Search arXiv⌕ Search

arXiv subjects

Eric Gilbert

Publications and source records attributed to Eric Gilbert.

At least 19 recordsLinked to original sources
arXiv2026

Information Discernment in Large Language Models

LLMs are increasingly used with external knowledge sources like the internet. Do they weigh information appropriately -- updating more for reliable sources (source discernment) and more when claims bring priors closer to the truth (truth discernment)? We formalize this as information discernment and introduce Learn2Discern (L2D), an experimental framework and benchmark grounded in three normative axioms with interpretable metrics. To establish external validity, a pre-registered, quota-matched user study (n=299) confirms that real LLM users endorse all three axioms and report that violations reduce their trust and usage intent. Across 13 models and nearly 670K trials, we find consistent failures across both dimensions: models perform near chance on source and truth discernment, rely on source popularity twice as much as source reliability, and update roughly equally whether a claim improves or worsens their position relative to the ground truth. Models integrate external knowledge most effectively on datasets where their priors are already the most accurate. Newer and larger models improve truth discernment but not source discernment, a blind spot that model complexity does not address. We identify simple inference-time interventions that improve both forms of discernment. We release our dataset and survey as a testbed for a core alignment property that scales in importance as LLMs replace traditional search.

cs.AI↗
arXiv2026

Position: AI/ML Deepfake Research is Misaligned with AI-Generated Non-Consensual Intimate Imagery (AIG-NCII)

AI-generated non-consensual intimate imagery (AIG-NCII) is not adequately addressed in AI/ML literature regarding AI-generated media, commonly referred to as "deepfakes". While research on deepfakes currently focuses on its epistemic harms -- or harms relating to truth and authenticity -- this is misaligned with the dominant reality of generative AI abuse involving sexualized imagery. We conduct a landscape analysis of highly-cited works to demonstrate that technical interventions addressing deepfakes almost entirely ignore AIG-NCII, limiting the research ecosystem to authenticity detection tools. In this position paper, we argue that existing interventions address viewer-centric epistemic harms, such as fraud or scams, but ignore subject-centric dignity harms, such as AIG-NCII. We illustrate that knowing an image is synthetic does not mitigate harms to subjects and may, in some cases, even exacerbate them. We conclude by offering recommendations to realign the field, including updating threat models to consider subject-centric harms and addressing AIG-NCII in AI safety research. Finally, we caution that researchers should only engage in this high-risk domain if they implement safety guardrails for both subjects and researchers and establish partnerships with domain experts in sexual violence prevention.

cs.AI↗
arXiv2025

Platforms as Crime Scene, Judge, and Jury: How Victim-Survivors of Non-Consensual Intimate Imagery Report Abuse Online

Non-consensual intimate imagery (NCII), also known as image-based sexual abuse (IBSA), is mediated through online platforms. Victim-survivors must turn to platforms to collect evidence and request content removal. Platforms act as the crime scene, judge, and jury, determining whether perpetrators face consequences and if harmful material is removed. We present a study of NCII victim-survivors' online reporting experiences, drawing on trauma-informed interviews with 13 participants. We find that platform reporting processes are hostile, opaque, and ineffective, often forcing complex harms into narrow interfaces, responding inconsistently, and failing to result in meaningful action. Leveraging institutional betrayal theory, we show how platforms' structures and practices compound harm, and, in doing so, surface concrete intervention points for redesigning reporting systems and shaping policy to better support victim-survivors

cs.HC↗
arXiv2025

Follow Nudges without Budges: A Field Experiment on Misinformation Followers Didn't Change Follow Networks

Can digital ads encourage users exposed to inaccurate information sources to follow accurate ones? We conduct a large-scale field experiment (N=28,582) on X, formerly Twitter, with users who follow accounts that spread health misinformation. Participants were exposed to four ad treatments varied on two dimensions: a neutral message versus a persuasive message appealing to values of independence, and a request to follow a health institution versus a request to follow a health influencer. We term this ad-based, social network intervention a follow nudge. The ad with a persuasive message to follow a well-known health institution generated significantly higher click-through rates than all other conditions (Bonferroni-corrected pairwise tests, all p<0.001). Given the overall low click-through rate across treatments and the high cost of digital advertising infrastructure on X, however, we conclude that our proposed intervention -- at least in its current ad-based format -- is not a cost-effective means to improve information environments online. We discuss challenges faced when conducting large-scale experiments on X following the platform's ownership change and subsequent restrictions on data access for research purposes.

