The editorial draws a careful distinction between two fundamentally different categories of AI use. The first involves non-generative tasks, such as data classification, pattern sorting, and copyediting. In these applications, AI functions much like a calculator or a statistical software package: it processes information according to defined procedures, and a human remains responsible for interpreting the results. The authors judge this kind of use to be broadly acceptable. It accelerates work that would otherwise consume countless hours, and it does so without displacing the judgment that makes research meaningful. A tool that cleans up prose or organizes a dataset does not decide what a study means or whether its conclusions are trustworthy.

The second category is far more troubling in the authors’ view. When researchers turn to generative AI to produce scientific ideas, seed research questions, or draft the interpretive substance of a manuscript, the machine stops being an instrument and starts being a substitute. The editorial argues that this kind of use deserves substantially more scrutiny than it currently receives, and that it should be disclosed in scientific manuscripts so that readers, reviewers, and the wider community can evaluate how much of the work originated in a human mind. Transparency, in this framework, is not a bureaucratic burden but a precondition for trust.

The authors go further than mere disclosure. They propose that certain uses of generative AI should be disallowed outright. Specifically, they identify the core intellectual, creative, ethical, interpretive, and accountability-bearing work of research as territory where AI must not replace human interpretation and judgment. A scientific paper is not merely a collection of sentences; it is a chain of responsibility that runs from the initial question through the gathering of evidence to the final claim about what that evidence shows. If any link in that chain is forged by a system that cannot be held accountable, the entire structure is weakened. Responsibility, the editorial insists, cannot be delegated to software.

Among the specific prohibitions the authors propose are some that may surprise researchers who have already begun experimenting with AI brainstorming tools. They argue that scientists should not use AI to generate their own initial research ideas or research questions. The origin of a study, they contend, should be a human act, rooted in curiosity, observation, and lived scientific experience. Likewise, they argue that ethical determinations, risk assessments, and policy recommendations should never be outsourced to generative systems. These are precisely the tasks where errors carry the greatest consequences for people, communities, and the environment, and where the absence of human accountability would be most damaging.

One of the most striking arguments in the editorial concerns serendipity, the happy accidents that have repeatedly driven scientific progress. The authors warn that if every researcher relies on the same AI tools with similar inputs, the diversity of thought that produces unexpected discoveries will shrink. Generative models are trained to produce probable outputs based on known patterns in existing data. They are, by design, engines of the average and the expected. Science, by contrast, has often advanced by departing from prior assumptions and breaking free of typical patterns. Penicillin, cosmic microwave background radiation, and countless other breakthroughs emerged not from the most probable answer but from an anomaly that a human noticed and refused to ignore. A research ecosystem optimized for probable outputs is an ecosystem primed to miss exactly those anomalies.

This concern leads the authors to a broader warning about optimization. Science, they argue, should not be unthinkingly optimized as if it were a manufacturing workflow producing standardized outputs. A factory benefits from consistency, repeatability, and the elimination of variation. Science requires the opposite: a wide diversity of inputs, approaches, and ideas. When the same models, trained on the same corpora, are applied across millions of research projects, the risk is a homogenization of the scientific imagination itself. Questions that fall outside the statistical grain of the training data become less likely to be asked at all, not because they are unimportant but because no human curiosity set them in motion.

The editorial does not call for a retreat from technology. Its authors are careful to preserve a role for AI that is genuinely augmentative. Non-generative applications, from classifying large datasets to polishing language, can free researchers to spend more time on the work that only they can do: conceiving questions, designing studies, weighing evidence, and taking responsibility for conclusions. The problem arises not from computation but from substitution, not from assistance but from abdication. In this sense, the proposed norms are less a restriction on AI than a defense of the human role that gives AI its value in the first place.

The authors distill their argument into a clear standard for the scientific enterprise. According to the editorial, the originating research question and the final evidentiary judgment in every scientific publication should be produced by human minds. Everything between those two poles may, with appropriate care and disclosure, involve machine assistance. But the beginning and the end of the scientific process, the spark of inquiry and the verdict on the evidence, must remain human. This framing offers journals, institutions, and funding agencies a practical benchmark against which policies on AI use can be measured, and it gives individual researchers a simple test to apply to their own workflows.

