Mohsen Mosleh
Current Organizations / Institutes
The organizations/institutes from the scholar's latest article
Research Subjects
Research Interests
Mohsen Mosleh has a total of 66 co-authors. Here are the top co-authors by number of shared publications:
- David Gertler Rand
Massachusetts Institute of Technology
45 shared publications - Gordon Pennycook
Department of Psychology, Cornell University
11 shared publications - Cameron Martel
Affiliation Sloan School of Management, Massachusetts Institute of Technology, Cambridge, Massachusetts, United States of America
9 shared publications - Babak Heydari
Northeastern University
6 shared publications Dean Eckles
5 shared publicationsQi Yang
Initiative on the Digital Economy, Massachusetts Institute of Technology, Cambridge, USA
5 shared publications
Latest Publications
- Open Access
- 13 November 2026
- Published by Center for Open Science in PsyArXiv
The Confidence – Information Quality Mismatch: Low Quality Posts Use More Confident Language
Misinformation on online platforms is of great concern for democracy, public health, and civic discourse. In fast-paced feeds, users often rely on heuristics when deciding what to attend to, believe, and share. One such heuristic is confidence, which, prior research shows, is treated as a cue of knowledge and credibility. Yet, we show that in over 4 million posts across seven social media platforms and one messaging app, confidence is negatively associated with information quality: posts written with greater confidence are more likely to link to lower-quality news domains and to contain false claims — even controlling for user baselines, toxicity, political lean, emotionality, and valence. Furthermore, we find that confident posts receive more engagement. Together, these results reveal a problematic asymmetry in online information environments: a widespread linguistic cue that is likely to be read as a signal of knowledge, and to attract engagement, is in fact associated with lower-quality information.
- 13 November 2026
- Published by Center for Open Science in PsyArXiv
Open AccessThe Confidence – Information Quality Mismatch: Low Quality Posts Use More Confident Language
Misinformation on online platforms is of great concern for democracy, public health, and civic discourse. In fast-paced feeds, users often rely on heuristics when deciding what to attend to, believe, and share. One such heuristic is confidence, which, prior research shows, is treated as a cue of knowledge and credibility. Yet, we show that in over 4 million posts across seven social media platforms and one messaging app, confidence is negatively associated with information quality: posts written with greater confidence are more likely to link to lower-quality news domains and to contain false claims — even controlling for user baselines, toxicity, political lean, emotionality, and valence. Furthermore, we find that confident posts receive more engagement. Together, these results reveal a problematic asymmetry in online information environments: a widespread linguistic cue that is likely to be read as a signal of knowledge, and to attract engagement, is in fact associated with lower-quality information.
- Open Access
- 2 November 2026
- Published by Center for Open Science in PsyArXiv
Can AI Scale Human Fact-Checking? Evidence from Community Notes
Community Notes---a crowd-based fact-checking program through which social media users can flag a post with a note explaining why the post is misleading, and the note is displayed if a diverse set of users rate the note as helpful---has become central to online content moderation. The program, however, faces substantial scalability constraints, as the generation of notes takes considerable time and effort, and the number of user participants is limited. Here, we investigate whether allowing AI models to write Community Notes helps to alleviate these constraints by analyzing the full corpus of AI-generated Community Notes on X through May 2026. Encouragingly, AI-generated notes are submitted more quickly, are longer, and cite more and higher-quality sources than human-written notes, while being rated helpful at a similar rate and being associated with similar decreases in reposts following the publication of a rated-helpful note. Furthermore, a pre-registered survey experiment finds no statistically detectable difference between AI- and human-generated fact-checks in their effect on belief updating, regardless of whether the author was revealed to participants. On the other hand, AI-generated note writing is highly concentrated, as a handful of accounts produce the large majority of notes, and are significantly less likely to engage with posts involving controversial and polarized topics such as politics and elections. While those findings suggest that AI may be a powerful tool for scaling crowd-based fact-checking systems such as Community Notes, this evolution also entails important risks: shifting the primary bottleneck from note production to note evaluation, moving the system from a decentralized toward a more centralized model, and creating incentives to prioritize claims that are more likely to achieve consensus over those of greater societal importance.
