The study focused on the earliest and most volatile days of the invasion. Because the Russian government blocked X just two days after the invasion began, sharply limiting the ability of Russian users to comment publicly, the researchers restricted their X dataset to the first 48 hours, collecting a total of 5,480 tweets tagged with the most active hashtags, #RussiaUkraineConflict and #RussiaUkraineWar. For Wikipedia, they cast a wider temporal net, gathering user interactions from one month before the invasion to one month after it, covering roughly 85 distinct discussion threads on the article’s talk pages. This asymmetry in data collection is itself revealing: it captures the moment when a fast-moving social media firestorm collided with a slower, more deliberate knowledge-building process.

To make sense of the X data, the team turned to natural language processing. They cleaned the corpus with stopword filtering and extracted discussion themes using bigrams, pairs of adjacent words whose co-occurrence frequencies reveal which concepts travel together in the same sentences. The conditional probability of one word following another was computed from the counts of word pairs, allowing the researchers to build networks in which the most frequently used words became nodes connected to their most common companions. To classify each tweet’s stance without importing the researchers’ own political judgments, they used Amazon Mechanical Turk workers, deliberately excluding anyone residing in Russia, Ukraine, or Belarus. Each sentence was evaluated by at least three independent annotators, and a tweet was labeled pro-war or anti-war only if at least two-thirds of its tags agreed. In the end, 78.4 percent of the sample received a reliable stance label, while the ambiguous remainder was excluded.

The results exposed a striking fragmentation. The pro-war community on X organized its discourse around geopolitical and ideological narratives, with NATO’s eastward expansion and the 2014 ‘Revolution of Dignity’ in Ukraine, framed by that community as a coup, serving as central justifying themes. The anti-war community, by contrast, largely ignored the NATO question and instead concentrated on humanitarian arguments: civilian casualties, the suffering of families under bombardment, and the characterization of Vladimir Putin as a war criminal. Crucially, the researchers found almost no thematic overlap between the two camps. The only shared topic, NATO, was itself so deeply polarized that it offered no common ground for dialogue. Without a single shared channel of discussion, the two communities talked past each other, each expressing opinions largely unsupported by sources, which further reduced the possibility of criticism, correction, or any meaningful exchange.

Wikipedia told a fundamentally different story. On the encyclopedia’s talk pages, where editors must arrive at wording that all parties can accept, the researchers found that almost none of the themes dominating X appeared at all. Instead, discussions revolved around editing requests, sourcing requirements, naming conventions, and factual accuracy. Editors debated how to frame events, which actors to identify, and how to maintain contextual precision, rather than whether the war was justified. The platform’s rules requiring verifiable citations and a neutral point of view transformed what had been an explosive ideological clash on X into a technical negotiation over evidence and language. Conflict did not vanish, but it was channeled into procedures designed to resolve it.

One of the study’s most intriguing findings emerged from a heatmap of Wikipedia editing activity. Two accounts, Tobby72 and the IP-based user 8.28.81.21, had been active on the article well before the invasion, even requesting that the page title be changed to ‘Russo-Ukrainian War’ months before Russian forces crossed the border. The authors suggest this early activity could be consistent with preparation for the invasion as part of a propaganda effort, though they present this as an observation rather than a proven conclusion. The heatmap also revealed a division of labor typical of large collaborative projects: a small core of experienced, named editor accounts performed most of the sustained work of reviewing and correcting the page, while many casual or anonymous contributors offered a single suggestion and never returned. Unanswered proposals, the researchers note, were treated as tacitly accepted modifications.

The most technically distinctive contribution of the paper lies in its measurement of linguistic entropy, a quantity borrowed from information theory that captures the unpredictability, and by extension the complexity, of the language used. Applying Shannon entropy alongside BERT-based sentiment and emotion models, the researchers found that sentiment levels were broadly similar across the two platforms, but entropy diverged dramatically. X exhibited significantly lower entropy, reflecting the short, punchy, emotionally direct statements that the platform’s character limit and real-time dynamics reward. Wikipedia showed markedly higher entropy, consistent with the formal, precise, and contextually rich prose that editors must produce to satisfy standards of verifiability and neutrality. In other words, the platforms were not merely hosting different opinions; they were inducing different kinds of language altogether.