cs.SI↗
arXiv2025

Introducing AI to an Online Petition Platform Changed Outputs but not Outcomes

The rapid integration of AI writing tools into online platforms raises critical questions about their impact on content production and outcomes. We leverage a unique natural experiment on Change$.$org, a leading social advocacy platform, to causally investigate the effects of an in-platform ''write with AI'' tool. To understand the impact of the AI integration, we collected 1.5 million petitions and employed a difference-in-differences analysis. Our findings reveal that in-platform AI access significantly altered the lexical features of petitions and increased petition homogeneity, but did not improve petition outcomes. We confirmed the results in a separate analysis of repeat petition writers who wrote petitions before and after introduction of the AI tool. The results suggest that while AI writing tools can profoundly reshape online content, their practical utility for improving desired outcomes may be less beneficial than anticipated, and introduce unintended consequences like content homogenization.

cs.CY↗
arXiv2025

Deep Value Benchmark: Measuring Whether Models Generalize Deep Values or Shallow Preferences

We introduce the Deep Value Benchmark (DVB), an evaluation framework that directly tests whether large language models (LLMs) learn fundamental human values or merely surface-level preferences. This distinction is critical for AI alignment: Systems that capture deeper values are likely to generalize human intentions robustly, while those that capture only superficial patterns in preference data risk producing misaligned behavior. The DVB uses a novel experimental design with controlled confounding between deep values (e.g., moral principles) and shallow features (e.g., superficial attributes). In the training phase, we expose LLMs to human preference data with deliberately correlated deep and shallow features -- for instance, where a user consistently prefers (non-maleficence, formal language) options over (justice, informal language) alternatives. The testing phase then breaks these correlations, presenting choices between (justice, formal language) and (non-maleficence, informal language) options. This design allows us to precisely measure a model's Deep Value Generalization Rate (DVGR) -- the probability of generalizing based on the underlying value rather than the shallow feature. Across 9 different models, the average DVGR is just 0.30. All models generalize deep values less than chance. Larger models have a (slightly) lower DVGR than smaller models. We are releasing our dataset, which was subject to three separate human validation experiments. DVB provides an interpretable measure of a core feature of alignment.

cs.AI↗
arXiv2025

Outsourcing an Information Operation: A Complete Dataset of Tenet Media's Podcasts on Rumble

Tenet Media, a U.S.-based, right-wing media company, hired six established podcasters to create content related to U.S. politics and culture during the 2024 U.S. presidential election cycle. After publishing content on YouTube and Rumble for nearly a year, Tenet Media was declared by the U.S. government to be funded entirely by Russia -- making it effectively an outsourced state-sponsored information operation (SSIO). We present a complete dataset of the 560 podcast videos published by the Tenet Media channel on the video-sharing platform Rumble between November 2023 and September 2024. Our dataset includes video metadata and user comments, as well as high-quality video transcriptions, representing over 300 hours of video content. This dataset provides researchers with material to study a Russian SSIO, and notably on Rumble, which is an understudied platform in SSIO scholarship.

cs.SI↗
arXiv2025

Capital and CHI: Technological Capture and How It Structures CHI Research

This paper advances a theoretical argument about the role capital plays in structuring CHI research. We introduce the concept of technological capture to theorize the mechanism by which this happens. Using this concept, we decompose the effect on CHI into four broad forms: technological capture creates market-creating, market-expanding, market-aligned, and externality-reducing CHI research. We place different CHI subcommunities into these forms -- arguing that many of their values are inherited from capital underlying the field. Rather than a disciplinary- or conference-oriented conceptualization of the field, this work theorizes CHI as tightly-coupled with capital via technological capture. The paper concludes by discussing some implications for CHI.