As generative AI becomes more capable and more deeply embedded in the research infrastructure, debates of this kind are likely to intensify. The PNAS Nexus editorial contributes a principled starting point: AI should amplify human science, not quietly absorb it. By distinguishing acceptable non-generative uses from the generative applications that demand scrutiny and disclosure, and by drawing firm lines around idea generation, ethical judgment, and accountability, Branas and Levine offer the scientific community a framework for preserving what makes research trustworthy. The message to researchers is unambiguous. Use the tools to work faster and see further, but never hand over the questions, the judgment, or the responsibility that define science itself.

Subject of Research: Norms for the appropriate use of artificial intelligence in scientific research

Article Title: Augmentation not abdication, when using AI in research

Article References: Augmentation not abdication, when using AI in research. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: artificial intelligence, generative AI, research ethics, scientific publishing, PNAS Nexus, authorship, serendipity, research norms, AI disclosure, human-led science, accountability, scientific method

The study, led by Lennart Luettgau, Hannah Kirk, Kobi Hackenburg, and colleagues, set out to test one of the most persistent fears about conversational AI: that hallucinations, biases, and confident-sounding errors could quietly distort what citizens believe to be true. Concerns about AI’s influence on democratic processes have intensified as large language models have become a common first stop for information seekers. Yet the empirical evidence about what actually happens when ordinary people use these tools to research politics has remained thin. This experiment was designed to fill that gap with controlled, measurable comparisons.

The researchers recruited UK residents online in two waves, drawing 1,711 participants in the first and 1,147 in the second. Each participant was asked to research true or false claims on two of four politically charged topics: climate change, immigration, criminal justice, and COVID-19 policy. The two topics a participant did not research served as within-subject controls, allowing the team to separate the effects of doing research from the effects of simply answering questions twice. This design gave the study a level of internal rigor that surveys and observational analyses of AI use often lack.

Participants were randomly assigned to one of four research tools: the chatbots GPT-4o, Claude-3.5, or Mistral, or conventional Google search. Before and after the research phase, everyone answered knowledge questions covering all four topics. By comparing belief in accurate statements and in misinformation across the before-and-after measurements, and across the researched and control topics, the researchers could quantify precisely how much genuine learning occurred under each condition — and whether any of the tools pushed people toward false beliefs.

The headline result is striking in its symmetry. Research increased belief in true information and decreased belief in misinformation regardless of which method participants used. Chatbot users did not walk away more misinformed than search users, and search users did not outperform chatbot users in factual accuracy. For a technology frequently accused of poisoning the information environment, the finding suggests that, at least for structured research tasks, large language models perform on par with the self-directed internet searching that has been the default for two decades.

There was one measurable difference between the modalities, and it favored the machines. The research process was 6 to 10 percent faster when participants used chatbots rather than Google search. That efficiency gain may sound modest, but multiplied across millions of information-seeking interactions, it hints at why conversational AI is displacing search as a research tool: it compresses the time between a question and a usable answer without, in this study, exacting a price in accuracy.

Just as notable is what did not change. Working with chatbots or with Google search did not alter participants’ trust in institutions, experts, the media, or technology itself. Critics have worried that AI tools might either inflate blind faith in expertise or erode it further; the data showed no such movement. The research phase left the landscape of institutional trust essentially untouched, suggesting that a single structured research session, however conducted, is not enough to shift these deeper attitudes.

The study did detect one shift that will draw scrutiny from across the political spectrum. Both chatbot and search conditions increased agreement with progressive views and decreased agreement with conservative views among all participants, regardless of the participants’ own political leaning. Because this pattern appeared in both modalities, it cannot be attributed to AI bias alone; something about actively researching claims on these topics moved people in the same ideological direction whether they consulted a language model or a search engine. The authors flag this as an important observation, but one that requires careful interpretation rather than alarm about chatbots specifically.

The researchers are candid about the limits of what their experiment can claim. They did not test models known to carry political biases, so the results cannot be generalized to every system on the market. Neither the chatbot condition nor the search condition reflected today’s AI-enhanced, personalized search experiences, which blend generative summaries into results pages and tailor output to individual users. And the study examined structured research tasks — verifying specific true or false claims — rather than the open-ended political conversations people increasingly have with AI assistants, where sycophancy, framing effects, and conversational drift could behave very differently.

