The prevalence of short form video platforms, combined with the ineffectiveness of age verification mechanisms, raises concerns about the potential harms facing children and teenagers in an algorithm-moderated online environment. We conducted multimodal feature analysis and thematic topic modeling of 4,492 short videos recommended to children and teenagers on Instagram Reels, TikTok, and YouTube Shorts, collected as a part of an algorithm auditing experiment. This feature-level and content-level analysis revealed that unsafe (i.e., problematic, mentally distressing) short videos (a) possess darker visual features and (b) contain explicitly harmful content and implicit harm from anxiety-inducing ordinary content. We introduce a useful framework of online harm (i.e., explicit, implicit, unintended), providing a unique lens for understanding the dynamic, multifaceted online risks facing children and teenagers. The findings highlight the importance of protecting younger audiences in critical developmental stages from both explicit and implicit risks on social media, calling for nuanced content moderation, age verification, and platform regulation.

This study explores whether social media platforms’ proprietary recommendation algorithms could possess actionable knowledge about the age and potential vulnerabilities of minors who misrepresent their age during account creation. Our results show that the algorithms of YouTube, Instagram, and TikTok quickly and confidently adjust (unsolicited) recommendations for accounts with behavioral traits from underage users after just one online session. Users behaving like 8-year-olds receive almost seven times more child-directed content than their 16-year-old peers. The tailoring of commercial offerings demonstrates that providers possess sufficient knowledge to prompt action. By law, children under the age of 13 are prohibited from being in a commercial space that collects personal data without parental consent. Algorithms also reacted to accounts behaving like struggling adolescents, which received over 30% more problematic- and over 70% more distressing content than their non-struggling peers. The visible adjustment of content offerings, based on this demonstrable detection ability, raises the question of whether and how this communicated information from the algorithms could be used to enhance protections of minors. This algorithm audit is part of a wider research agenda showcasing the feasibility and value of independent audits to address (often unintended) loopholes stemming from automated algorithmic decision-making processes

Currently, a series of lawsuits and bills are addressing whether social media companies should increase their duty of care toward children aged 13 or younger, as well as vulnerable adolescents. This study investigates whether social media platforms' proprietary recommendation algorithms possess actionable knowledge about the age and potential vulnerabilities of minors who misrepresent their age during account creation, and if so, how this knowledge is utilized in personalizing (unsolicited) recommendations. We demonstrate that after just one online session, YouTube, Instagram, and TikTok recommender algorithms significantly adjust their content recommendations based on user behavior. Our results show how quickly and confidently these algorithms adjust content for underage users, presenting almost seven times more child-directed content to users who exhibit the behavior of 8-year-olds than to the 16-year-old control group. Contrary to expectations, algorithmic insights are not used to protect minors. The proportion of problematic and distressing content shown to those exhibiting child-like behavior was similar to that encountered by adolescents twice their age (12%), while struggling adolescents received over 30% more problematic content and over 70% more distressing content than their same-aged peers in the control group. Our algorithmic audit aligns with audits that, in other areas, are orchestrated by public agencies that oversee the consumption of products and services that could harm users. It is part of a generation of studies that show that such oversight audits are practically feasible and useful to gain insights that could be used to protect minors.

We conducted a quantitatively coarse-grained, but wide-ranging evaluation of the frequency recommender algorithms provide ‘good’ and ‘bad’ recommendations, with a focus on the latter. We found 151 algorithmic audits from 33 studies that report fitting risk-utility statistics from YouTube, Google Search, Twitter, Facebook, TikTok, Amazon, and others. Our findings indicate that roughly 8–10% of algorithmic recommendations are ‘bad’, while about a quarter actively protect users from self-induced harm (‘do good’). This average is remarkably consistent across the audits, irrespective of the platform nor on the kind of risk (bias/ discrimination, mental health and child harm, misinformation, or political extremism). Algorithmic audits find negative feedback loops that can ensnare users into spirals of ‘bad’ recommendations (or being ‘dragged down the rabbit hole’), but also highlight an even larger likelihood of positive spirals of ‘good recommendations’. While our analysis refrains from any judgment of the causal consequences and severity of risks, the detected levels surpass those associated with many other consumer products. They are comparable to the risk levels of generic food defects monitored by public authorities such as the FDA or FSIS in the United States. Consequently, our findings inform the ongoing discussion regarding regulatory oversight of the potential risks posed by recommender algorithms.

This study investigates how anticipating an artificial intelligence agent versus human information source moderates the risk information seeking and processing model. It focuses on a behavioral proxy of seeking intention—how long a participant waited for an online consultant whose identity was manipulated. In two samples ( N 1 = 182 students and N 2 = 800 mturkers), the source identity consistently moderated the model in two ways: First, informational subjective norms encouraged seeking from humans but discouraged seeking from AI agents. Second, information insufficiency drove favoritism toward humans–when perceived information-gathering capacity was high. When the capacity was low, AI agents were favored.