- 2 November 2026
- Published by Center for Open Science in PsyArXiv
Open AccessCan AI Scale Human Fact-Checking? Evidence from Community Notes
Community Notes---a crowd-based fact-checking program through which social media users can flag a post with a note explaining why the post is misleading, and the note is displayed if a diverse set of users rate the note as helpful---has become central to online content moderation. The program, however, faces substantial scalability constraints, as the generation of notes takes considerable time and effort, and the number of user participants is limited. Here, we investigate whether allowing AI models to write Community Notes helps to alleviate these constraints by analyzing the full corpus of AI-generated Community Notes on X through May 2026. Encouragingly, AI-generated notes are submitted more quickly, are longer, and cite more and higher-quality sources than human-written notes, while being rated helpful at a similar rate and being associated with similar decreases in reposts following the publication of a rated-helpful note. Furthermore, a pre-registered survey experiment finds no statistically detectable difference between AI- and human-generated fact-checks in their effect on belief updating, regardless of whether the author was revealed to participants. On the other hand, AI-generated note writing is highly concentrated, as a handful of accounts produce the large majority of notes, and are significantly less likely to engage with posts involving controversial and polarized topics such as politics and elections. While those findings suggest that AI may be a powerful tool for scaling crowd-based fact-checking systems such as Community Notes, this evolution also entails important risks: shifting the primary bottleneck from note production to note evaluation, moving the system from a decentralized toward a more centralized model, and creating incentives to prioritize claims that are more likely to achieve consensus over those of greater societal importance.
- Open Access
- 30 September 2026
- Published by Center for Open Science in PsyArXiv
Fact-Checking Using AI
Large language models (LLMs) can both correct inaccurate beliefs (1, 2, 3) and manipulate attitudes (4, 5, 6). As LLMs are increasingly deployed as fact-checkers on social media, their impact depends not only on the models’ technical capability, but also on the choices of their designers and the attitudes and behaviors of platform users. Here, we shed light on all three dimensions of this sociotechnical system. We analyze an exhaustive sample of 1,555,281 English-language fact-check requests sent to Grok and Perplexity bots on X from February through September 2025 and a random sample of 184,124 fact-checks from Grok from September 2025 to August 2026. We also benchmark 19 frontier LLMs against professional fact-checkers’ evaluations of hundreds of news headlines, and conduct a preregistered survey experiment (N=1,592) to evaluate correction efficacy. On the user side, Grok requesters skew heavily Republican; familiarity with and trust in specific LLMs vary substantially, such that trust in AI is no longer a unitary construct; and revealing that a fact-check comes from Grok reduces belief updating among Democrats while increasing belief updating among Republicans. Examining LLM outputs, we find that from a technical perspective, LLM fact-checks can match or exceed agreement among professional fact-checkers. Yet we also show that this capacity is readily undermined by designers’ choices: for example, restricting sources (e.g. Truth Social AI) or disabling search can degrade accuracy. Designers and users, more so than technical capability, are now the primary binding constraints.
- 30 September 2026
- Published by Center for Open Science in PsyArXiv
Open AccessFact-Checking Using AI
Large language models (LLMs) can both correct inaccurate beliefs (1, 2, 3) and manipulate attitudes (4, 5, 6). As LLMs are increasingly deployed as fact-checkers on social media, their impact depends not only on the models’ technical capability, but also on the choices of their designers and the attitudes and behaviors of platform users. Here, we shed light on all three dimensions of this sociotechnical system. We analyze an exhaustive sample of 1,555,281 English-language fact-check requests sent to Grok and Perplexity bots on X from February through September 2025 and a random sample of 184,124 fact-checks from Grok from September 2025 to August 2026. We also benchmark 19 frontier LLMs against professional fact-checkers’ evaluations of hundreds of news headlines, and conduct a preregistered survey experiment (N=1,592) to evaluate correction efficacy. On the user side, Grok requesters skew heavily Republican; familiarity with and trust in specific LLMs vary substantially, such that trust in AI is no longer a unitary construct; and revealing that a fact-check comes from Grok reduces belief updating among Democrats while increasing belief updating among Republicans. Examining LLM outputs, we find that from a technical perspective, LLM fact-checks can match or exceed agreement among professional fact-checkers. Yet we also show that this capacity is readily undermined by designers’ choices: for example, restricting sources (e.g. Truth Social AI) or disabling search can degrade accuracy. Designers and users, more so than technical capability, are now the primary binding constraints.
- Open Access
- 29 June 2026
- Published by Oxford University Press (OUP) in PNAS Nexus
- Vol. 5 (7)
Confident judgments of (mis)information veracity are more, rather than less, accurate
Does confidence help or hinder the recognition of misinformation? Prior work has reached opposing conclusions, in part because it has not separated confidence in a specific judgment from confidence in one's abilities in general. Here, we separate these constructs and test them side-by-side. In a large, preregistered study on Lucid (n = 503) where participants judged the accuracy of news headlines, confidence in specific judgments predicted greater accuracy. In contrast, general confidence predicted greater “inaccuracy.” In a preregistered prolific replication (n = 498), the confidence-in-judgment effect replicated, whereas the confidence-in-general effect was less consistent. Together, these results indicate that when people say they are confident in a specific judgment, they tend to be right—but people who are generally confident in themselves are no better at spotting misinformation and may even be worse.