Even with those caveats, the findings land at a consequential moment. Debates about AI governance and democratic integrity are unfolding in legislatures and standards bodies around the world, often driven by worst-case assumptions about what chatbots do to political knowledge. This study suggests that for one of the most common use cases — a citizen researching political claims before forming a judgment — conversational AI may perform on par with self-directed Google search, with no apparent cost to factual political knowledge. If that pattern holds as models and search experiences evolve, the policy conversation may need to shift from whether people should use AI for political research to how the tools should be designed, audited, and monitored as they become an ordinary part of democratic life.

Subject of Research: Effects of AI chatbot use on accurate political knowledge and beliefs compared with internet search

Article Title: AI chatbots may increase accurate political knowledge

Article References: AI chatbots may increase accurate political knowledge. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: AI chatbots, large language models, political knowledge, misinformation, GPT-4o, Claude-3.5, Mistral, Google search, UK election 2024, democracy, PNAS Nexus, AI governance

The research focuses on three distinct categories of hydrological whiplash. The first, drought–flood alternation events, describes sequences in which prolonged water scarcity and inundation repeatedly trade places within a defined period. The second, drought-to-flood transitions, captures situations in which a drought is abruptly terminated by heavy rainfall, often overwhelming soils that have lost their capacity to absorb water. The third, flood-to-drought transitions, follows the opposite trajectory, with floodwaters receding into dry spells that can be equally damaging to agriculture and water supplies. By modeling all three types separately, the authors were able to map not just whether extremes are intensifying, but how the temporal structure of hydrological risk is shifting, a subtler question with profound implications for planning and adaptation.

The physical mechanisms behind these oscillations are well grounded in hydrology. During a drought, soils dry out, vegetation cover declines, and the land surface loses much of its sponge-like ability to soak up precipitation. When intense rainfall finally arrives, much of the water runs off directly into rivers rather than infiltrating the ground, amplifying flood peaks and increasing the risk of flash flooding and landslides. Conversely, a flood can saturate reservoirs and aquifers only briefly; if the wet episode is short and followed by sustained heat and evaporation, the apparent water surplus evaporates quickly, leaving communities exposed to renewed scarcity. This is why the sequence of extremes matters as much as their individual magnitude. A region that experiences a flood and a drought in the same season faces a fundamentally different risk profile than one that experiences either event alone.

The socioeconomic consequences of such rapid switching are equally important. Back-to-back extremes place compounding stress on crops, which may be damaged first by waterlogging and then by desiccation, or vice versa, within a single growing season. Yield losses under these conditions can exceed what would be expected from either hazard in isolation, because plants have little time to recover between shocks. Infrastructure faces a parallel problem. Flood defenses, drainage systems, and irrigation networks designed around the assumption of recovery time between events can be degraded by one disaster and then overwhelmed by the next. Emergency response capacity, financial reserves, and household savings are similarly depleted when disasters arrive in quick succession, leaving people and institutions with diminished ability to prepare for the following emergency.

Using projections of future climate alongside detailed hydrological and socioeconomic data, the researchers found that all three types of switching events are increasing as temperatures rise. The mean frequency increase of 10.09 percent per 0.5 degrees Celsius of warming provides a quantifiable scaling relationship between global temperature and the tempo of hydrological whiplash. Equally significant is the finding on spatial extent. The area affected each year by these events is projected to grow, meaning that more communities will be drawn into the cycle of alternating scarcity and inundation even if they have never previously considered themselves at risk from compound hydrological extremes.

The study identifies two regions that consistently emerge as national hotspots: the Yangtze Basin and the North China Plain. These are not random locations. The Yangtze Basin is one of the most densely populated and economically productive regions in the world, home to vast agricultural lands, major industrial centers, and an extensive network of rivers, lakes, and reservoirs. The North China Plain, meanwhile, is a critical grain-producing region that already faces chronic water stress, with groundwater depletion and variable monsoon rainfall making it acutely sensitive to any intensification of the drought–flood cycle. The convergence of high exposure, dense population, and agricultural importance in these two basins means that the projected increases in switching events carry implications that extend well beyond local boundaries, touching national food security and supply chains.

