Stefaan Verhulst

Book by Daniel P. Aldrich: “What if society could move past traditional “gray infrastructure” approaches—like seawalls and prisons—to address problems such as climate change, terrorism, and crime? In Beyond Common Ground, Daniel P. Aldrich argues that social infrastructure—physical and virtual spaces including parks, libraries, and radio programs—offer a more effective alternative by fostering coproduction and multiple benefits.

Drawing on qualitative and quantitative evidence from nine countries across Africa, Asia, and North America, this book demonstrates how these systems build social capital and resilience—and proposes practical policies for implementation. Case studies show that facilities in Japan, such as the elder-led Ibasho center, reduced mortality during the 2011 tsunami and accelerated recovery. Greening initiatives in Philadelphia mitigated crime, while radio countered extremist recruitment in the Sahel. Beyond Common Ground positions social infrastructure as a “polysolution” for our interconnected crises, urging society to stop treating it as a Cinderella service and prioritize equitable distribution…(More)”.

Paper by Andrew Caplin: “Many important economic decisions depend not only on acquiring information, but on determining which questions are worth asking before committing to an action. Individuals and organizations must organize inquiry: they select questions, interpret answers, and update beliefs in ways that determine whether decisive distinctions are identified or missed. Despite its importance, the organization of inquiry is largely absent from formal economic analysis.

This paper studies decision-making under \emph{diagnostic uncertainty}, in which agents must learn not only about payoff-relevant states, but about which questions are informative. We develop a model in which agents choose questions sequentially prior to commitment, while facing uncertainty over a latent diagnostic structure that governs how questions generate answers. Before each inquiry step, agents may incur cognitive cost to refine beliefs about this diagnostic structure. The resulting problem is a finite-horizon dynamic program over inquiry, in which costly attention is allocated to belief transformations over diagnostic structure rather than directly to payoff-relevant states.

Two canonical diagnostic geometries isolate distinct planning margins. In one, value depends on locating a decisive question; in the other, on maintaining a correct sequence of questions. Both environments admit closed-form solutions and yield a common representation in which optimal inquiry increases the probability of success relative to uninformed search.

The framework identifies a distinct margin of economic behavior—planning under diagnostic uncertainty—that becomes increasingly important in environments where answers are abundant but the organization of inquiry remains scarce…(More)”.

Article by Stephen Sims: “On June 13, 2025, Iran’s air defense network was largely silent in the face of an intense Israeli bombing campaign. Just before the attack, swarms of explosive quadcopter drones, launched by Israel from inside Iranian territory and acting on vast troves of intelligence sifted with the use of AI to select targets, had taken out Iran’s radar systems and numerous missile sites. Israel’s one-two punch made Iran an object lesson in how a combination of AI and drones is blazing a new trajectory for international politics.

Not long before, on June 1, Ukraine had employed a strikingly similar tactic, using cargo trucks with false inventories to smuggle drones deep into Russian territory. The drones had been trained using AI to recognize Tu-95 “Bear” bombers based on photographs taken of a decommissioned version in a Ukrainian air museum and to recognize the weakest point of the bombers, often the fuel tanks in the wings. This allowed the drones, flying first autonomously and then with human pilots, to strike Russian bombers with high precision as far away as Siberia.

In the grand scheme of geopolitics, these events were small. The conflict between Iran and Israel ended up being more like glorified shadowboxing than real war, and the Ukrainian strike on Russia did nothing to change the relentless, grinding attrition of the front line. These events are not obvious ruptures in international politics, as when nuclear fire consumed Hiroshima and Nagasaki in August 1945. That moment announced with dreadful clarity that the future of war and strategy would never be the same. The use of AI coupled with drones, however, is more like Sputnik in 1957, a seemingly small event that nevertheless drastically altered the human relationship to technology.

Heidegger once remarked that the first images of Earth from the Moon shocked him because they revealed a new way of grasping the human condition, drained of direct human experience. AI-enabled drone strikes carry a similar symbolic charge: they represent war drained of direct human contact…(More)”.