cs.HC↗
arXiv2024

Plurals: A System for Guiding LLMs Via Simulated Social Ensembles

Recent debates raised concerns that language models may favor certain viewpoints. But what if the solution is not to aim for a 'view from nowhere' but rather to leverage different viewpoints? We introduce Plurals, a system and Python library for pluralistic AI deliberation. Plurals consists of Agents (LLMs, optionally with personas) which deliberate within customizable Structures, with Moderators overseeing deliberation. Plurals is a generator of simulated social ensembles. Plurals integrates with government datasets to create nationally representative personas, includes deliberation templates inspired by deliberative democracy, and allows users to customize both information-sharing structures and deliberation behavior within Structures. Six case studies demonstrate fidelity to theoretical constructs and efficacy. Three randomized experiments show simulated focus groups produced output resonant with an online sample of the relevant audiences (chosen over zero-shot generation in 75% of trials). Plurals is both a paradigm and a concrete system for pluralistic AI. The Plurals library is available at https://github.com/josh-ashkinaze/plurals and will be continually updated.

cs.CL↗
arXiv2024

A Law of One's Own: The Inefficacy of the DMCA for Non-Consensual Intimate Media

Non-consensual intimate media (NCIM) presents internet-scale harm to individuals who are depicted. One of the most powerful tools for requesting its removal is the Digital Millennium Copyright Act (DMCA). However, the DMCA was designed to protect copyright holders rather than to address the problem of NCIM. Using a dataset of more than 54,000 DMCA reports and over 85 million infringing URLs spanning over a decade, this paper evaluates the efficacy of the DMCA for NCIM takedown. Results show that for non-commercial requests, while more than half of URLs are deindexed from Google Search within 48 hours, the actual removal of content from website hosts is much slower. The median infringing URL takes more than 45 days to be removed from website hosts, and only 5.39% URLs are removed within the first 48 hours. Additionally, the most frequently reported domains for non-commercial NCIM are smaller websites, not large platforms. We stress the need for new laws that ensure a shorter time to takedown that are enforceable across big and small platforms alike.

cs.CY↗
arXiv2024

Reporting Non-Consensual Intimate Media: An Audit Study of Deepfakes

Non-consensual intimate media (NCIM) inflicts significant harm. Currently, victim-survivors can use two mechanisms to report NCIM - as a non-consensual nudity violation or as copyright infringement. We conducted an audit study of takedown speed of NCIM reported to X (formerly Twitter) of both mechanisms. We uploaded 50 AI-generated nude images and reported half under X's "non-consensual nudity" reporting mechanism and half under its "copyright infringement" mechanism. The copyright condition resulted in successful image removal within 25 hours for all images (100% removal rate), while non-consensual nudity reports resulted in no image removal for over three weeks (0% removal rate). We stress the need for targeted legislation to regulate NCIM removal online. We also discuss ethical considerations for auditing NCIM on social platforms.

cs.CY↗
arXiv2024

Wikipedia in Wartime: Experiences of Wikipedians Maintaining Articles About the Russia-Ukraine War

How do Wikipedians maintain an accurate encyclopedia during an ongoing geopolitical conflict where state actors might seek to spread disinformation or conduct an information operation? In the context of the Russia-Ukraine War, this question becomes more pressing, given the Russian government's extensive history of orchestrating information campaigns. We conducted an interview study with 13 expert Wikipedians involved in the Russo-Ukrainian War topic area on the English-language edition of Wikipedia. While our participants did not perceive there to be clear evidence of a state-backed information operation, they agreed that war-related articles experienced high levels of disruptive editing from both Russia-aligned and Ukraine-aligned accounts. The English-language edition of Wikipedia had existing policies and processes at its disposal to counter such disruption. State-backed or not, the disruptive activity created time-intensive maintenance work for our participants. Finally, participants considered English-language Wikipedia to be more resilient than social media in preventing the spread of false information online. We conclude by discussing sociotechnical implications for Wikipedia and social platforms.

cs.SI↗
arXiv2024

Seeing Like an AI: How LLMs Apply (and Misapply) Wikipedia Neutrality Norms

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.

