Published Sep 2, 2026Added Sep 7, 2026
The machinery that builds cooperation can also harden group boundaries
An agent-based study of cultural transmission shows how the same processes that stabilize cooperation inside a group can also make group boundaries more behaviorally important. In the model, norm transmission, migration, and intergroup competition jointly shaped whether cooperative groups persisted. The result is a useful warning against treating cohesion as unconditionally positive, but it remains a theoretical and simulation result rather than an identified causal effect in real populations.
Published Sep 1, 2026Added Sep 7, 2026
U.S. Army Pacific is treating information advantage as part of deterrence
At INFOPAC 2026, U.S. Army Pacific commander Gen. Ronald P. Clark described the information environment as decisive terrain and argued that cyber, electronic warfare, deception, influence, resilient networks, and visible military capability should work together to shape adversary decisions before conflict. The important point is the explicit deterrence theory being articulated. The speech shows how the command says the mechanism should work; it does not by itself demonstrate that adversary behavior has changed.
Published Aug 5, 2026Added Sep 7, 2026
Spite can be measured as a behavioral response, not just a trait
A new behavioral study turned spite into a measurable choice pattern: participants could accept a personal cost to impose a loss on someone they disliked. Spite generally fell as social distance increased, and some results suggested sensitivity to how much harm the other person would experience. The work offers a way to model costly antagonistic behavior more precisely, although the choices were hypothetical and need replication in consequential settings.
Published Aug 3, 2026Added Sep 7, 2026
Earlier social-media adoption was linked to worse later learning outcomes
A longitudinal study of 5,227 students used matched Difference-in-Differences comparisons to examine when students first opened social-media accounts. Earlier adopters, particularly those starting in grades 6 or 7, later performed worse in Italian and mathematics than matched late adopters, while English showed no comparable effect. The design is stronger than a cross-sectional screen-time correlation, but it remains quasi-experimental and the proposed smartphone-distraction mechanism was not causally established.
Published Jul 30, 2026Added Sep 7, 2026
A “valid” silicon sample can depend heavily on how you configure it
A methods study tested hundreds of defensible ways to generate LLM-based silicon samples and found that conclusions about human similarity changed substantially with model, prompt, demographic, reasoning, and sampling choices. In one reanalysis, human-synthetic association correlations ranged from .23 to .84. The result does not show that synthetic respondents are inherently invalid; it shows that their apparent validity can be highly sensitive to researcher choices that need to be governed and reported.
Published Jun 5, 2026Added Sep 7, 2026
AI can become “too caring to be real”
A mixed-method study found that moderate autonomy-supportive AI communication increased intimacy, conversational elaboration, and voluntary re-engagement, but highly consistent or fluent support could begin to feel scripted and reduce authenticity. Human escalation and transparent AI disclosure softened that penalty. The practical lesson is that empathic AI may need calibration rather than maximization, although the findings come from the studied AI-consumer interaction contexts and may not generalize everywhere.
Published Apr 15, 2026Added Sep 7, 2026
One important behavior improved while most process measures stayed null
A 27-school cluster randomized trial found that adding R-CITY equity-focused lessons and teacher coaching to an existing social-emotional program improved one observed student outcome—physical aggression—and one teacher-reported outcome, while most other process and system measures did not significantly change. The result is valuable precisely because the nulls remain visible: a successful intervention package does not automatically reveal which component or mechanism produced the behavioral effect.
Published Mar 31, 2026Added Sep 7, 2026
More realistic agents can make a model harder to identify
A study of agent-based model calibration found that adding heterogeneous agent parameters made it harder to recover the true parameter values from model outputs. The problem is not simply computational complexity: different parameter combinations can become observationally indistinguishable when the available summary statistics are not informative enough. The result is a useful warning that richer agent descriptions do not automatically improve explanatory value; model detail has to be supported by observations that can identify it.