The authors frame this difference through the social roles each platform assigns to its users. On X, participants behave as citizens exercising free expression with minimal constraint, an arrangement that favors immediacy and provocation over nuance. On Wikipedia, participants function as volunteers embedded in a structured culture of editorial guidelines, peer review, and dispute resolution mechanisms that mirror organized labor. This culture, the study argues, creates a collaborative environment in which disagreement is processed through rules rather than raw confrontation, allowing contributors who hold opposing views about the war itself to cooperate on the narrower question of how to document it accurately. The requirement to cite sources also opens every claim to criticism, which paradoxically strengthens rather than weakens the possibility of agreement.

The study is candid about its limitations. The absence of Russian voices on X after the government blockade restricts the comparison, a larger dataset might have revealed temporal fluctuations in sentiment and entropy tied to specific events, and the possible presence of bots in the X data cannot be excluded. The researchers also acknowledge that they could not verify whether any X users and Wikipedia editors were the same individuals, since neither platform exposes account-level identifiers that would permit reliable cross-platform matching. Sentiment and emotion detection, applied through validated pretrained models rather than ones retrained on war discourse, may also have missed domain-specific expressions such as sarcasm or euphemism. Yet even with these caveats, the findings carry a clear message for anyone concerned with the health of the digital public sphere. Platform design, community norms, and the presence or absence of shared evidentiary standards can determine whether a polarizing event fractures a community into warring echo chambers or becomes the occasion for structured, evidence-based collaboration. As debates over algorithmic amplification and content moderation intensify, this comparison of X and Wikipedia suggests that the most powerful algorithmic influence on online conflict may not be the recommendation engine at all, but the deeper architecture that decides what kind of conversation a platform makes possible.

Subject of Research: Comparative analysis of conflict and cooperation dynamics on X and Wikipedia during the Russia-Ukraine war

Article Title: Algorithmic influence on conflict and cooperation in digital communities

Article References: Algorithmic influence on conflict and cooperation in digital communities. (n.d.). https://doi.org/10.1371/journal.pcsy.0000087

Image Credits: AI Generated

DOI: 10.1371/journal.pcsy.0000087

Keywords: algorithmic influence, X platform, Wikipedia, polarization, natural language processing, Shannon entropy, Russia-Ukraine war, online cooperation, echo chambers, sentiment analysis, platform design, digital communities

The paper, titled How AI companions could deepen social inequality, introduces a new framework for understanding how AI companionship may reinforce existing social inequalities across three interconnected dimensions: the characteristics of users themselves, the design choices made by platforms, and the governance structures that surround the technology. Rather than treating risks in isolation, the framework shows how these factors interact to produce unequal outcomes, and it proposes practical interventions at each level. The study was led by first author Zhang Qiyang, Assistant Professor in Learning Analytics at SMU’s College of Integrative Studies, together with Zhang Renwen, Nanyang Assistant Professor at Nanyang Technological University’s Wee Kim Wee School of Communication and Information, and senior author Liu Nan, Associate Professor at Duke-NUS Medical School’s Centre for Biomedical Data Science.

At the heart of the research is a striking observation about human relationships that the authors describe as a rich-get-richer dynamic. People who already enjoy strong family ties and robust social networks are more likely to use AI companions as a supplement to their existing relationships. For these users, the technology can serve as a rehearsal space, a place to practise difficult conversations or manage stress before facing real interactions. Their social skills remain sharpened by regular human contact, and the AI adds convenience rather than replacing genuine connection.

The picture looks very different for people who are lonely, socially isolated or have limited access to mental health support. According to the study, these users are more likely to turn to AI companions as a substitute for human connection rather than a complement to it. Over time, this reliance may contribute to what the researchers call social deskilling, the gradual erosion of interpersonal skills that comes from having fewer opportunities to practise authentic human interaction. In this way, the technology may deliver its greatest advantages to those who need it least, while leaving the most vulnerable users increasingly exposed to harm.

To explain how such inequalities emerge, the study adapts the Swiss cheese model, a framework widely used in engineering and safety science to describe how multiple small failures align to produce larger systemic risks. The model, familiar from aviation and healthcare safety analysis, pictures each protective layer as a slice of cheese containing holes. Disasters occur only when the holes in successive slices line up, allowing a hazard to pass through every barrier at once. Applied to AI companionship, the framework suggests that harm rarely stems from a single cause.

