Cognition in Online Environments

Reviewing and Testing Individual-Level Interventions Against Misinformation
Understanding Susceptibility to Online Misinformation
Improving Public Discourse Online
Hypercustomization in Generative Artificial Intelligence: Risks and Mitigation Strategies
Delegating to AI Can Erode Ethics
Arresting Democratic Backsliding Driven by Elite Norm Violations
Social Media and the Mental Health of Children and Adolescents: The Leopoldina Discussion Paper (2025)

Reviewing and Testing Individual-Level Interventions Against Misinformation

In today's digital age, misinformation spreads rapidly. The rise of false and misleading information online can undermine trust in democratic institutions, increase social discord, and threaten public health. As these challenges mount, researchers and policymakers are focusing increasingly on finding effective ways to counteract misinformation. What can individuals do to protect themselves against online misinformation?

Kozyreva et al. (2024) developed a toolbox of individual-level interventions against online misinformation (Figure 1). It provides a review of behavioral and cognitive interventions to counter online misinformation and includes empirically validated tools designed to equip individuals with the skills and tools necessary to identify, resist, and refute misinformation. Based on findings from 81 scientific papers, this resource provides a detailed overview of different types of interventions, their effectiveness, and the contexts in which they have been tested. Each intervention serves a unique purpose in helping users to navigate the complex digital landscape. The interventions range from competence-building boosts, such as lateral reading, media literacy tips, and verification strategies, to strategies that require the involvement of experts, policymakers, or platforms, such as debunking, accuracy prompts, and fact-checking labels.

ARC is actively testing boosting tools designed to help users verify information themselves, including source-focused lateral reading and claim-focused online search strategies. Ongoing work also investigates how self-selection into interventions shapes their effectiveness, and whether hybrid approaches combining AI with human crowds can reliably distinguish false from true news.

Key Reference

Understanding Susceptibility to Online Misinformation

Despite a large body of research on misinformation, it is unclear who is susceptible to it and why. Sultan et al. (2024) conducted a systematic participant-level meta-analysis covering over a quarter million ratings made by over 11,500 US participants across 31 studies (Figure 2). The analysis distinguishes between two cognitive constructs: discrimination ability (whether a person can distinguish true from false news) and response bias (whether a person is generally biased to label news as true or false).

Discrimination Ability and Response Bias

Older adults, Democrats (compared to Republicans), and individuals with higher analytical thinking skills demonstrated greater discrimination ability. Contrary to common assumptions, individuals with higher education did not. But discrimination ability is only half the picture. The authors found that ideological congruency—when news aligns with a user's politics—and familiarity with news are associated with a "true-news bias": If something is ideologically congruent and/or familiar, you are more likely to accept it as true regardless.

The authors also settled a longstanding debate in the field, finding small but consistent evidence for motivated reflection. Individuals with higher analytical thinking skills and individuals who exhibited more skepticism fell prey to stronger congruency effects. Paradoxically, cognitive sophistication is associated with protecting one's own partisan beliefs, leaving people more entrenched in their partisan beliefs.

Adolescents' Unique Potential in the Face of Misinformation

A recent perspective (Ma et al., 2026) argues that today's digital natives require distinct research attention: Adolescents, social media's most avid users, face unique susceptibilities to misinformation given their rapid social, cognitive, and emotional development. At the same time, adolescents also bring distinct strengths that interventions could leverage, including heightened open-mindedness, a drive for autonomy, and motivation to master digital environments.

Misinformation and Metacognition

Another critical question is how well people know whether their judgments are actually correct or incorrect (confidence and metacognitive insight) and how this differs across people. In a large representative US sample, Geers et al. (2024) found that the metacognitive performance of Republicans was substantially lower than that of Democrats when the information in a headline challenged their ideology. This finding highlights the role of metacognition in perpetuating and exacerbating ideological divides.

Key References

Improving Public Discourse Online

Online discussions are often shaped by a small, highly active minority, while most users remain silent. This dynamic can distort perceptions of public opinion and intensify polarization. Yet little is known about what drives this imbalance, whether silent users can be activated, and how the resulting participation gaps shape perceptions of public opinion.