Published Mar 16, 2026Added Sep 7, 2026
Hierarchical FCMs can compress many indicators without pretending the links are causal
A hierarchical Fuzzy Cognitive Map grouped large numbers of financial indicators into a smaller set of interpretable concepts, then propagated risk across that concept layer. On two bankruptcy datasets it remained competitive with stronger nonlinear models while improving interpretability and monitoring performance. The architecture is useful for complex decision support because it separates indicator detail from higher-level reasoning, but the learned concept links are predictive or associational rather than causally identified.
Published Sep 6, 2026Added Sep 6, 2026
Germany is turning hybrid-threat attribution into a broader defensive architecture
Germany is planning a wider anti-sabotage package combining mobile counter-drone units, stronger authorities for critical-infrastructure operators, a national cyber sensor network, and expanded AI-supported surveillance. The development is notable because the response is becoming a sociotechnical defense system rather than a statement of condemnation. It is still a plan, however: implementation quality, public reassurance, adversary behavior, legal sufficiency, and privacy consequences remain separate outcomes that cannot be inferred from the announcement.
Published Sep 5, 2026Added Sep 6, 2026
Influence can begin as a job, trip, or opportunity—not an argument
Taiwan’s Straits Exchange Foundation warned that youth-facing job offers, tourism, and cultural exchanges involving China can combine economic opportunity with scam, detention, data, and political-influence risks. The useful analytical point is the pathway: an initial behavioral ask can look ordinary long before any ideological or political objective becomes visible. The warning comes from a Taiwan institution engaged in cross-strait competition, so legitimate exchanges, criminal scams, political influence, and state direction should remain separate evidentiary questions.
Published Sep 1, 2026Added Sep 6, 2026
A digital twin needs a continuing connection to the real world
A new architecture paper draws a useful boundary between a crowd simulation and a crowd digital twin. A true twin requires sensing and synchronization with the physical crowd and its environment, alongside prediction and decision support; representational detail alone is not enough. The framework is a roadmap rather than evidence that a fully validated crowd twin already exists, but it provides a stricter standard for claims about synthetic audiences and digital twins.
Published Aug 25, 2026Added Sep 6, 2026
Synthetic populations are starting to include values and ideology—not just demographics
A Bayesian population-synthesis framework adds values, ideology, opinions, and life priorities to the demographic attributes normally used to construct synthetic populations. In a Barcelona demonstration, the method reproduced important demographic and motivational distributions and dependencies across large samples. This creates a more empirical starting point for agent models than demographic personas alone, but it does not show that the resulting synthetic people will make human-like decisions or respond correctly to interventions.
Published Jul 29, 2026Added Sep 6, 2026
Exposure to untrustworthy sources fell about 70%; measured political attitudes did not
A multimonth randomized Meta field experiment reduced Facebook and Instagram feed exposure to repeatedly untrustworthy sources by about 70%. Despite that large change in what people saw, researchers found no measurable effects across numerous preregistered political attitudes and beliefs, including among users with high prior exposure. The result does not show that misinformation never matters. It shows why exposure is an intermediate outcome: a technically successful feed intervention can still produce a null at the attitude level.
Published May 18, 2026Added Sep 6, 2026
Let an FCM learn without letting it silently invent the graph
A human-centric Fuzzy Cognitive Map framework separates reasoning over a fixed graph from a learning loop that updates weights and from explicit bounded expert correction. That architecture allows adaptation while preserving inspectable structure and a record of human intervention. It offers a useful governance pattern for decision support, but the medical benchmarks do not establish that learned edges are causal or that the method transfers unchanged to behavioral or national-security systems.
Published Mar 27, 2026Added Sep 6, 2026
Ordinary citizens persuaded political opponents surprisingly often—but could not tell when they were good at it
An incentivized persuasion study asked ordinary citizens to write messages for people who disagreed with them politically. Nearly 30% of the messages shifted opinion in the intended direction and about 11% backfired. Perspective-taking and personal narratives were the strongest predictors of success, yet senders were poor judges of their own persuasive ability. The study shows meaningful message-level heterogeneity, but those features were not independently randomized, so they should not yet be treated as proven causal ingredients.