Instead, the researchers argue that negative outcomes occur when several protective layers fail simultaneously. These layers include users’ own AI literacy, the strength of their social support networks, the design choices made by platform developers and the quality of regulatory oversight. Each layer contains weaknesses, and when those weaknesses align, vulnerable users become significantly more exposed to harmful outcomes. The authors present this systems perspective as a significant shift from existing approaches, which tend to focus on individual users, technology companies or regulation in isolation rather than examining how these elements interact.

Among the four protective layers examined, the study identifies governance as the most pressing concern. Despite being increasingly used for emotional support and mental well-being, AI companions currently occupy a regulatory grey area in many countries. Most are governed as consumer applications rather than as technologies with potential psychosocial or mental health impacts, a classification that leaves substantial gaps in protection. The researchers argue that recognising AI companions as health-related technologies could unlock stronger safeguards, including age-appropriate design requirements, transparent disclosure that users are interacting with AI rather than a human, limits on emotionally manipulative engagement features, clearer data governance and privacy standards, and mandatory crisis-response protocols for users who disclose self-harm or other high-risk situations.

The urgency of these questions is particularly visible in Singapore, which the study highlights as both well positioned and uniquely exposed to the rapid rise of AI companionship. The country’s advanced digital infrastructure, strong policymaking capabilities and history of proactive technology governance position it to lead internationally in developing responsible safeguards. At the same time, high smartphone adoption, widespread AI usage, an ageing population, growing numbers of older adults living alone, and increasing concerns around youth mental health and loneliness create conditions in which AI companions are likely to be adopted quickly, particularly among vulnerable groups. Zhang Qiyang is also developing a global policy dashboard tracking countries’ governance frameworks for AI mental health technologies, and early findings indicate that many jurisdictions have yet to establish dedicated policies or guidelines for AI companions.

Building on its framework, the study outlines practical measures for policymakers. First, it argues that AI companions should be recognised as technologies with psychosocial and mental health implications rather than regulated solely as consumer or entertainment applications, an approach that could support age-appropriate design standards, clearer disclosure of AI identity, stronger privacy protections and consistent crisis-response protocols. Second, it calls for a layered governance approach, arguing that no single intervention can adequately address the risks of AI companionship. Policymakers should combine regulation with efforts to strengthen AI literacy, encourage responsible platform design and support real-world social connection, reducing the likelihood that vulnerable users become overly reliant on AI for emotional support.

The researchers emphasise that the trajectory of AI companionship is not fixed by the technology itself but by the choices society makes about how to design, govern and use it. Liu Nan, who also directs the Duke-NUS AI + Medical Sciences Initiative, notes that the rapid evolution of AI companions presents an opportunity to shape their role in society before widespread adoption outpaces governance, and that young people in particular may be less equipped to recognise the limitations or commercial incentives behind these systems, making thoughtful safeguards especially important. The hope, the authors conclude, is that with thoughtful regulation, responsible product design and stronger public AI literacy, AI companions can complement human relationships and support well-being while strengthening, rather than replacing, meaningful human connection, and without deepening the social inequalities that already divide communities.

Subject of Research: How AI companions may deepen social inequality through user, design and governance factors

Article Title: SMU-Duke-NUS study offers new framework for understanding how AI companions may shape social inequality

Article References: SMU-Duke-NUS study offers new framework for understanding how AI companions may shape social inequality. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: AI companions, social inequality, Swiss cheese model, governance, mental health, social deskilling, Nature Human Behaviour, AI literacy, platform design, regulation, loneliness, Singapore

At the heart of the Series lies a new umbrella review, itself a review of published reviews, covering data on nine million people under the age of 25 across 121 individual reviews. The analysis finds consistent associations between what the authors define as problematic digital media use and a range of adverse health outcomes for children, adolescents, and young adults. Crucially, the Series defines problematic use not by the sheer number of hours spent online, but by addiction-like patterns of engagement, in which digital media use becomes compulsive, displaces other activities, and proves difficult to control. Under this definition, problematic use is linked to a higher risk of common mental health problems, sleep problems, and reduced physical activity.