- 29 June 2026
- Published by Oxford University Press (OUP) in PNAS Nexus
- Vol. 5 (7)
Open AccessConfident judgments of (mis)information veracity are more, rather than less, accurate
Does confidence help or hinder the recognition of misinformation? Prior work has reached opposing conclusions, in part because it has not separated confidence in a specific judgment from confidence in one's abilities in general. Here, we separate these constructs and test them side-by-side. In a large, preregistered study on Lucid (n = 503) where participants judged the accuracy of news headlines, confidence in specific judgments predicted greater accuracy. In contrast, general confidence predicted greater “inaccuracy.” In a preregistered prolific replication (n = 498), the confidence-in-judgment effect replicated, whereas the confidence-in-general effect was less consistent. Together, these results indicate that when people say they are confident in a specific judgment, they tend to be right—but people who are generally confident in themselves are no better at spotting misinformation and may even be worse.
- Open Access
- 25 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 5546-5569
Fact-Checkers Navigating Generative AI: Practices, Boundaries, and Design Implications
- 25 June 2026
- Published by Association for Computing Machinery (ACM)
- p. 5546-5569
Open AccessFact-Checkers Navigating Generative AI: Practices, Boundaries, and Design Implications
- Open Access
- 19 May 2026
- Published by Center for Open Science in PsyArXiv
Human-AI dialogues can reduce feelings of partisan animosity
Affective polarization – and especially dislike of opposing partisans – is a source of major concern. Consequently, there is great interest in identifying effective approaches for depolarization. Here, we ask whether recent advances in artificial intelligence (AI), namely the development of conversational large language models, can be leveraged to reduce affective polarization. Prior work suggests that such AI-powered chatbots can be powerful tools for changing factual beliefs. However, it remains unclear whether dialogues with an AI chatbot can reduce feelings of animosity towards members of one’s political outgroup. To answer this question, we assign N = 3,874 US-based participants to one of three treatments in which an AI chatbot tries to reduce affective polarization, to a control where they have an irrelevant dialogue with the AI, or to a benchmark condition consisting of the top-performing non-AI intervention identified in a recent mega-study. We find that all of the human-AI dialogue treatments significantly reduce animosity towards opposing partisans and discomfort with having opposing partisans as neighbors or close friends, with the most effective AI treatment achieving treatment effects roughly 70% the size of the effect of the top-performing non-AI intervention benchmark. In addition, we find that the human-AI dialogues increase the likelihood that individuals reciprocate social ties with hypothetical social media profiles associated with their political outgroup by approximately five percentage points, performing at least as well as the non-AI intervention benchmark. We find that supplying the AI chatbot with specific facts about the outgroup increases effectiveness. These findings indicate that providing factual information is an important route to depolarization, and that AI chatbots offer an effective approach for reducing affective polarization that has the potential to be deployed at scale online.
- 19 May 2026
- Published by Center for Open Science in PsyArXiv
Open AccessHuman-AI dialogues can reduce feelings of partisan animosity
Affective polarization – and especially dislike of opposing partisans – is a source of major concern. Consequently, there is great interest in identifying effective approaches for depolarization. Here, we ask whether recent advances in artificial intelligence (AI), namely the development of conversational large language models, can be leveraged to reduce affective polarization. Prior work suggests that such AI-powered chatbots can be powerful tools for changing factual beliefs. However, it remains unclear whether dialogues with an AI chatbot can reduce feelings of animosity towards members of one’s political outgroup. To answer this question, we assign N = 3,874 US-based participants to one of three treatments in which an AI chatbot tries to reduce affective polarization, to a control where they have an irrelevant dialogue with the AI, or to a benchmark condition consisting of the top-performing non-AI intervention identified in a recent mega-study. We find that all of the human-AI dialogue treatments significantly reduce animosity towards opposing partisans and discomfort with having opposing partisans as neighbors or close friends, with the most effective AI treatment achieving treatment effects roughly 70% the size of the effect of the top-performing non-AI intervention benchmark. In addition, we find that the human-AI dialogues increase the likelihood that individuals reciprocate social ties with hypothetical social media profiles associated with their political outgroup by approximately five percentage points, performing at least as well as the non-AI intervention benchmark. We find that supplying the AI chatbot with specific facts about the outgroup increases effectiveness. These findings indicate that providing factual information is an important route to depolarization, and that AI chatbots offer an effective approach for reducing affective polarization that has the potential to be deployed at scale online.