Published Mar 24, 2026Added Sep 6, 2026
AI labels can change how people engage with the same content
Two randomized experiments found that labeling social-media content as AI-generated reduced affective and intended behavioral engagement, especially for emotional posts. Disclosure timing also mattered for some AI-enhanced content. The finding suggests that provenance labels can change the experience of a message rather than simply describe it. Because the evidence comes from immediate experimental responses, it does not establish durable changes in trust or observed behavior across an entire platform.
Published Feb 4, 2026Added Sep 6, 2026
Better information reduced party-cue influence; simply thinking harder did not
Two large experiments separated access to policy information from prompts to think more effortfully. Providing substantive information reduced reliance on party cues by roughly 35% on average, while encouraging additional thought alone produced little independent effect. The result suggests that some apparent heuristic reasoning reflects an information constraint rather than simple unwillingness to deliberate. The outcomes were policy opinions, not durable political behavior, and effects varied across issues.
Published date unavailableAdded Sep 6, 2026
A government nudge changed when civil servants complied more than whether they complied
A Buenos Aires field experiment redesigned the notice attached to real freedom-of-information requests using clarity, salience, deterrence, and social-norm cues. Overall compliance did not significantly increase, but responses exactly at the final legal deadline rose by 10.7 percentage points. Government training workshops diluted the treatment contrast and the researchers also found intra-agency spillovers. The study shows how a treatment can shift the timing distribution while concurrent institutional activity becomes part of the causal story.
Published date unavailableAdded Sep 6, 2026
Causal discovery can help explain an LLM simulation before it explains society
CAMO is a new framework for tracing macro-level outcomes in LLM-agent simulations back to smaller sets of simulated mechanisms. It extracts factors from simulation records, learns compact explanatory graphs, and uses simulator-internal counterfactuals to probe ambiguous relationships. That is useful for auditing complex agent systems and identifying model-internal levers. The critical boundary is external validity: a causal pathway recovered inside a simulator is evidence about how that simulator works, not proof that humans follow the same pathway.
Published Sep 4, 2026Added Sep 5, 2026
Pacific leaders’ missile-test response shows why reassurance is part of strategic signaling
Pacific Islands Forum leaders expressed concern over China’s ICBM test and called for greater transparency and at least 24 hours’ advance notification, while Nauru dissented and Beijing rejected the criticism. The episode shows that a military demonstration communicates more than capability: warning, consultation, predictability, and perceived respect shape how regional audiences interpret it. Public concern is observable, but the episode does not yet establish a durable change in alignment, deterrence, or security behavior.
Published Sep 3, 2026Added Sep 5, 2026
Movement and environmental feedback can reverse cooperation dynamics
A mathematical public-goods model shows how spatial movement can reverse seemingly intuitive cooperation dynamics. Under some conditions, diffusion increased cooperation and public-good availability, while strong movement of cooperators toward resource-rich areas reduced both cooperation and environmental quality. The result is theoretical rather than a human behavioral estimate, but it illustrates an important systems lesson: changing where people cluster, connect, or move can alter the incentives and feedback that determine collective outcomes.
Published Sep 2, 2026Added Sep 5, 2026
Rich LLM digital twins still predicted individual human behavior weakly
A Science Advances mega-study built individualized LLM digital twins from more than 500 prior answers per person and tested them across 19 preregistered studies, 164 outcomes, and 1,784 participants. Average twin–human prediction correlations were only about .20, with systematic problems including stereotyping, ideological bias, hyper-rationality, and overly compressed responses. The study tests current approaches rather than placing a ceiling on future systems, but it is strong evidence that rich personal context and fluent personas do not automatically produce reliable individual behavioral prediction.
Published Aug 27, 2026Added Sep 5, 2026
Saying “you” can trigger reactance and persuade less than saying “I”
Across seven experiments, recommendations framed as “you’ll like this” were often less persuasive than first-person statements such as “I like this.” Second-person wording increased psychological reactance, particularly among people already prone to reactance and when the recommender had little reason to know the audience’s preferences. The effect largely disappeared for close friends. The finding suggests that personalization without legitimate audience knowledge can feel controlling rather than relevant, although most outcomes were evaluations or intentions rather than durable real-world behavior.
