Centre for Democracy and Information Integrity (CDI)

News

2026

• Faisal Feroz & Jonas R. Kunst

  Can AI Debias the News? LLM Interventions Improve Cross-Partisan Receptivity but LLMs Overestimate Their Own Effectiveness

  Computers in Human Behavior Reports, 2026, 101285, ISSN 2451-9588

  Abstract: Partisan news media erode cross-partisan trust, but large language models (LLMs) offer a potential means of debiasing such content at scale. Across two pre-registered experiments, we tested whether LLM-generated debiasing of liberal news headlines could improve conservative readers' trust-relevant judgments. Study 1 found that subtle lexical debiasing (replacing emotive words with moderate synonyms) had no effect on any outcome. Study 2 found that a more substantive reframing intervention significantly increased conservatives' perceived trustworthiness, completeness, and willingness to engage with liberal news headlines, without producing a backfire effect among liberals. In Study 1, the intervention produced robust effects across silicon participants simulated with six different models (o3-mini, o3, GPT-4o mini, GPT-4o, GPT-5 mini, and GPT-5), whereas it had no impact on human readers. In Study 2, the intervention's effects among silicon participants generally aligned directionally with human responses but were significantly larger for some outcomes, and three models (o3, GPT-4o mini, and GPT-4o) incorrectly predicted a liberal backfire effect absent in humans. Moderation analyses tentatively suggest that the models’ implicit theory of who responds to debiasing diverged from the psychological profile that actually predicted human responsiveness. Most strikingly, in Study 2, participants simulated by each of the six models suggested that debiasing effects would be stronger among participants high in political in-group identification. Yet, no such moderation was observed among human participants. Collectively, these findings suggest that LLM-based debiasing can improve cross-partisan receptivity when the intervention targeted ideological framing but not when it was limited to the minimal surface-level substitutions tested in this research — and that current models lack both the quantitative accuracy and qualitative psychological fidelity to evaluate their own interventions without human oversight.

• Inesa Zayets et al., with Suzanne van Gils

  Think Political Leader — Think He or She? A Multi-Country Test of a Gender Fair Language Intervention for Stereotype Bias

  PsyArXiv (OSF Preprints)

  Abstract: Women are greatly underrepresented in positions of political leadership around the world. In seeking to explain this underrepresentation, some researchers have pointed to people’s tendencies to stereotype leaders as more similar to men than women as these tendencies can support the belief that women are unsuited to leadership. This project aimed to test whether a subtle linguistic intervention was able to ameliorate this gender bias in political leadership stereotypes across different languages and national contexts (country N = 42, sample N = 22,995). Specifically, this project examined whether gender fair language (e.g., the use of paired pronouns ‘he or she’) can reduce the tendency for people to stereotype political leaders as more similar to men than women. To increase the rigor with which these stereotypes were measured, this project complemented the dominant ‘cheap talk’ measure of personal stereotype content with a novel incentivised community measure of this content. As expected, participants stereotyped political leaders as more similar to men than women; this pattern was stronger on the incentivised measure of community stereotypes. Unexpectedly, there was no evidence that paired pronouns reduced this bias (there was weak exploratory evidence that paired nouns increased participants’ tendencies to stereotype political leaders as more similar to women). This project suggests that the ability of gender fair language interventions to ameliorate gender bias in the political leadership domain may be limited.

• Ningzhi Liu et al., with Yannic Hinrichs and Jonas R. Kunst

  The persuasive power of large language models does not depend on their perceived national origin

  Arxiv.org, Human-Computer Interaction

  Abstract: Conversational AI developed by geopolitical rivals reaches citizens worldwide, raising concerns that it could sway public opinion or be rejected as foreign propaganda, with consequences for democratic discourse and information sovereignty. Yet, whether an AI's perceived national origin shapes its persuasive power is unknown. In a preregistered randomized experiment, 403 adults from a nationally representative United States sample held a three-round debate with a chatbot introduced as either American ("DiscoveryAI") or Chinese ("ZhengheAI"), discussing a political or non-political topic. In all conditions, participants actually conversed with the same model (GPT-4o), instructed to argue against their initial position. We combined pre- and post-conversation self-reports of attitudes, trust, and collective narcissism with computational analyses of 1,209 participant turns, including LLM-coded stance and argumentative conduct, stance-sensitive embeddings, and keyword-masked emotion and toxicity classifiers. The conversations produced substantial attitude changes in every condition. Critically, the nationality label affected neither self-reported attitude change nor expressed stance, concessions, counterarguing, or affect, and equivalence tests and Bayes factors largely supported these null effects. The label's only reliable footprint was lower pre-conversation human-like trust in the Chinese model, whereas functionality trust was unaffected. Political topics slowed stance movement toward the AI's position, and collective narcissism predicted less attitude change regardless of origin, acting as a general barrier rather than an out-group filter. Users thus initially withhold social trust from a rival's AI yet still assimilate its arguments; origin labeling and transparency requirements alone may offer weak protection against foreign influence operations conducted through conversational AI.