The picture is not uniformly negative. The umbrella review also identifies beneficial associations, including a link between moderate digital media use and higher wellbeing, and between active video game use and both higher wellbeing and healthier lifestyles. These findings underscore the authors’ insistence that the goal is not to demonise technology but to reshape the digital environment so that young people can benefit from online opportunities without substantial risk of harm. Digital media, as the Series frames it, encompasses the content people watch, make, or share online, the platforms through which it is delivered, and the devices used to access it, meaning that interventions can target any of these layers rather than treating screen time as a single undifferentiated exposure.

The authors are candid about the limitations of the evidence base. The certainty of the findings for both harms and benefits is rated low, and the associations between problematic use and health harms are modest in size, resting on observational data that cannot establish cause and effect. Series author Louise Holly of the Digital Transformations for Health Lab in the United Kingdom argues that this uncertainty is itself a product of the current system, noting that the most valuable data on how platforms actually affect users remain hidden within technology companies, limiting independent scrutiny and slowing scientific progress. A key recommendation of the Series is therefore that governments should require technology companies to share data on how their features affect users’ health with vetted public-interest researchers, and should provide the legal and technical infrastructure to make such sharing possible.

To illustrate what even a small individual risk could mean at population scale, the authors modelled the potential burden of high rates of digital media use combined with modestly elevated risks. Their estimates suggest that spending more than two hours a day on screens could be associated with roughly 2.8 million healthy years of life lost to depression and 240,000 healthy years lost to obesity worldwide each year. Dr Falk Müller-Riemenschneider of the National University of Singapore stresses that these projections are intended to be illustrative of the possible scale of harm rather than an accurate measurement, since they are based on low-certainty evidence, and that more research is needed to provide accurate estimates of health impacts at the population level. Even so, the arithmetic illustrates a core principle of public health: small risks applied across billions of users can translate into enormous burdens.

The Series places much of the responsibility not on users but on the architecture of the digital environment itself. Most digital platforms, the authors argue, are designed by companies to capture attention and collect data rather than to protect health. Features such as infinite scroll, autoplay, and personalised recommendation feeds are deliberate design choices engineered to maximise engagement, not accidents of technology that individuals can simply self-regulate away. Dr Jay Shaw of the University of Toronto calls for far more research into how these specific design choices influence the way people use technology, rather than continuing to investigate only the impact of how long people spend online. This reframing shifts the analytical focus from user behaviour to product design, echoing earlier public health battles over products engineered for dependence.

Indeed, the Series draws explicit parallels with tobacco and fossil fuels, arguing that technology companies are drawing on a well-worn corporate defence playbook to delay, weaken, or shape regulation in response to evidence that their products may harm health. Professor Barbara Prainsack of the University of Vienna describes the tactics as manufacturing doubt, discrediting critics, and framing regulation as a threat to freedom. A public health approach counters this, she argues, by assessing what the evidence suggests about harm across whole populations rather than what any single case can prove. Prainsack also contends that good regulation is not the enemy of innovation but a driver of good innovation, channelling it toward safety and trust, and that where regulation is absent, it does not mean a technology is ungoverned, only that companies rather than public policy are setting the rules.

On the policy side, the Series is emphatic that no single measure is sufficient. Age limits alone, a favoured tool of several national governments, cannot address the structural drivers of digital harm. Instead, the authors call for wide-ranging, coordinated action at the company, national, and global levels. Among their specific proposals are requirements that technology companies assess new products or features for harm before releasing them, analogous to pre-market safety testing in other industries, alongside mandatory data sharing with vetted researchers. They further propose a global governance treaty through the World Health Organization to set baseline standards, preventing weak regulation in one country from undermining progress elsewhere, and a dedicated international body to monitor harms to children and young people from artificial intelligence. The Series also highlights how litigation is emerging as a route to hold technology companies accountable, with Professor Ilona Kickbusch, the Series chair, noting that legal action can compel companies to disclose internal documents and shift the public conversation about responsibility for harm, though she cautions that litigation is not a substitute for regulation and should be considered one route to accountability among several.