Article by Emanuel Maiberg: “After months of heated debate and previous attempts to restrict the use of large language models on Wikipedia, on March 20 volunteer editors accepted a new policy that prohibits using them to create articles for the online encyclopedia. “Text generated by large language models (LLMs) often violates several of Wikipedia’s core content policies,” Wikipedia’s new policy states. “For this reason, the use of LLMs to generate or rewrite article content is prohibited, save for the exceptions given below.” The new policy, which was accepted in an overwhelming 40 to 2 vote among editors, allows editors to use LLMs to suggest basic copyedits to their own writing, which can be incorporated into the article or rewritten after human review if the LLM doesn’t generate entirely new content on its own. “Caution is required, because LLMs can go beyond what you ask of them and change the meaning of the text such that it is not supported by the sources cited,” the policy states. “The use of LLMs to translate articles from another language’s Wikipedia into the English Wikipedia must follow the guidance laid out at Wikipedia:LLM-assisted translation.” I previously reported about editors using LLMs to translate Wikipedia articles and  Wikipedia editor, Ilyas Lebleu, who goes by Chaotic Enby on Wikipedia and who proposed the guideline said that it seemed unlikely the policy will last because previously the editor community has been divided on the issue…(More) “.

Book by Marina Nitze, Matthew Weaver, and Mikey Dickerson: “When the system breaks, what do you do? You’re in the middle of a meltdown. The platform is down, the phones are ringing, the headlines are brutal, and your team is looking to you for answers. The usual playbooks—careful planning, expert consultation, bold strategy—aren’t working. What if we told you that instead of the end of the world, this is your moment to create lasting, transformative change?     
 
Crisis Engineering is your field guide to leading through the chaos—and coming out stronger than before. Drawing on decades of experience inside some of the most complex systems in industry and government, Marina Nitze, Matthew Weaver, and Mikey Dickerson, of the crisis engineering firm Layer Aleph, reveal their powerful, hands-on framework for navigating high-stakes crises.
From the rescue of HealthCare.gov to wildfire response and pandemic logistics, this book offers real-world stories, practical tools, and hard-won insights into how complex systems fail—and how to help them recover. You’ll learn: 

• How to identify the 5 signals of a crisis—and use them to your advantage
• Why traditional leadership instincts fail under pressure—and what to do instead
• How to stand up your own crisis engineering effort when it matters most

Whether you’re in tech, government, healthcare, or any other critical system, Crisis Engineering gives you the mindset, tools, and vocabulary to lead with clarity and create lasting change…(More)”.

Paper by Adnan Firoze, et al: “Historically, only resource-rich U.S. cities have collected data about where their public trees are, usually through labor-intensive manual surveys or via coarse canopy-cover estimation. However, a significant portion of city trees are on private property, making them difficult to quantify with surveys, yet they contribute uniquely to species diversity and ecosystem service distribution. Further, canopy-cover estimation cannot provide information about tree density, locations of trees across different land types, or changes in tree counts. Cities are under continual change, and the mean mortality rate of urban trees is twice that of rural trees.Thus, frequent updating of tree analytics is critical for sustainable, habitable cities.

Method.  Recent advances in computing—in particular, generative artificial intelligence (AI)—have enabled our multidisciplinary team, spanning computer science, engineering, and forestry, to develop a first-of-its-kind computational method that can individually locate and maintain an inventory of trees in at least 330 U.S. cities (Figure  1). Using satellite data, this approach can complete the inventory process in less than a day of automated computing. Individual trees are challenging to discern in satellite images due to occlusion and resolution limitations, which in turn limits traditional segmentation-based approaches. Our approach leverages several key insights to enable a scalable generative AI solution. First, a frequent capture rate of satellite imagery (e.g., daily, monthly, etc.) provides spatiotemporal vegetation footprints, yielding richer information than single images. Our method includes a deep spatiotemporal vegetation cover classification using satellite images that classifies a city into tree, grass, and background, followed by a cluster-creation process and then individual tree localization using a set of conditional generative adversarial networks (cGANs). Further, our method can be applied to current or archived satellite imagery, allowing for change detection and historical analysis…(More)”.

Article by Sébastien Bourdon and Antoine Schirer: “On the Strava fitness app, there are individual cases that raise concerns, some of which Le Monde has already documented. For example, the naval officer who went jogging in the middle of the Mediterranean Sea, thereby revealing, at the moment of recording his run, the exact position of the Charles de Gaulle aircraft carrier strike group. Or the military personnel who exercised on the docks of the Ile Longue base, where France’s nuclear ballistic missile submarines (SNLE) operate, and whose activities provide information about patrol schedules. Or even the bodyguards of the French, American and Russian presidents, whose sporting activities make it possible to track these heads of state and, in some cases, to anticipate their movements.