Perhaps the most sobering result concerns the distribution of risk. The analysis shows that low-income populations are likely to experience the steepest relative increases in exposure to rapid hydrological transitions. This pattern reflects a combination of geography and vulnerability. Poorer communities are often located in flood-prone lowlands, on unstable slopes, or in areas with inadequate drainage and water storage infrastructure. They also tend to have fewer resources for recovery, meaning that each successive disaster erodes their capacity to withstand the next. Wealthier households and municipalities, by contrast, can invest in flood barriers, insurance, backup water supplies, and rapid reconstruction, effectively insulating themselves from the worst consequences. The authors describe this dynamic as climate change acting as a regressive force, deepening the divide between the resilient wealthy and the vulnerable poor.

This regressive framing has important implications for how adaptation policy is designed. If exposure to compound extremes is concentrated among those least able to cope, then aggregate national statistics on disaster risk can mask severe inequities. A country may appear to be adapting well on average while its poorest regions fall further behind. The findings suggest that adaptation investments, from reservoir management and early warning systems to crop insurance and social safety nets, will need to be deliberately targeted at the communities and basins where the relative increase in switching-event exposure is greatest. They also suggest that the timing dimension of adaptation deserves more attention: systems designed to handle single hazards at a time may need to be rethought around the reality of back-to-back events with little recovery interval.

The methodological approach of the study also offers a template for future research. By integrating climate model projections with hydrological modeling and socioeconomic mapping, the authors moved beyond hazard assessment alone to examine who is exposed and how that exposure changes under warming. This integrated framing is increasingly recognized as essential, because the damage caused by an extreme event depends not only on its physical intensity but on the vulnerability and adaptive capacity of the population in its path. Quantifying how exposure shifts with income level, as this study does, provides policymakers with a concrete basis for prioritizing interventions rather than relying on broad generalizations about climate risk.

As global temperatures continue to climb, the study’s scaling relationship offers a sobering benchmark. Each half degree of warming does not merely add a proportional increment of risk; it accelerates the oscillation between opposite extremes, compressing the time available for ecosystems, economies, and communities to recover. The Yangtze Basin and the North China Plain may be the clearest hotspots in this analysis, but the underlying physics of dried soils amplifying floods and brief wet spells failing to replenish depleted reserves applies across many of the world’s densely populated river basins. The message of the research is ultimately one of urgency and equity: the whiplash between drought and flood is speeding up, and without deliberate, targeted adaptation, the burden of that acceleration will fall hardest on those with the fewest means to bear it.

Subject of Research: Increasing frequency and socioeconomic exposure of drought–flood alternation events under climate warming in China

Article Title: Drought-flood alternation under climate change

Article References: Drought-flood alternation under climate change. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: climate change, drought, flood, hydrological extremes, PNAS Nexus, Yangtze Basin, North China Plain, compound events, socioeconomic exposure, adaptation, food security, climate inequality

Alexandre S. Pires and colleagues set out to map what might be called the social norms of large language models. The research team examined 21 different LLMs, probing each system with a deceptively simple task: judge fictional people as good or bad after learning how those people behaved in a series of interpersonal dilemmas. The scenarios covered familiar moral terrain, such as whether a person chose to spend time and energy helping someone with a problem, or whether they gave food or money to another person in need. By varying the reputations of the people involved, the researchers could determine whether a model rewarded generosity universally or conditioned its approval on the character of the recipient.

The results revealed a broad consensus on one point and deep disagreement on another. Across the board, the models approved of helping and sharing with people who had already demonstrated themselves to be good. Cooperation with well-reputed individuals was, in the eyes of every system tested, virtuous conduct. The trouble began when the fictional actors interacted with individuals of ill repute. Here the models split into distinct camps, each implicitly endorsing a different philosophy about how decent people should treat those who have behaved badly.

Most, but not all, of the LLMs tested assigned a good reputation to people who cooperated with bad recipients. In this view, extending help even to a known defector is a laudable act, perhaps reflecting an ethic of unconditional kindness or a hope that generosity might reform the wrongdoer. But two models stood apart. Gemma 2 27B IT and Llama 3.1 8B tended to endorse shunning bad individuals, judging that cooperating with them was itself bad behavior. For these systems, morality is not just about what you give but about who deserves to receive it, and associating with the wrong people can stain your own standing.