The Series was informed by a Youth Advisory Group of six young adults aged 23 to 30, who worked with the authors to articulate aspects of lived experience that the research alone could not capture. Group member Dr Salman Fitrat Khan of India argues that the health risks associated with digital systems are preventable, and that with better governance and by empowering citizens, especially young people, to co-shape the digital systems around them, the online world can deliver connection, information, and opportunity without putting health at risk. His message to governments is blunt: they should not wait for perfect evidence before acting for this generation and the ones that follow.

Kickbusch, who leads the Digital Transformations for Health Lab in Switzerland, summarises the Series’ core claim in stark terms: digital media use is now woven into the fabric of childhood and young adult life, yet governments have allowed technology companies to write their own rules, producing business models built to capture attention and collect data with health treated as an afterthought. The evidence linking problematic use to poorer mental health, sleep, and physical activity may not be conclusive, she acknowledges, but in her view it justifies coordinated, global action on precautionary grounds. The Series, published as part of The Lancet’s examination of the digital determinants of health, signals a shift in the scientific and policy conversation, from debates about individual screen time toward a systems-level analysis of how digital environments are designed, regulated, and held to account. Whether governments respond with the coordinated framework the authors demand, or continue to defer to industry self-regulation, will determine whether the next generation inherits a digital world shaped for health or for profit.

Subject of Research: Digital media use and health impacts in children and young people

Article Title: The Lancet: Scale of young people’s digital media use and emerging evidence of harm justify action now, experts argue

Article References: The Lancet: Scale of young people’s digital media use and emerging evidence of harm justify action now, experts argue. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: digital media, young people, public health, The Lancet, mental health, screen time, technology regulation, platform design, sleep problems, physical activity, global governance, umbrella review

The study’s strength lies in its design. Rather than running a small laboratory experiment, the team exploited what statisticians call a natural experiment: Change.org introduced its ‘write with AI’ feature in some countries before others, creating a clean before-and-after boundary. The researchers collected roughly 1.5 million petitions and applied a difference-in-differences analysis, a causal inference technique that compares changes over time between groups that received the AI tool and groups that did not. This approach, widely used in economics and built on methods formalized by Callaway and Sant’Anna for settings with multiple time periods, allows researchers to isolate the effect of the intervention from broader trends that would have happened anyway. Because the rollout timing was determined by the platform rather than by users’ characteristics, the comparison approximates the rigor of a randomized controlled trial at population scale.

The first major finding concerns the texture of the text itself. After the AI tool became available, the lexical features of petitions shifted measurably at the platform level. Petitions grew longer and their language changed in ways consistent with the stylistic fingerprints of large language models: smoother phrasing, more standardized vocabulary, and altered readability characteristics. These are not trivial cosmetic shifts. Lexical measures such as vocabulary diversity and readability scores have long been used in computational text analysis to characterize how writing differs across authors, genres, and eras, and the fact that a single platform feature could move these metrics across an entire ecosystem of user-generated content is striking. In effect, the introduction of the tool bent the collective writing style of a global advocacy community.

But the second major finding is where the story turns. None of these stylistic improvements translated into better petition outcomes. Petitions written with access to the AI assistant did not attract more signatures, achieve their goals more often, or otherwise perform better than petitions written without it. For a tool presumably offered to help advocates succeed, that null result carries real weight. It echoes a growing body of evidence that generative AI can boost individual productivity in controlled settings, as shown in the well-known 2023 Science experiment by Noy and Zhang on professional writing tasks, while failing to deliver equivalent gains when success depends on persuading other humans in a noisy, competitive environment.

The researchers did not stop at the platform-level analysis. A persistent worry with observational studies is that the people who choose to use a new feature differ systematically from those who do not, contaminating any comparison. To address this, the team conducted a separate analysis of repeat petition writers: 4,611 users who had written at least one petition before the AI tool was introduced and another one after. This within-person comparison holds individual ability, motivation, and topic preferences roughly constant, isolating the effect of AI access on the same writers over time. The pattern that emerged was telling. On average, these returning writers produced longer petitions when they had access to AI, yet their second petitions fared worse in terms of outcomes than their earlier work. The very users most likely to benefit from accumulated platform experience saw no advantage, and by some measures a disadvantage, from the writing assistant.