• Petter Bae Brandtzaeg et al., with Jonas R. Kunst

  Professional and community-based fact-checking show different strengths, but neither performs strongly across trust, scalability, and impact

  Harvard Kennedy School (HKS) Misinformation Review

  Abstract: Fact-checking on social media matters more than ever. When Meta abandoned professional fact checkers in 2025 in favour of crowd-sourced annotations, it bet that community participation could substitute for institutional expertise. Our systematic review of 21 empirical studies of social media fact-checking (2022–2025) suggests the picture is more complicated than it appears. Community-based fact-checking is faster and scalable, whereas professional fact checkers are generally more trusted but often slow. Neither model, on its own, performs strongly across trust, scalability, and impact. The evidence suggests that hybrid systems combining community speed with expert verification are likely to be the most effective approach.

• Jonas R. Kunst et al.

  How cyborg propaganda reshapes collective action

  ArXiv.org, Computers and Society

  Abstract: The distinction between genuine grassroots activism and automated influence operations is collapsing. While contemporary policy debates prioritize fully autonomous generative agents and synthetic content, this paper offers a conceptual contribution: we develop 'cyborg propaganda,' a closed-loop architecture combining verified human accounts with algorithmic automation to generate personalized content at scale, as a distinct and undertheorized threat to democratic discourse. By relying on verified citizens to ratify AI-generated messages, these campaigns exploit a regulatory gray zone that frameworks built on the human/bot binary (including the EU AI Act and Section 230) are structurally unable to address. Drawing on a conceptual analysis of coordination platforms and comparative examination of governance frameworks across democratic and non-democratic contexts, we analyze this paradox across micro, meso, and macro levels. We examine whether cyborg propaganda democratizes political power by unionizing influence or reduces citizens to cognitive proxies of a hidden directive, arguing that it shifts political discourse from a contest of ideas to a battle of algorithmic campaigns. We propose three regulatory responses: classifying coordination hubs as political action committees to enforce supply-chain transparency; mandating researcher access to platform data through DSA-style mechanisms; and establishing risk standards penalizing amplification of synthetically coordinated content. Comparative analysis reveals that viability varies structurally. Democratic states are simultaneously the most capable of regulation and the most rule-of-law constrained. By contrast, non-democratic actors face no comparable accountability, making international risk standards the primary cross-border enforcement mechanism.

• Save the date: CDI kick-off conference, fall 2026

  CDI will mark its launch with a kick-off conference in the fall of 2026, co-organized with Oslo Science City, convening academics, policymakers, and practitioners. Date: TBA soon. Registration link coming soon.

• New tool: compare misinformation interventions—for whom they work best

  Alongside the Journal of Experimental Psychology: General paper on false-news sharing, the team released an open-source interactive app that lets researchers and policymakers compare the leading interventions—accuracy prompts, warning labels, social-norm nudges, and media-literacy tips—for a chosen audience, and see how each one reshapes the underlying decision to share.

• Journal Anton Gollwitzer et al.

  Toward a mechanistic understanding of false news sharing: Which interventions work best, for whom, and why

  Journal of Experimental Psychology: General, 155(6), 1550–1574.