- Open Access
- 24 April 2026
- Published by Center for Open Science in PsyArXiv
Different ideology, same psychology: Evidence for shared psychological pathways to violence across three extremist movements
Violent extremist movements differ dramatically in ideology, target audiences, and rhetorical style, yet converging evidence suggests that they may share underlying psychological processes of radicalisation towards violent outcomes. Testing this hypothesis at scale has been limited by the difficulty of simultaneously measuring ideological content and psychological processes across multiple movements. Here, we analyse 742,264 messages from 42 Telegram channels spanning three ideologically distinct extremist movements: 764 (a youth-targeting sextortion network), Terrorgram (accelerationist neo-Nazi channels), and white supremacist channels. Topic modelling reveals that the three movements discuss fundamentally different subjects, with distinct vocabulary and thematic profiles. In sharp contrast, annotation of the Deadly Cocktail Model's four psychological process indicators (identity fusion, outgroup othering, existential threat perception, and violence-condoning norms) reveals that all four are present across every movement, co-occur within messages far more than expected by chance, and follow an identical rank ordering. Temporal analyses reveal parallel dynamics and directed dependencies between these processes across movements, too. In a fragmented and fast-evolving ideological landscape, where ideological affiliation alone is an increasingly weak predictor of violence, we argue that process-based indicators, grounded in psychological theory and operationalised computationally, offer a more generalisable framework for understanding and detecting radicalisation across ideological boundaries.
- 24 April 2026
- Published by Center for Open Science in PsyArXiv
Open AccessDifferent ideology, same psychology: Evidence for shared psychological pathways to violence across three extremist movements
Violent extremist movements differ dramatically in ideology, target audiences, and rhetorical style, yet converging evidence suggests that they may share underlying psychological processes of radicalisation towards violent outcomes. Testing this hypothesis at scale has been limited by the difficulty of simultaneously measuring ideological content and psychological processes across multiple movements. Here, we analyse 742,264 messages from 42 Telegram channels spanning three ideologically distinct extremist movements: 764 (a youth-targeting sextortion network), Terrorgram (accelerationist neo-Nazi channels), and white supremacist channels. Topic modelling reveals that the three movements discuss fundamentally different subjects, with distinct vocabulary and thematic profiles. In sharp contrast, annotation of the Deadly Cocktail Model's four psychological process indicators (identity fusion, outgroup othering, existential threat perception, and violence-condoning norms) reveals that all four are present across every movement, co-occur within messages far more than expected by chance, and follow an identical rank ordering. Temporal analyses reveal parallel dynamics and directed dependencies between these processes across movements, too. In a fragmented and fast-evolving ideological landscape, where ideological affiliation alone is an increasingly weak predictor of violence, we argue that process-based indicators, grounded in psychological theory and operationalised computationally, offer a more generalisable framework for understanding and detecting radicalisation across ideological boundaries.
- Open Access
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-22
Request a Note: How the Request Function Shapes X's Community Notes System
- 13 April 2026
- Published by Association for Computing Machinery (ACM)
- p. 1-22
Open AccessRequest a Note: How the Request Function Shapes X's Community Notes System
- Open Access
- 11 November 2025
- Published by Center for Open Science in OSF Preprints
Promoting engagement with social fact-checks online: Investigating the roles of social connection and shared partisanship
Social corrections – where users correct each other – can help rectify inaccurate beliefs. However,social corrections are often ignored. Here we ask under what conditions social corrections promoteengagement from corrected users, allowing for greater insight into how users respond to debunkingmessages (even if such responses are negative). Prior work suggests two key factors may helppromote engagement with corrections – partisan alignment between users, and social connectionsbetween users. We investigate these factors here. First, we conducted a field experiment on Twitter(X) using human-looking bots to examine how shared partisanship and prior social connectionaffect correction engagement. We randomized whether our accounts identified as Democrat orRepublican, and whether they followed Twitter users and liked three of their tweets beforecorrecting them (creating a minimal social connection). We found that shared partisanship had nosignificant effect in the baseline (no social connection) condition. Interestingly, social connectionincreased engagement with corrections from co-partisans. Effects in the social counter-partisancondition were ambiguous. Follow-up survey experiments largely replicated these results andfound evidence for a generalized norm of responding, wherein people feel more obligated torespond to people who follow them – even outside the context of misinformation correction. Ourfindings have important implications for increasing engagement with social corrections online.