In a group-based field experiment on Reddit, Oswald et al. (2025) investigated what distinguishes silent "lurkers" from highly active "power users," and which interventions might encourage broader participation (see Figure 3). The authors created six private subreddits, each with up to 100 participants who discussed 20 political topics over 4 weeks and completed weekly surveys. Participants who perceived discussions as toxic, disrespectful, or highly polarized were more likely to remain silent. Paradoxically, these same conditions were associated with higher activity among already active users, suggesting that heated environments further amplify the voices of the few. The most active participants were more likely to be male, highly interested in politics, and predisposed to comment online.

The study also tested interventions to activate lurkers. Financial incentives increased participation and positive feedback through upvotes was linked to higher next-day activity. Norm-based appeals had little effect. The goal of these interventions was not to maximize participation at all costs, but to examine how the dominance of a small number of highly active users distorts perceived public opinion for both individuals and researchers. Comment sections often give a misleading impression of what most people think.

In a related study, Oswald (2025) used a survey experiment to examine whether behavioral interventions can reduce the toxicity of replies. Consistent with the limited effects of norm appeals in the Reddit field experiment, simple nudges to engage in empathy and to take different perspectives when confronted with opposing political views did not reduce the toxicity of replies. More complex boosting interventions were more promising, highlighting the role of political education in an age of rapidly transforming information environments.

Key References

Hypercustomization in Generative Artificial Intelligence: Risks and Mitigation Strategies

Generative Artificial Intelligence (GenAI) is transforming how people seek information: through search engines, social media, and GenAI-powered chatbots. Similarly to how social media algorithms personalize online experiences (e.g., by tailoring feed-ranking algorithms based on users' interaction with the feed), GenAI systems can align with users' preferences to customize how they respond. However, through training, fine-tuning, and prompting, GenAI systems can introduce a qualitatively stronger form of customization: hypercustomization.

By adaptively tailoring their responses to users' explicit and implicit preferences (e.g., as revealed through users' writing), hypercustomization can reinforce people's preconceptions or false beliefs because GenAI systems, by default, tend to agree with users (i.e., AI sycophancy). It can also provide bespoke responses (rather than, e.g., simply ranking posts in a news feed). This creates a feedback loop that can reinforce preconceptions and validate falsehoods, potentially accelerating misinformation spread and deepening societal polarization.

Two recent papers (Abels et al., 2025; Lopez-Lopez et al., 2025) conceptually explore the challenges and consequences of GenAI hypercustomization, as well as behavioral and regulatory approaches to mitigating the risks posed by hypercustomization (Figure 4). Challenges include a lack of transparency, the opaque nature of user–GenAI interactions, users' overreliance on GenAI, and the limited efficacy of warning messages. Proposed mitigation strategies include public black-box testing, issue-reporting platforms, data donations, boosting GenAI literacy, and in-application prompts that encourage users to reflect on AI-generated content. The authors also urge the research and policy community to proactively engage with those challenges and not repeat the slow and reactive response to earlier challenges posed by social media.

Key References

Delegating to AI Can Erode Ethics

As AI increasingly permeates everyday life, AI companies are developing and promoting the use of agentic AI systems that autonomously execute tasks on people's behalf. The promise is that such agents will simplify people's lives by completing mundane tasks like sending emails, booking restaurants, and managing schedules. But AI companies will not stop at mundane tasks—indeed, AI agents are already integrated into platforms offering higher-stakes services and activities, including generating tax returns and investing finances. As AI agents proliferate, a stronger understanding is needed of how they impact ethical behavior. Currently, little is known about how people's ethical intentions shift when delegating to AI compared to doing the task themselves. Furthermore, to what extent would current AI agents comply with unethical instructions?

A team of researchers co-led by ARC's Zoe Rahwan (Köbis et al., 2025) investigated these questions in a study in which human participants instructed AI agents to perform different tasks on their behalf where there were incentives to cheat. The authors found that while a minority (5% of people) were dishonest when completing a task themselves, this number was substantially higher when people delegated to an AI agent (Figure 5). The way people programmed the agent mattered: Setting specific rules caused a 5-fold increase in dishonesty, and setting a high-level goal (e.g., "maximize profit")—which can obfuscate unethical intentions—encouraged the most dishonesty, causing a 17-fold increase.

