problems with fixed-sized approaches. We implement our approach in CrowdMask, a system that allows images with potentially sensitive content to be masked by appearing in progressively larger, more identifiable segments, and masking portions of the image as soon as a risk is identified. Our experiments with 4134 Mechanical Turk workers show that CrowdMask can effectively mask private content from images without revealing sensitive content to constituent workers, while still enabling future systems to use the filtered result.

We present WatchWriter, a finger operated keyboard that supports both touch and gesture typing with statistical decoding on a smartwatch. Just like on modern smartphones, users type one letter per tap or one word per gesture stroke on WatchWriter but in a much smaller spatial scale. WatchWriter demonstrates that human motor control adaptability, coupled with modern statistical decoding and error correction technologies developed for smartphones, can enable a surprisingly effective typing performance despite the small watch size. In a user performance experiment entirely run on a smartwatch, 36 participants reached a speed of 22-24 WPM with near zero error rate.

Coding behavioral video is an important method used by researchers to understand social phenomenon. Unfortunately, traditional hand-coding approaches can take days or weeks of time to complete. Recent work has shown that these tasks can be completed quickly by leveraging the parallelism of large online crowds, but using the crowd introduces new concerns about accuracy, reliability, privacy, and cost. To explore these issues, we conducted interviews with 12 researchers who frequently code behavioral video, to investigate common practices and challenges with video coding. We find accuracy and privacy to be the researchers' primary concerns. To explore this more concretely, we used sample videos to investigate whether crowds can accurately recognize instances of commonly coded behaviors, and show that the crowd yields accurate results. Then, we demonstrate a method for obfuscating participant identity with a video blur filter, and find, as expected, that workers' ability to identify participants decreases as blur level increases. The workers' ability to accurately and reliably code behaviors also decreases, but not as steeply as the identity test. This trade-off between coding quality and privacy protection suggests that researchers can use online crowds to code for some key behaviors in video without compromising participant identity. We conclude with a discussion of how researchers can balance privacy and accuracy on their own data using a system we introduce called Incognito.

A common way to learn is by studying written step-by-step tutorials such as worked examples. However, tutorials for computer programming can be tedious to create since a static text-based format cannot convey what happens as code executes. We created a system called Codepourri that enables people to easily create visual coding tutorials by annotating steps in an automatically-generated program visualization. Using Codepourri, we developed a novel crowdsourcing workflow where learners who are visiting an educational Web site (www. pythontutor.com) collectively create a tutorial by annotating execution steps in a piece of code and then voting on the best annotations. Since there are far more learners than experts, using learners as a crowd is a potentially more scalable way of creating tutorials. Our experiments with 4 expert judges and 101 learners adding 145 raw annotations to two pieces of textbook Python code show the learner crowd's annotations to be accurate, informative, and containing some insights that even experts missed.

Online discussion forums are one of the most ubiquitous kinds of resources for people who are learning computer programming. However, their user interface - a hierarchy of textual threads - has not changed much in the past four decades. We argue that generic forum interfaces are cumbersome for learning programming and that there is a need for a domain-specific visual discussion forum for programming. We support this argument with an empirical study of all 5,377 forum threads in Introduction to Computer Science and Programming Using Python, a popular edX MOOC. Specifically, we investigated how forum participants were hampered by its text-based format. Most notably, people often wanted to discuss questions about dynamic execution state - what happens “under the hood” as the computer runs code. We propose that a better forum for learning programming should be visual and domain-specific, integrating automatically-generated visualizations of execution state and enabling inline annotations of source code and output.

Evaluating the results of user accessibility testing on the web can take a significant amount of time, training, and effort. Some of this work can be offloaded to others through coding video data from user tests to systematically extract meaning from subtle human actions and emotions. However, traditional video coding methods can take a considerable amount of time. We have created Glance, a tool that uses the crowd to allow researchers to rapidly query, sample, and analyze large video datasets for behavioral events that are hard to detect automatically. In this abstract, we discuss how Glance can be used to quickly code video of users with special needs interacting with a website by coding for whether or not websites conform with accessibility guidelines, in order to evaluate how accessible a website is and where potential problems lie.

Behavioral researchers code video to extract systematic meaning from subtle human actions and emotions. While this has traditionally been done by analysts within a research group, recent methods have leveraged online crowds to massively parallelize this task and reduce the time required from days to seconds. However, using the crowd to code video increases the risk that private information will be disclosed because workers who have not been vetted will view the video data in order to code it. In this Work-in-Progress, we discuss techniques for maintaining privacy when using Glance to code video and present initial experimental evidence to support them.

Behavioral researchers spend considerable amount of time coding video data to systematically extract meaning from subtle human actions and emotions. In this paper, we present Glance, a tool that allows researchers to rapidly query, sample, and analyze large video datasets for behavioral events that are hard to detect automatically. Glance takes advantage of the parallelism available in paid online crowds to interpret natural language queries and then aggregates responses in a summary view of the video data. Glance provides analysts with rapid responses when initially exploring a dataset, and reliable codings when refining an analysis. Our experiments show that Glance can code nearly 50 minutes of video in 5 minutes by recruiting over 60 workers simultaneously, and can get initial feedback to analysts in under 10 seconds for most clips. We present and compare new methods for accurately aggregating the input of multiple workers marking the spans of events in video data, and for measuring the quality of their coding in real-time before a baseline is established by measuring the variance between workers. Glance's rapid responses to natural language queries, feedback regarding question ambiguity and anomalies in the data, and ability to build on prior context in followup queries allow users to have a conversation-like interaction with their data - opening up new possibilities for naturally exploring video data.