- 11 November 2025
- Published by Center for Open Science in OSF Preprints
Open AccessPromoting engagement with social fact-checks online: Investigating the roles of social connection and shared partisanship
Social corrections – where users correct each other – can help rectify inaccurate beliefs. However,social corrections are often ignored. Here we ask under what conditions social corrections promoteengagement from corrected users, allowing for greater insight into how users respond to debunkingmessages (even if such responses are negative). Prior work suggests two key factors may helppromote engagement with corrections – partisan alignment between users, and social connectionsbetween users. We investigate these factors here. First, we conducted a field experiment on Twitter(X) using human-looking bots to examine how shared partisanship and prior social connectionaffect correction engagement. We randomized whether our accounts identified as Democrat orRepublican, and whether they followed Twitter users and liked three of their tweets beforecorrecting them (creating a minimal social connection). We found that shared partisanship had nosignificant effect in the baseline (no social connection) condition. Interestingly, social connectionincreased engagement with corrections from co-partisans. Effects in the social counter-partisancondition were ambiguous. Follow-up survey experiments largely replicated these results andfound evidence for a generalized norm of responding, wherein people feel more obligated torespond to people who follow them – even outside the context of misinformation correction. Ourfindings have important implications for increasing engagement with social corrections online.
- Open Access
- 30 October 2025
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 122 (44)
Divergent patterns of engagement with partisan and low-quality news across seven social media platforms
In recent years, social media has become increasingly fragmented, as platforms evolve and new alternatives emerge. Yet most research studies a single platform—typically X/Twitter, or occasionally Facebook—leaving little known about the broader social media landscape. Here, we shed light on patterns of cross-platform variation in the high-stakes context of news sharing. We examine the relationship between user engagement and news domains’ political orientation and quality across seven platforms: X/Twitter, BlueSky, TruthSocial, Gab, GETTR, Mastodon, and LinkedIn. Using an exhaustive sample, we analyze all (over 10 million) posts containing links to news domains shared on these platforms during January 2024. We find that news shared on platforms with more conservative user bases is significantly lower quality on average. Turning to engagement, we find—contrary to hypotheses of a consistent “right-wing advantage” on social media—that the relationship between political lean and engagement is strongly heterogeneous across platforms. Conservative news posts receive more engagement on platforms where most content is conservative, and vice versa for liberal news posts, consistent with an “echo platform” perspective. In contrast, the relationship between news quality and engagement is strikingly consistent: Across all platforms examined, a given user’s lower-quality news posts received higher average engagement, even though higher-quality news is substantially more prevalent and garners far more total engagement across posts. This pattern holds when accounting for poster-level variation and is observed even in the absence of ranking algorithms, suggesting that user preferences—not algorithmic bias—may underlie the underperformance of higher-quality news.
- 30 October 2025
- Published by National Academy of Sciences in Proceedings of the National Academy of Sciences
- Vol. 122 (44)
Open AccessDivergent patterns of engagement with partisan and low-quality news across seven social media platforms
In recent years, social media has become increasingly fragmented, as platforms evolve and new alternatives emerge. Yet most research studies a single platform—typically X/Twitter, or occasionally Facebook—leaving little known about the broader social media landscape. Here, we shed light on patterns of cross-platform variation in the high-stakes context of news sharing. We examine the relationship between user engagement and news domains’ political orientation and quality across seven platforms: X/Twitter, BlueSky, TruthSocial, Gab, GETTR, Mastodon, and LinkedIn. Using an exhaustive sample, we analyze all (over 10 million) posts containing links to news domains shared on these platforms during January 2024. We find that news shared on platforms with more conservative user bases is significantly lower quality on average. Turning to engagement, we find—contrary to hypotheses of a consistent “right-wing advantage” on social media—that the relationship between political lean and engagement is strongly heterogeneous across platforms. Conservative news posts receive more engagement on platforms where most content is conservative, and vice versa for liberal news posts, consistent with an “echo platform” perspective. In contrast, the relationship between news quality and engagement is strikingly consistent: Across all platforms examined, a given user’s lower-quality news posts received higher average engagement, even though higher-quality news is substantially more prevalent and garners far more total engagement across posts. This pattern holds when accounting for poster-level variation and is observed even in the absence of ranking algorithms, suggesting that user preferences—not algorithmic bias—may underlie the underperformance of higher-quality news.