The disagreements grew sharper when the researchers examined the opposite case: refusing to help or share with bad people. GPT-4o tended to penalize people for not cooperating with bad individuals, effectively treating withholding aid as a moral failing regardless of the recipient’s history. Llama 3.3 70B took the opposite stance, typically arguing that bad people should be punished for their prior lack of cooperation, so refusing to help them is entirely fine. Meanwhile, Gemini 1.5-Pro and Grok 2 proved inconsistent in their judgements, sometimes approving and sometimes condemning the same pattern of conduct depending on details of the scenario. For users seeking moral guidance, this means that identical questions posed to different popular chatbots can yield substantively different ethical verdicts.

Despite the noise, the researchers identified a discernible trend. Most LLM model families appear to be evolving toward a norm that social scientists call Simple Standing. Under this norm, cooperating is always good, and defecting against bad individuals is also deemed good. Simple Standing is permissive: it never requires anyone to punish or exclude wrongdoers, but it gives everyone license to walk away from them without guilt. It stands in contrast to stricter reputational codes found in human societies, such as norms under which failing to sanction a defector makes you complicit in their defection.

Why does this taxonomy matter beyond the philosophy seminar? Because norms are not merely descriptions of behavior; they are engines of it. The authors modeled what would happen if the various philosophies underpinning the LLMs’ judgements were universalized across an entire population. If everyone operated under the rules of Simple Standing, cooperation would indeed ensue, but it would not reach the high levels achieved under a sterner framework, one in which you should always fail to help bad people or be judged as bad yourself. That harsher regime, sometimes associated with norms of punitive standing, sustains near-universal cooperation by making association with defectors costly. In other words, the gentle, forgiving norm that most AI systems seem to be converging on produces a decent but less cooperative world than the unforgiving alternative.

The stakes rise when one considers how people actually use these systems. As more and more individuals turn to LLMs for advice on interpersonal matters, the ethical foundations behind the tools’ replies increasingly have the power to shape individual lives and, at scale, the social fabric itself. A person deciding whether to cut off a difficult relative, whether to give a second chance to an unreliable business partner, or whether to help a stranger with a bad reputation may be quietly absorbing the normative leanings of whichever model they consult. If a generation of users internalizes Simple Standing because their assistants endorse it, patterns of human cooperation, punishment, and reconciliation could shift in ways that no democratic process ever debated.

The study also uncovered subtler biases in how models render their judgements. LLM-based assessments depended on the gender and perceived cultural background of the recipients as well as the overall context of the fictional situation. The same act of helping or withholding could receive different moral evaluations depending on who was being helped, a finding that raises questions about consistency and fairness in AI moral reasoning. If a model’s verdict on your behavior shifts with the demographics of the people involved, the advice it gives may encode stereotypes that users never consented to and may not notice.

Can the problem be fixed with a simple instruction? Apparently not easily. The authors report that prompting interventions, such as instructing a model to adopt norms that promote overall cooperation, have a limited and inconsistent effect across LLMs. A nudge that steers one system toward a cooperative norm may leave another untouched or produce erratic behavior. This fragility suggests that the moral dispositions of these models are baked deeply into their training and alignment, and that shaping them deliberately will require more than clever phrasing in a user prompt. As AI systems become ubiquitous advisers on human relationships, the study implies, society will need to pay attention not only to what these machines know but to what, implicitly, they believe is right.

Subject of Research: Social norms and moral judgements of large language models regarding human cooperation

Article Title: AI’s social norms and their implications for society

Article References: AI’s social norms and their implications for society. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: large language models, social norms, human cooperation, AI ethics, PNAS Nexus, reputation, Simple Standing, moral judgement, GPT-4o, Llama, Gemini, prompting interventions

A team of researchers from the Massachusetts Institute of Technology, Princeton University, Queen’s University, and Penn State University believes it has found a way around that problem. In a new open access paper published in PNAS Nexus, the group describes a series of experiments conducted in a pressurized wind tunnel, a facility that allows scaled-down turbines to reproduce the flow physics of full-size machines operating in the open atmosphere. By raising the air pressure inside the tunnel, the researchers dramatically increased the density of the air, and with it the inertia of the flow around their model turbine, closing a gap between laboratory experiments and field conditions that has hampered wind energy research for decades.