  Abstract: False news—given its capacity to distort public opinion and erode trust—has prompted extensive research on potential countermeasures. Yet, there has been no systematic, comparative, and computational investigation of false news sharing and how best to curb it. To address this gap, we apply a semi-integrative experimental approach that (a) compares multiple existing false news interventions, (b) examines how individual and news-level factors predict false news sharing and shape intervention efficacy, and (c) uses drift-diffusion modeling to uncover the decision-making processes underlying all these effects. We find warning labels and media literacy tips to substantially improve news-sharing quality, whereas social norm cues exert a comparatively modest effect, and accuracy prompts yield only subtle benefits. Although numerous individual factors (e.g., age, political conservatism, social media use) predicted news-sharing quality, the observed intervention effects remained broadly robust across these factors, proving effective even within at-risk populations. Intervention outcomes were likewise robust to news-level variation, such as the believability, sensationalism, and political congruence of news content. Despite this robustness, we find each intervention to operate via distinct decision-making pathways. Warning labels shift initial sharing intentions toward sharing higher quality news, whereas media literacy tips operate later, enhancing the processing of news content and increasing cautiousness before making sharing decisions. By applying a multicomponent experimental framework, this work clarifies the risk factors and decision-making processes driving false news sharing and pinpoints which interventions work best, how they operate at the process level, and in which contexts they should be most effective.

• Article Jonas R. Kunst et al.

  Intelligent Systems, Vulnerable Minds: A Framework for Radicalization to Violence in the Age of AI

  Personality and Social Psychology Review, 30(3), 395-426.

  Abstract: Advances in AI require a revision of the psychological and socio-technical dynamics by which individuals are radicalized to embrace violent extremism. This review synthesizes process models of radicalization with research on social and personality risk factors, AI, and psychological mechanisms to propose a four-stage framework mapping the AI architecture of radicalization: (1) Exposure, where recommender systems and virality features create initial attraction to extreme content; (2) Reinforcement, where filter bubbles and group recommendations leverage biases to strengthen extremist beliefs and create echo chambers; (3) Group Integration, where ideologically homogenous clusters, AI bot swarms and companions foster group belonging and readiness for action; cumulatively resulting in (4) Violent Extremist Action. We examine how established social, cognitive, personality, and contextual vulnerability factors heighten psychological risk in the AI-driven radicalization process, as well as the emerging role of generative AI. We conclude by outlining a stage-based framework for governance and future research.

• Announcing the Center for Democracy & Information Integrity

  BI Norwegian Business School has established the Center for Democracy & Information Integrity (CDI), co-directed by Anton Gollwitzer and Jonas R. Kunst.

  The center brings together faculty from across five BI departments—spanning psychology, political science, communication, economics, law, and data science—to study misinformation, polarization, institutional trust, and the effects of AI on democratic life, and to translate that research into practical tools for policymakers, platforms, and the public.

2026

• Faisal Feroz & Jonas R. Kunst

  Can AI Debias the News? LLM Interventions Improve Cross-Partisan Receptivity but LLMs Overestimate Their Own Effectiveness

  Computers in Human Behavior Reports, 2026, 101285, ISSN 2451-9588

  Abstract: Partisan news media erode cross-partisan trust, but large language models (LLMs) offer a potential means of debiasing such content at scale. Across two pre-registered experiments, we tested whether LLM-generated debiasing of liberal news headlines could improve conservative readers' trust-relevant judgments. Study 1 found that subtle lexical debiasing (replacing emotive words with moderate synonyms) had no effect on any outcome. Study 2 found that a more substantive reframing intervention significantly increased conservatives' perceived trustworthiness, completeness, and willingness to engage with liberal news headlines, without producing a backfire effect among liberals. In Study 1, the intervention produced robust effects across silicon participants simulated with six different models (o3-mini, o3, GPT-4o mini, GPT-4o, GPT-5 mini, and GPT-5), whereas it had no impact on human readers. In Study 2, the intervention's effects among silicon participants generally aligned directionally with human responses but were significantly larger for some outcomes, and three models (o3, GPT-4o mini, and GPT-4o) incorrectly predicted a liberal backfire effect absent in humans. Moderation analyses tentatively suggest that the models’ implicit theory of who responds to debiasing diverged from the psychological profile that actually predicted human responsiveness. Most strikingly, in Study 2, participants simulated by each of the six models suggested that debiasing effects would be stronger among participants high in political in-group identification. Yet, no such moderation was observed among human participants. Collectively, these findings suggest that LLM-based debiasing can improve cross-partisan receptivity when the intervention targeted ideological framing but not when it was limited to the minimal surface-level substitutions tested in this research — and that current models lack both the quantitative accuracy and qualitative psychological fidelity to evaluate their own interventions without human oversight.