In separate experiments where people instructed AI agents using natural language prompts, the authors identified a separate ethical risk: AI compliance with unethical instructions. Despite inbuilt guardrails, instructions given to leading commercial models at the time (GPT‑4, Claude, and Llama) complied with unambiguous requests to cheat in the majority of cases. This contrasted sharply with humans, who resisted requests to cheat. Even with additional guardrails, compliance with fully unethical requests could not always be eliminated, and the most effective guardrails proved impractical.

As agentic AI scales into classrooms, workplaces, and public services, understanding and mitigating these dynamics will be crucial to preventing a quiet drift toward normalized, AI-mediated dishonesty.

Key Reference

Arresting Democratic Backsliding Driven by Elite Norm Violations

Despite a substantial body of research on democratic backsliding, the mechanisms through which democracies erode—particularly the gradual concentration of power in the executive and the weakening of democratic institutions—are not well understood. To address this gap, Abels et al. (2024) examined democratic near misses: cases in which countries briefly drifted toward autocratic governance but either rapidly reversed course or avoided full backsliding at the last moment.

Drawing on near-miss analysis in sociotechnical systems (e.g., nuclear power plants), the authors adapted the drift-to-danger model—where failures emerge through slow, often imperceptible processes of normalization and boundary-pushing rather than sudden collapse—to democratic systems. Democratic backsliding typically unfolds incrementally, leading not only to the erosion of free and fair elections but also to the degradation of core democratic freedoms, including freedom of speech, assembly, and the press, as well as protections for individual and minority rights (Figure 6).

Across all near-miss cases, elite norm violations were a central driver of backsliding. Because democratic systems heavily depend on unwritten norms where formal rules leave gaps, they are vulnerable to elites who deliberately test or violate these normative constraints.

Democratic backsliding, however, is neither linear nor predictable. Rather, it emerges from dynamic interactions between political elites and the public. Elites who violate democratic norms may feel emboldened by public support that frames such violations as justified or necessary. Conversely, elite behavior can signal the acceptability of norm violations, potentially fostering anti-democratic beliefs and behaviors among citizens.

The authors identified key risk factors of democratic backsliding—including right-wing populism, voter disenchantment, misinformation, and rising social and political polarization—and outlined interventions informed by behavioral science that aim at strengthening democratic resilience and preventing democratic drift toward autocracy, such as inoculation against misinformation, intergroup contact programs, and experiential simulations that make the risks of democratic decline tangible.

Key References

Social Media and the Mental Health of Children and Adolescents: The Leopoldina Discussion Paper (2025)

The impact of social media on the mental health of children and adolescents is a major concern in Germany and worldwide. In response, the German National Academy of Sciences Leopoldina published a discussion paper in August 2025 that examines this issue and formulates concrete policy recommendations for German policy makers. Led by Leopoldina members Ralph Hertwig and Johannes Buchmann, the initiative included an interdisciplinary group of scholars from psychology, computer science, communication science, philosophy, and law.

Drawing on the current state of research, the paper shows that while much of the available evidence is correlational and does not yet establish clear causal pathways, a robust association exists between intensive—particularly risky or addictive—social media use and increased mental health burdens among children and adolescents. Findings from longitudinal studies and intervention research further suggest that social media use may play a causal role in mental health issues among intensive users.

Given the potential severity and irreversibility of mental health impairments in children and teenagers, as well as the rapid pace of technological innovation, the authors argue for the application of the precautionary principle: Preventive policy action is warranted even without definitive causal proof. The paper also stresses the need for continued research to better understand risks and to assess the effectiveness of protective interventions.

Key recommendations include prohibiting social media use for children under 13 years of age, introducing privacy-preserving age verification (e.g., via the European Digital Identity Wallet) and requiring age-appropriate platform design for users aged 13 to 17 years. Complementary measures include restricting smartphone use in schools, comprehensive media literacy education from kindergarten through secondary education, and public health campaigns targeting families and adults who regularly interact with children and adolescents, such as educators and sports clubs.

Beyond protection, the report aims to empower young people to engage with social media in a competent, reflective, and sovereign manner. Since its publication, the discussion paper has received wide public attention in Germany among parents, educators, and policymakers.

Key Reference