cs.CL↗
arXiv2024

The Sociotechnical Stack: Opportunities for Social Computing Research in Non-consensual Intimate Media

Non-consensual intimate media (NCIM) involves sharing intimate content without the depicted person's consent, including "revenge porn" and sexually explicit deepfakes. While NCIM has received attention in legal, psychological, and communication fields over the past decade, it is not sufficiently addressed in computing scholarship. This paper addresses this gap by linking NCIM harms to the specific technological components that facilitate them. We introduce the sociotechnical stack, a conceptual framework designed to map the technical stack to its corresponding social impacts. The sociotechnical stack allows us to analyze sociotechnical problems like NCIM, and points toward opportunities for computing research. We propose a research roadmap for computing and social computing communities to deter NCIM perpetration and support victim-survivors through building and rebuilding technologies.

cs.CY↗
arXiv2024

Feminist Interaction Techniques: Deterring Non-Consensual Screenshots with Interaction Techniques

Non-consensual Intimate Media (NCIM) refers to the distribution of sexual or intimate content without consent. NCIM is common and causes significant emotional, financial, and reputational harm. We developed Hands-Off, an interaction technique for messaging applications that deters non-consensual screenshots. Hands-Off requires recipients to perform a hand gesture in the air, above the device, to unlock media -- which makes simultaneous screenshotting difficult. A lab study shows that Hands-Off gestures are easy to perform and reduce non-consensual screenshots by 67 percent. We conclude by generalizing this approach and introduce the idea of Feminist Interaction Techniques (FIT), interaction techniques that encode feminist values and speak to societal problems, and reflect on FIT's opportunities and limitations.

cs.HC↗
arXiv2024

The Future of Research on Social Technologies: CCC Workshop Visioning Report

Social technologies are the systems, interfaces, features, infrastructures, and architectures that allow people to interact with each other online. These technologies dramatically shape the fabric of our everyday lives, from the information we consume to the people we interact with to the foundations of our culture and politics. While the benefits of social technologies are well documented, the harms, too, have cast a long shadow. To address widespread problems like harassment, disinformation, information access, and mental health concerns, we need to rethink the foundations of how social technologies are designed, sustained, and governed. This report is based on discussions at the Computing Community Consortium Workshop, The Future of Research on Social Technologies, that was held November 2-3, 2023 in Washington, DC. The visioning workshop came together to focus on two questions. What should we know about social technologies, and what is needed to get there? The workshop brought together over 50 information and computer scientists, social scientists, communication and journalism scholars, and policy experts. We used a discussion format, with one day of guiding topics and a second day using an unconference model where participants created discussion topics. The interdisciplinary group of attendees discussed gaps in existing scholarship and the methods, resources, access, and collective effort needed to address those gaps. We also discussed approaches for translating scholarship for various audiences including citizens, funders, educators, industry professionals, and policymakers. This report presents a synthesis of major themes during our discussions. The themes presented are not a summary of what we know already, they are an exploration of what we do not know enough about, and what we should spend more effort and investment on in the coming years.

cs.SI↗
arXiv2024

The Dynamics of (Not) Unfollowing Misinformation Spreaders

Many studies explore how people 'come into' misinformation exposure. But much less is known about how people 'come out of' misinformation exposure. Do people organically sever ties to misinformation spreaders? And what predicts doing so? Over six months, we tracked the frequency and predictors of ~900K followers unfollowing ~5K health misinformation spreaders on Twitter. We found that misinformation ties are persistent. Monthly unfollowing rates are just 0.52%. In other words, 99.5% of misinformation ties persist each month. Users are also 31% more likely to unfollow non-misinformation spreaders than they are to unfollow misinformation spreaders. Although generally infrequent, the factors most associated with unfollowing misinformation spreaders are (1) redundancy and (2) ideology. First, users initially following many spreaders, or who follow spreaders that tweet often, are most likely to unfollow later. Second, liberals are more likely to unfollow than conservatives. Overall, we observe a strong persistence of misinformation ties. The fact that users rarely unfollow misinformation spreaders suggests a need for external nudges and the importance of preventing exposure from arising in the first place.

cs.SI↗