And most significantly, Le Monde has uncovered several thousand servicemembers, of all ranks and specialties, who, by exercising and running in far-flung corners of the world and sharing their performances on their public Strava profiles, reveal the activities of the French armed forces as a whole, from the most banal to the most sensitive…(More)”.

Article by Ridhi Purohit, and Judah Axelrod: “Before building a new house or apartment building, residents and developers must first ensure their project adheres to local zoning rules. 

But zoning documents are often difficult for both applicants and local government staff to navigate. This can lead to extended back and forth between staff and applicants, slowing down the permit process at a time when most communities don’t have enough housing.

Some local leaders have expressed interest to Urban researchers in the promise of using generative AI to make zoning codes easier to understand, whether through screener tools that scan permitting applications to ensure they are complete, or chatbots that can answer development questions. However, leaders are hesitant to adopt tools that haven’t been properly vetted for the quality of the information they produce.

To test how well generative AI tools could interpret zoning codes, we ran a benchmarking exercise that evaluated the capabilities of various large language models (LLMs), building on previous work that explored whether machine learning could help automate the collection of standardized zoning data. To do this, we developed a set of zoning and permitting queries for Minneapolis—a city with a complex, 467-page zoning code and with zoning processes familiar to our team…(More)”.

Article by Karine Perset and Anna Pietikäinen: “…Six insights about AI regulatory sandboxes from around the globe

1. AI sandboxes are not uniform

• According to the Datasphere Initiative, three primary types of sandboxes are emerging worldwide, especially within the context of AI. Regulatory sandboxes: Collaborative processes where regulators work with innovators to test innovations under regulatory supervision.
• Operational sandboxes: Testing environments and infrastructure where data can be hosted and accessed in controlled conditions.
• Hybrid models: Combining regulatory oversight with operational capabilities, sometimes offering infrastructure and operational spaces for testing and experimentation (e.g., “supercharged sandbox” in the UK).

These models intervene at different phases of the policy and regulatory lifecycle. Some are employed before formal regulation to identify gaps and suggest necessary updates. Others operate during the development process, supporting iterative regulatory design. Some focus on helping understand legal obligations and ensure regulatory compliance, such as under the EU AI Act. Sector-specific sandboxes are also common, with countries adopting different approaches depending on regulatory priorities and institutional settings. Across these models, regulatory waivers are frequently used to enable experimentation under regulatory supervision. 

Several experimentation-related initiatives, such as regulatory testbeds, living labs, or policy prototyping, share certain features and objectives with regulatory sandboxes. What truly distinguishes sandboxes is that they are the most institutionalised form of regulatory experimentation, usually led by regulators and integrated with regulatory supervision..(More)”.

Article by Ian Leavitt: “Connecticut is one of many U.S. states increasingly using cross-sector data-sharing to improve public health outcomes, with the Prevention Data Portal a prime example of the state’s efforts. The portal demonstrates how state leadership buy-in, paired with expertise from nongovernmental partners, can overcome common barriers to cross-sector data-sharing and produce clear and meaningful information to help shape state policy—without requiring large new investments in data infrastructure or centralized control. Insights gained from Connecticut’s experience can inform the work of other state public health agencies as they look to expand cross-sector data-sharing through creative partnerships, simple use cases that show value early, and other avenues within their own states.

The Prevention Data Portal, launched in 2018 by Connecticut’s State Epidemiological Outcomes Workgroup (SEOW), houses data from local, state, and federal sources to advance health promotion and substance use prevention in the state. It provides free access to a wide range of data-driven products about populations in Connecticut, including epidemiological profiles, data stories using multiple streams of information, and infographics on mental health, substance use, suicide, gambling, and other public health topics. The portal is supported through federal block grant funding and partnerships between the SEOW and the Connecticut State Department of Mental Health and Addiction Services (DMHAS), the Center for Prevention Evaluation and Statistics (CPES) at UConn Health, and the Connecticut Data Collaborative (CTData).

In an ongoing collaboration with states to improve public health data, The Pew Charitable Trusts commissioned interviews with Connecticut officials and other participants in the conceptualization, creation, and use of the SEOW Prevention Data Portal. Those interviews produced several key takeaways…(More)”.