• Inesa Zayets et al., with Suzanne van Gils

  Think Political Leader — Think He or She? A Multi-Country Test of a Gender Fair Language Intervention for Stereotype Bias

  PsyArXiv (OSF Preprints)

  Abstract: Women are greatly underrepresented in positions of political leadership around the world. In seeking to explain this underrepresentation, some researchers have pointed to people’s tendencies to stereotype leaders as more similar to men than women as these tendencies can support the belief that women are unsuited to leadership. This project aimed to test whether a subtle linguistic intervention was able to ameliorate this gender bias in political leadership stereotypes across different languages and national contexts (country N = 42, sample N = 22,995). Specifically, this project examined whether gender fair language (e.g., the use of paired pronouns ‘he or she’) can reduce the tendency for people to stereotype political leaders as more similar to men than women. To increase the rigor with which these stereotypes were measured, this project complemented the dominant ‘cheap talk’ measure of personal stereotype content with a novel incentivised community measure of this content. As expected, participants stereotyped political leaders as more similar to men than women; this pattern was stronger on the incentivised measure of community stereotypes. Unexpectedly, there was no evidence that paired pronouns reduced this bias (there was weak exploratory evidence that paired nouns increased participants’ tendencies to stereotype political leaders as more similar to women). This project suggests that the ability of gender fair language interventions to ameliorate gender bias in the political leadership domain may be limited.

• Ningzhi Liu et al., with Yannic Hinrichs and Jonas R. Kunst

  The persuasive power of large language models does not depend on their perceived national origin

  Arxiv.org, Human-Computer Interaction

  Abstract: Conversational AI developed by geopolitical rivals reaches citizens worldwide, raising concerns that it could sway public opinion or be rejected as foreign propaganda, with consequences for democratic discourse and information sovereignty. Yet, whether an AI's perceived national origin shapes its persuasive power is unknown. In a preregistered randomized experiment, 403 adults from a nationally representative United States sample held a three-round debate with a chatbot introduced as either American ("DiscoveryAI") or Chinese ("ZhengheAI"), discussing a political or non-political topic. In all conditions, participants actually conversed with the same model (GPT-4o), instructed to argue against their initial position. We combined pre- and post-conversation self-reports of attitudes, trust, and collective narcissism with computational analyses of 1,209 participant turns, including LLM-coded stance and argumentative conduct, stance-sensitive embeddings, and keyword-masked emotion and toxicity classifiers. The conversations produced substantial attitude changes in every condition. Critically, the nationality label affected neither self-reported attitude change nor expressed stance, concessions, counterarguing, or affect, and equivalence tests and Bayes factors largely supported these null effects. The label's only reliable footprint was lower pre-conversation human-like trust in the Chinese model, whereas functionality trust was unaffected. Political topics slowed stance movement toward the AI's position, and collective narcissism predicted less attitude change regardless of origin, acting as a general barrier rather than an out-group filter. Users thus initially withhold social trust from a rival's AI yet still assimilate its arguments; origin labeling and transparency requirements alone may offer weak protection against foreign influence operations conducted through conversational AI.

• Petter Bae Brandtzaeg et al., with Jonas R. Kunst

  Professional and community-based fact-checking show different strengths, but neither performs strongly across trust, scalability, and impact

  Harvard Kennedy School (HKS) Misinformation Review

  Abstract: Fact-checking on social media matters more than ever. When Meta abandoned professional fact checkers in 2025 in favour of crowd-sourced annotations, it bet that community participation could substitute for institutional expertise. Our systematic review of 21 empirical studies of social media fact-checking (2022–2025) suggests the picture is more complicated than it appears. Community-based fact-checking is faster and scalable, whereas professional fact checkers are generally more trusted but often slow. Neither model, on its own, performs strongly across trust, scalability, and impact. The evidence suggests that hybrid systems combining community speed with expert verification are likely to be the most effective approach.

• Jonas R. Kunst et al.

