The safety alignment of current Large Language Models (LLMs) is vulnerable. Simple attacks, or even benign fine-tuning, can jailbreak aligned models. We note that many of these vulnerabilities are related to a shared underlying issue: safety alignment can take shortcuts, wherein the alignment adapts a model's generative distribution primarily over only its very first few output tokens. We unifiedly refer to this issue as shallow safety alignment. In this paper, we present case studies to explain why shallow safety alignment can exist and show how this issue universally contributes to multiple recently discovered vulnerabilities in LLMs, including the susceptibility to adversarial suffix attacks, prefilling attacks, decoding parameter attacks, and fine-tuning attacks. The key contribution of this work is that we demonstrate how this consolidated notion of shallow safety alignment sheds light on promising research directions for mitigating these vulnerabilities. We show that deepening the safety alignment beyond the first few tokens can meaningfully improve robustness against some common exploits. We also design a regularized fine-tuning objective that makes the safety alignment more persistent against fine-tuning attacks by constraining updates on initial tokens. Overall, we advocate that future safety alignment should be made more than just a few tokens deep.

Despite huge gains in performance in natural language understanding via large language models in recent years, voice assistants still often fail to meet user expectations. In this study, we conducted a mixed-methods analysis of how voice assistant failures affect users’ trust in their voice assistants. To illustrate how users have experienced these failures, we contribute a crowdsourced dataset of 199 voice assistant failures, categorized across 12 failure sources. Relying on interview and survey data, we find that certain failures, such as those due to overcapturing users’ input, derail user trust more than others. We additionally examine how failures impact users’ willingness to rely on voice assistants for future tasks. Users often stop using their voice assistants for specific tasks that result in failures for a short period of time before resuming similar usage. We demonstrate the importance of low stakes tasks, such as playing music, towards building trust after failures.

Preethi Lahoti, Nicholas Blumm, Xiao Ma, Raghavendra Kotikalapudi, Sahitya Potluri, Qijun Tan, Hansa Srinivasan, Ben Packer, Ahmad Beirami, Alex Beutel, Jilin Chen. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing. 2023.

We are entering an era of AI-Mediated Communication (AI-MC) where interpersonal communication is not only mediated by technology, but is optimized, augmented, or generated by artificial intelligence. Our study takes a first look at the potential impact of AI-MC on online self-presentation. In three experiments we test whether people find Airbnb hosts less trustworthy if they believe their profiles have been written by AI. We observe a new phenomenon that we term the Replicant Effect: Only when participants thought they saw a mixed set of AI- and human-written profiles, they mistrusted hosts whose profiles were labeled as or suspected to be written by AI. Our findings have implications for the design of systems that involve AI technologies in online self-presentation and chart a direction for future work that may upend or augment key aspects of Computer-Mediated Communication theory.

Trust facilitates cooperation and supports positive outcomes in social groups, including member satisfaction, information sharing, and task performance. Extensive prior research has examined individuals' general propensity to trust, as well as the factors that contribute to their trust in specific groups. Here, we build on past work to present a comprehensive framework for predicting trust in groups. By surveying 6,383 Facebook Groups users about their trust attitudes and examining aggregated behavioral and demographic data for these individuals, we show that (1) an individual's propensity to trust is associated with how they trust their groups, (2) smaller, closed, older, more exclusive, or more homogeneous groups are trusted more, and (3) a group's overall friendship-network structure and an individual's position within that structure can also predict trust. Last, we demonstrate how group trust predicts outcomes at both individual and group level such as the formation of new friendship ties.

As any savvy online shopper knows, second-hand peer-to-peer marketplaces are filled with images of mixed quality. How does image quality impact marketplace outcomes, and can quality be automatically predicted? In this work, we conducted a large-scale study on the quality of user-generated images in peer-to-peer marketplaces. By gathering a dataset of common second-hand products (≈75,000 images) and annotating a subset with human-labeled quality judgments, we were able to model and predict image quality with decent accuracy (≈87%). We then conducted two studies focused on understanding the relationship between these image quality scores and two marketplace outcomes: sales and perceived trustworthiness. We show that image quality is associated with higher likelihood that an item will be sold, though other factors such as view count were better predictors of sales. Nonetheless, we show that high quality user-generated images selected by our models outperform stock imagery in eliciting perceptions of trust from users. Our findings can inform the design of future marketplaces and guide potential sellers to take better product images.

