Arvind Narayanan's research while affiliated with Princeton University and other places

... Separately, the adoption of benchmarking best practices from predictive modeling fields, including machine learning (Weber et al., 2019;Mangul et al., 2019;Mitchell et al., 2019;Kapoor and Narayanan, 2023), can help facilitate progress in explanatory modeling. The following list describes three important examples of these practices: ...

... At the extreme, some of these forms of benefit-such as the economic reward of selling AI systems-are consistent with the proliferation of AI applications that don't actually perform their stated task and do not provide a direct benefit to anyone impacted by their operation (Raji et al., 2022). This prospect echoes recent criticisms that frameworks for ethical and responsible AI are too centrally focused on the values and interests of technology creators (Birhane et al., 2022) and on the way that relevant moral values are instantiated within technologies rather than on the larger impacts of those technologies (Wang et al., 2022;Laufer et al., 2022), including questions of authority and power dynamics, human sovereignty and autonomy, and democratic values like liberty and freedom (Washington and Kuo, 2020;Barabas et al., 2020). * Authors contributed equally to this work. ...

... For example, Mathur et al. use structural topic modeling to study manipulative tactics in a corpus of U.S. political campaign emails from the 2020 election cycle. One of their findings is that the median active sender of campaign emails uses "sensationalist clickbait" in 37% of those emails [167]. A question about the external validity of this claim might be: is the frequency of sensationalist clickbait in campaign emails similar in other U.S. election cycles? ...

... We first examine the proposed methods capability of improving node classification tasks. We run each experiment with three trial runs and report the average test set F1-score of all 5 GNNs and random forest models in the timesteps from [34][35][36][37][38][39][40][41][42][43][44][45][46][47][48][49] separated into four intervals. The best results in each model group are underlined and shown in bold in Table 1. ...

... Compared to older statistical methods, they offer increased predictive accuracy [1], the ability to process large amounts of data [12], and the ability to use different types of data for scientific research, such as text, images, and video [7]. However, the rapid uptake of ML methods has been accompanied by concerns of validity, reproducibility, and generalizability [13][14][15][16][17][18][19]. There are several reasons for concern. ...

... However, this resulting uncertainty subsequently erodes trust in news disseminated via social media. Papakyriakopoulos et al. [29] also offer a large-scale data analysis of publicly available data, aiming to provide a critical assessment of how these platforms either amplified or moderated the distribution of political advertisements. Additionally, Ma et al. [20] engage in a conceptual replication and extension of this influential study. ...

... While algorithmic bias mitigation is well explored (Barocas et al. 2019;Ntoutsi et al. 2020;Mitchell et al. 2021;Mehrabi et al. 2021), few works tackle bias detection, especially in the CV domain. A common strategy (Fabbrizzi et al. 2022) is to extract information from the visual data, arrange it into tabular form, and quantify and mitigate bias on this data type (Zhao et al. 2017;Buolamwini and Gebru 2018;Merler et al. 2019;Wang et al. 2022). However, the reduction does not necessarily preserve all biased characteristics of the visual data space. ...

... The authors investigated the readability and text features from 2009 to 2019. The authors concluded that the privacy policies are long, hard to read, and lack transparency in some areas such as tracking cookies and data sharing with third parties [28]. Bateni compared the privacy policies before and after the General Data Protection Regulation (GDPR) [24,27,29]. ...

... Fairness Toolkits. Several analysis tools [3,8,14,17,18,37] exist to discover biases in datasets and models. For example, Google's Know Your Data [8] focuses on discovering correlations between attribute labels and metadata labels in TensorFlow image datasets. ...

... With recent [2], [1] and imminent restrictions [34] on third-party cookies, trackers are moving towards alternative methods of identifying users, which include the use of firstparty cookies [10], [8], [7], personally identifiable information (PII) [35], [36], and device/browser fingerprinting [37], [38], [39]. In contrast to third-party cookies that are automatically included as headers in outgoing HTTP requests, these alternative tracking methods rely on link decoration for identifier exfiltration. ...