  How cyborg propaganda reshapes collective action

  ArXiv.org, Computers and Society

  Abstract: The distinction between genuine grassroots activism and automated influence operations is collapsing. While contemporary policy debates prioritize fully autonomous generative agents and synthetic content, this paper offers a conceptual contribution: we develop 'cyborg propaganda,' a closed-loop architecture combining verified human accounts with algorithmic automation to generate personalized content at scale, as a distinct and undertheorized threat to democratic discourse. By relying on verified citizens to ratify AI-generated messages, these campaigns exploit a regulatory gray zone that frameworks built on the human/bot binary (including the EU AI Act and Section 230) are structurally unable to address. Drawing on a conceptual analysis of coordination platforms and comparative examination of governance frameworks across democratic and non-democratic contexts, we analyze this paradox across micro, meso, and macro levels. We examine whether cyborg propaganda democratizes political power by unionizing influence or reduces citizens to cognitive proxies of a hidden directive, arguing that it shifts political discourse from a contest of ideas to a battle of algorithmic campaigns. We propose three regulatory responses: classifying coordination hubs as political action committees to enforce supply-chain transparency; mandating researcher access to platform data through DSA-style mechanisms; and establishing risk standards penalizing amplification of synthetically coordinated content. Comparative analysis reveals that viability varies structurally. Democratic states are simultaneously the most capable of regulation and the most rule-of-law constrained. By contrast, non-democratic actors face no comparable accountability, making international risk standards the primary cross-border enforcement mechanism.

• Save the date: CDI kick-off conference, fall 2026

  CDI will mark its launch with a kick-off conference in the fall of 2026, co-organized with Oslo Science City, convening academics, policymakers, and practitioners. Date: TBA soon. Registration link coming soon.

• New tool: compare misinformation interventions—for whom they work best

  Alongside the Journal of Experimental Psychology: General paper on false-news sharing, the team released an open-source interactive app that lets researchers and policymakers compare the leading interventions—accuracy prompts, warning labels, social-norm nudges, and media-literacy tips—for a chosen audience, and see how each one reshapes the underlying decision to share.

• Announcing the Center for Democracy & Information Integrity

  BI Norwegian Business School has established the Center for Democracy & Information Integrity (CDI), co-directed by Anton Gollwitzer and Jonas R. Kunst.

  The center brings together faculty from across five BI departments—spanning psychology, political science, communication, economics, law, and data science—to study misinformation, polarization, institutional trust, and the effects of AI on democratic life, and to translate that research into practical tools for policymakers, platforms, and the public.

2026

• Article Jonas R. Kunst et al.

  Intelligent Systems, Vulnerable Minds: A Framework for Radicalization to Violence in the Age of AI

  Personality and Social Psychology Review, 30(3), 395-426.

  Abstract: Advances in AI require a revision of the psychological and socio-technical dynamics by which individuals are radicalized to embrace violent extremism. This review synthesizes process models of radicalization with research on social and personality risk factors, AI, and psychological mechanisms to propose a four-stage framework mapping the AI architecture of radicalization: (1) Exposure, where recommender systems and virality features create initial attraction to extreme content; (2) Reinforcement, where filter bubbles and group recommendations leverage biases to strengthen extremist beliefs and create echo chambers; (3) Group Integration, where ideologically homogenous clusters, AI bot swarms and companions foster group belonging and readiness for action; cumulatively resulting in (4) Violent Extremist Action. We examine how established social, cognitive, personality, and contextual vulnerability factors heighten psychological risk in the AI-driven radicalization process, as well as the emerging role of generative AI. We conclude by outlining a stage-based framework for governance and future research.

2026

• Journal Anton Gollwitzer et al.

  Toward a mechanistic understanding of false news sharing: Which interventions work best, for whom, and why

  Journal of Experimental Psychology: General, 155(6), 1550–1574.

  Abstract: False news—given its capacity to distort public opinion and erode trust—has prompted extensive research on potential countermeasures. Yet, there has been no systematic, comparative, and computational investigation of false news sharing and how best to curb it. To address this gap, we apply a semi-integrative experimental approach that (a) compares multiple existing false news interventions, (b) examines how individual and news-level factors predict false news sharing and shape intervention efficacy, and (c) uses drift-diffusion modeling to uncover the decision-making processes underlying all these effects. We find warning labels and media literacy tips to substantially improve news-sharing quality, whereas social norm cues exert a comparatively modest effect, and accuracy prompts yield only subtle benefits. Although numerous individual factors (e.g., age, political conservatism, social media use) predicted news-sharing quality, the observed intervention effects remained broadly robust across these factors, proving effective even within at-risk populations. Intervention outcomes were likewise robust to news-level variation, such as the believability, sensationalism, and political congruence of news content. Despite this robustness, we find each intervention to operate via distinct decision-making pathways. Warning labels shift initial sharing intentions toward sharing higher quality news, whereas media literacy tips operate later, enhancing the processing of news content and increasing cautiousness before making sharing decisions. By applying a multicomponent experimental framework, this work clarifies the risk factors and decision-making processes driving false news sharing and pinpoints which interventions work best, how they operate at the process level, and in which contexts they should be most effective.