We build on the increasing availability of Virtual Reality (VR) devices and Web technologies to conduct behavioral experiments in VR using crowdsourcing techniques. A new recruiting and validation method allows us to create a panel of eligible experiment participants recruited from Amazon Mechanical Turk. Using this panel, we ran three different crowdsourced VR experiments, each reproducing one of three VR illusions: place illusion, embodiment illusion, and plausibility illusion. Our experience and worker feedback on these experiments show that conducting Web-based VR experiments using crowdsourcing is already feasible, though some challenges---including scale---remain. Such crowdsourced VR experiments on the Web have the potential to finally support replicable VR experiments with diverse populations at a low cost.

In recent years, several mobile applications allowed individuals to anonymously share information with friends and contacts, without any persistent identity marker. The functions of these "tie-based" anonymity services may be notably different than other social media services. We use semi-structured interviews to qualitatively examine motivations, practices and perceptions in two tie-based anonymity apps: Secret (now defunct, in the US) and Mimi (in China). Among the findings, we show that: (1) while users are more comfortable in self-disclosure, they still have specific practices and strategies to avoid or allow identification; (2) attempts for deidentification of others are prevalent and often elaborate; and (3) participants come to expect both negativity and support in response to posts. Our findings highlight unique opportunities and potential benefits for tie-based anonymity apps, including serving disclosure needs and social probing. Still, challenges for making such applications successful, for example the prevalence of negativity and bullying, are substantial.

Many location-based dating applications allow users to search for potential matches who are physically proximate. A recent mobile dating application, happn, adds a temporal dimension to location-based dating, showing users the number of times that they crossed path with each other, as well as the location of the most recent overlap. We conducted qualitative interviews with 15 happn users to understand how people make sense of crossed paths, and assess the meanings they assign to these location overlaps. Building on Uncertainty Reduction Theory, we show the various outcomes of the crossed paths and how they play a role in uncertainty reduction. In particular, the warranting power of the device-driven location data was accepted as valuable, and generated little concern about misrepresentation. Moreover, people assigned significant meaning to the minimal cues available from the overlap data. In addition, the location overlap data was useful in allowing users to estimate convenience in meeting and establish common ground. On the other hand, concerns of security and recognition by known others persisted in the happn app. Our findings suggest the potential for utilizing location data outside of the domain of online dating.

Online peer-to-peer platforms like Airbnb allow hosts to list a property (e.g. a house, or a room) for short-term rentals. In this work, we examine how hosts describe themselves on their Airbnb profile pages. We use a mixed-methods study to develop a categorization of the topics that hosts self-disclose in their profile descriptions, and show that these topics differ depending on the type of guest engagement expected. We also examine the perceived trustworthiness of profiles using topic-coded profiles from 1,200 hosts, showing that longer self-descriptions are perceived to be more trustworthy. Further, we show that there are common strategies (a mix of topics) hosts use in self-disclosure, and that these strategies cause differences in perceived trustworthiness scores. Finally, we show that the perceived trustworthiness score is a significant predictor of host choice--especially for shorter profiles that show more variation. The results are consistent with uncertainty reduction theory, reflect on the assertions of signaling theory, and have important design implications for sharing economy platforms, especially those facilitating online-to-offline social exchange.

Self-disclosure is rewarding and provides significant benefits for individuals, but it also involves risks, especially in social media settings. We conducted an online experiment to study the relationship between content intimacy and willingness to self-disclose in social media, and how identification (real name vs. anonymous) and audience type (social ties vs. people nearby) moderate that relationship. Content intimacy is known to regulate self-disclosure in face-to-face communication: people self-disclose less as content intimacy increases. We show that such regulation persists in online social media settings. Further, although anonymity and an audience of social ties are both known to increase self-disclosure, it is unclear whether they (1) increase self-disclosure baseline for content of all intimacy levels, or (2) weaken intimacy's regulation effect, making people more willing to disclose intimate content. We show that intimacy always regulates self-disclosure, regardless of settings. We also show that anonymity mainly increases self-disclosure baseline and (sometimes) weakens the regulation. On the other hand, an audience of social ties increases the baseline but strengthens the regulation. Finally, we demonstrate that anonymity has a more salient effect on content of negative valence.The results are critical to understanding the dynamics and opportunities of self-disclosure in social media services that vary levels of identification and types of audience.