We obtain a quantitatively coarse-grained, but wide-ranging evaluation of the frequency recommender algorithms provide ‘good’ and ‘bad’ recommendations. We found 146 algorithmic audits from 32 studies that report fitting risk-utility statistics from YouTube, Google Search, Twitter, Facebook, TikTok, Amazon, and others. The vast majority of algorithmic recommendations do no harm (around 90 %), while about a quarter of recommendations safeguard humans from self-induced harm (‘do good’). The frequency of ‘bad’ recommendations is around 7-10 % on average, which poses a potential risk that is notably higher than risks posed by other consumer products. This average is remarkably robust across the audits and neither depends on the platform nor on the kind of harm (bias/ discrimination, mental health and child harm, misinformation, or political extremism). Algorithmic audits find negative feedback loops that lock users into spirals of ‘bad’ recommendations (or being ‘dragged down the rabbit hole’), but they find an even larger probability of positive spirals of ‘good recommendations’. Our analysis refrains from any qualitative judgment of the severity of different harms. As concerns for ‘AI alignment’ with human values grow, necessitating more algorithmic audits, our study offers preliminary figures for quantitative comparison to other contemporary risks of modern life.

We hypothesize that individuals with psychological traits typical for self-transcended individuals show natural resistance to psychological harms from online social media. We theorize that this 'digital immunity' is a manifestation of people's consciousness development.Social media’s persuasive technologies rely on the predictability of an un-reflective or automatic reaction to a previously identified digital stimulus. Tailor-made triggers of predictable responses are often identified with machine learning based on big data traces. This mechanism works best when the user unconsciously follows the stimulus-response chain. The psychological concept of self-transcendence characterizes a later stage in the developmental evolution of human consciousness that is thought to be highly correlated with the exhibition of self-reflection/self-knowing, the realization of interconnectedness of the self and other complex adaptive systems, an expansion of personal boundaries through a nondual understanding of reality, and a present moment awareness that provides cognitive space (less dependence) with causal event sequences in time. We hypothesized that this autonomy lessens conditioning on unconscious stimuli-response pairing. In other words, we hypothesized that individuals who started to transcend the self-identification with their own mind (self-identified thoughts and emotions) have evolved a natural immunity to digital manipulations.This report shows the results of the first study on “Digital Immunity”, consisting of a 1.5h online experiment with 1,164 high quality contributions, collected between May 4 and November 24, 2021. It includes the background information for all measured variables, a plan for analysis (in form of an integrative modelling framework), and theoretical explorations on the underlying proposed generative mechanism (self-transcendence), all of which had already been published on November 5, 2021, prior to any human observation of the data (part of the data existed but had not yet been quantified, constructed, analyzed, or reported by anyone) and can found here:Hilbert, M., Thakur, A., Repetto, A., & Weisman, W. (2021). Supporting Information for “Digital Immunity…”, Nov. 5, 2021, (SSRN Scholarly Paper No. 3957595). https://doi.org/10.2139/ssrn.3957595This current report includes the analysis and (most of the) results of the data collection of the 2021 survey experiment (N = 1,164), including our exploratory and confirmatory analyses, following an integrative modelling framework. It ends by preregistering a follow-up study, currently ongoing.

Although artificial intelligence is blamed for many societal challenges, it also has underexplored potential in political contexts online. We rely on six preregistered experiments in three countries (N = 6,728) to test the expectation that AI and AI-assisted humans would be perceived more favorably than humans (a) across various content moderation, generation, and recommendation scenarios and (b) when exposing individuals to counter-attitudinal political information. Contrary to the preregistered hypotheses, participants see human agents as more just than AI across the scenarios tested, with the exception of news recommendations. At the same time, participants are not more open to counter-attitudinal information attributed to AI rather than a human or an AI-assisted human. These findings, which-with minor variations-emerged across countries, scenarios, and issues, suggest that human intervention is preferred online and that people reject dissimilar information regardless of its source. We discuss the theoretical and practical implications of these findings. Lay Summary In the era of unprecedented political divides and misinformation, artificial intelligence (AI) and algorithms are often seen as the culprits. In contrast to these dominant narratives, we argued that AI might be seen as being less biased than a human in online political contexts. We relied on six preregistered experiments in three countries (the United Sates, Spain, Poland) to test whether internet users perceive AI and AI-assisted humans more favorably than simply humans; (a) across various distinct scenarios online, and (b) when exposing people to opposing political information on a range of contentious issues. Contrary to our expectations, human agents were consistently perceived more favorably than AI except when recommending news. These findings suggest that people prefer human intervention in most online political contexts.

Thanks to advanced sensing and logging technology, automatic personality assessment (APA) with users' behavioral data in the workplace is on the rise. While previous work has focused on building APA systems with high accuracy, little research has attempted to understand users' perception towards APA systems. To fill this gap, we take a mixed-methods approach: we (1) designed a survey (n=89) to understand users'social workplace behavior both online and offline and their privacy concerns; (2) built a research probe that detects personality from online and offline data streams with up to 81.3% accuracy, and deployed it for three weeks in Korea (n=32); and (3) conducted post-interviews (n=9). We identify privacy issues in sharing data and system-induced change in natural behavior as important design factors for APA systems. Our findings suggest that designers should consider the complex relationship between users' perception and system accuracy for a more user-centered APA design.