Perhaps the most consequential finding is the increase in homogeneity. Using dynamic difference-in-differences estimates, the researchers showed that platform-wide homogeneity among petitions rose after the AI tool’s introduction. In plain terms, petitions began to sound more like one another. This result aligns with a rapidly expanding literature on what some researchers call generative monoculture. A 2024 study in Science Advances by Doshi and Hauser found that generative AI enhanced individual creativity while reducing the collective diversity of novel content. Work presented at the 2025 CHI Conference by Agarwal, Naaman, and Vashistha showed that AI writing suggestions can homogenize text toward Western stylistic conventions and erode cultural nuance. Padmakumar and He asked directly whether writing with language models reduces content diversity and found evidence that it does. The Change.org study extends these laboratory and small-scale findings to a genuine production environment used by millions.

Why should homogenization matter for something as seemingly benign as petition language? Advocacy platforms depend on differentiation. A petition competes for attention in a crowded feed, and its distinctive voice, urgency, and local specificity are part of what makes a reader stop and sign. When an algorithmic assistant nudges thousands of writers toward the same polished register, the collective signal degrades even if each individual text improves. Researchers have raised parallel concerns in other domains: a 2026 Nature analysis by Hao and colleagues found that AI tools expanded individual scientists’ output while contracting the focus of science as a whole, and work on Stack Overflow suggested that large language models may threaten the value of community-generated digital public goods. The pattern that emerges across these studies is consistent: individual-level gains, community-level costs.

The null effect on outcomes also invites a deeper question about what petition success actually depends on. Prior research on e-petitions, including studies of linguistic cues and multidimensional time-series predictors of success, has suggested that factors such as topic salience, timing, social networks, and media coverage often dwarf the influence of prose quality. A beautifully written petition about an issue nobody cares about will still fail; a clumsy petition riding a wave of public outrage may succeed regardless. The Change.org results are consistent with this view. If outcomes are driven primarily by the underlying demand for a cause rather than by the craft of the appeal, then a tool that polishes craft without touching demand should indeed leave outcomes unchanged. The study cannot rule out subtler possibilities, such as offsetting effects in which longer, more polished text helps in some respects and hurts in others, but the headline conclusion stands: no measurable outcome benefit at scale.

There are also social and perceptual dimensions that the authors situate within a broader research program. Studies have shown that human heuristics for detecting AI-generated language are unreliable, that there can be a social evaluation penalty for using AI, and that perceptions of AI-mediated communication affect trustworthiness and perceived authenticity. For petition signers, the suspicion that a heartfelt appeal was machine-generated could matter as much as the text itself. Related work on the ‘AI ghostwriter effect’ shows that users often fail to feel ownership of AI-generated text while still claiming authorship, raising questions about authenticity in civic expression. A platform that encourages AI-assisted advocacy may therefore be trading not only diversity of style but also the perceived sincerity that makes grassroots appeals compelling.

The practical implications reach well beyond one platform. Companies including Amazon and Meta have embedded generative writing tools into seller listings and social posts, making the Change.org experiment a preview of dynamics likely to unfold across the internet. The study’s message for platform designers is that embedding AI writing assistance is not a neutral upgrade: it reshapes the linguistic character of an entire content ecosystem and can introduce unintended consequences such as homogenization without delivering the hoped-for performance gains. For researchers, the work demonstrates the power of natural experiments and difference-in-differences designs for evaluating AI interventions in the wild, complementing randomized trials that capture only narrow populations and short horizons. And for the millions of people who turn to online platforms to press for change, the findings carry a quietly empowering lesson: the cause, the community, and the timing still matter more than the polish of the prose. The data and code behind the analysis have been released through the Open Science Framework, allowing other researchers to reproduce the results and probe further into how generative AI is quietly rewriting the way the public speaks online.

Subject of Research: Causal effects of an in-platform AI writing tool on petition content and outcomes on Change.org

Article Title: Introducing AI to an online petition platform changed outputs but not outcomes

Article References: Corpus, I., Gilbert, E., Koenecke, A., & Naaman, M. (2026). Introducing AI to an online petition platform changed outputs but not outcomes. Nature Human Behaviour. https://doi.org/10.1038/s41562-026-02580-8

Image Credits: AI Generated

DOI: 10.1038/s41562-026-02580-8

Keywords: generative AI, Change.org, online petitions, difference-in-differences, content homogenization, large language models, computational text analysis, Nature Human Behaviour, platform design, AI writing tools, natural experiment, digital activism