The logic behind the approach rests on a principle known as dynamic similarity. A small turbine spinning in ordinary air does not behave like a giant machine standing in a windy plain, because the balance of inertial and aerodynamic forces is different at different scales. But when the air in the test chamber is pressurized, it becomes far denser, and the scaled model suddenly experiences forces that mirror those acting on a much larger turbine. In the new study, the team used a turbine measuring just 15 centimeters in diameter, but at pressures of up to 240 atmospheres the air density increased by a factor of 100 to 220 times. According to lead author John Kurelek, an assistant professor at Queen’s University, that means the experiments effectively tested a turbine that is, all else being equal, 15 to 20 meters in diameter, with the capability of reaching the equivalent of 35 meters.

The motivation for this experimental paradigm comes in part from the frustrations of field research. Michael Howland, MIT’s Jeffrey Cheah Career Development Professor and a senior figure in the project, has spent years developing models to simulate wind farm performance and to increase power output. In 2022, he demonstrated that accounting for the wake of individual turbines when controlling an entire wind farm could significantly boost electricity production. But that work demanded a lengthy and costly field campaign in which turbines at a real wind farm were intentionally misaligned with the wind for months, temporarily lowering the farm’s output, in order to understand how misalignment affects performance.

Wind energy, Howland explains, sits at a unique intersection of environmental flow, mechanics, aerodynamics, and meteorology, and that makes controlled experimentation extraordinarily difficult. Researchers want to isolate the effect of a single change, but wind farms operate in chaotic, turbulent environments where the weather is constantly evolving and turbines must react to conditions beyond anyone’s control. This complexity has left engineers uncertain about how precisely changes in the alignment between wind and turbine, or variations in the turbine’s tip speed relative to the wind, alter the power a machine produces. Many predictive models in use today are built on the assumption that turbines always face perfectly perpendicular to the wind, an assumption that is almost never true in the real world, even for modern turbines that continuously adjust their orientation in response to rapid directional shifts in the wind.

That uncertainty is exactly what the pressurized experiments were designed to resolve. Marcus Hultmark, a professor at Princeton University whose laboratory pioneered the study of scaled turbines in pressurized wind tunnels, notes that researchers have long debated which models best describe the output of wind farms, but without reliable experimental data to compare against, the field has struggled to advance. The new paper, he says, attempts to do both: settle questions about turbine performance and provide the missing experimental benchmark. Over several weeks of testing, the team ran the model turbine through a systematic series of operating conditions, varying its alignment with the wind and its control strategy to isolate how each factor influences power production.

The results carry direct implications for the economics of wind energy. The researchers found that power output could be significantly increased by adjusting the turbine’s tip speed based on its misalignment angle with the wind, a control strategy that is rarely employed at wind farms today but that could deliver performance gains at minimal added cost. Kurelek describes this as the clearest output of the experiments: new power maximums become achievable when a turbine is misaligned with the wind, simply through changes to the tip speed. Combined with optimization of the turbine’s alignment relative to the wind and the blade pitch angles, which control the airfoil’s angle of attack, the researchers estimate that such improvements could yield tens of thousands of dollars per turbine every year in additional revenue.

The experiments also served a second, equally important purpose: validating a new predictive model. Howland and his colleagues developed what they call a unified wind turbine model, building on previous work that produced a more general aerodynamic theory for wind turbines. The model simulates turbine performance across operating conditions without relying on the empirical corrections that have historically been baked into wind power models. Before running the experiments, Kurelek sent the dimensions of the wind tunnel and turbine setup to Howland, who used them to calculate the expected aerodynamics, forces, and power production. The close agreement between prediction and measurement marks the first experimental validation of the unified momentum model’s predictions about how turbines should be controlled under misalignment to achieve maximum power production.

One of the most striking features of the validated model is its computational efficiency. According to Howland, the model is fast enough to run on ordinary laptop computers, which means engineers designing and operating wind turbines anywhere in the world can use it to test different turbine designs and wind farm control strategies without access to supercomputers. That accessibility could accelerate the adoption of the control strategies identified in the study, allowing existing wind farms to be retuned for higher output rather than waiting for new hardware to be installed. Because the approach improves and validates the models that feed directly into wind turbine control protocols, its immediate impact applies to the vast installed base of turbines already generating power today.

Beyond the immediate gains, the researchers see the pressurized wind tunnel as a way to fill a persistent void in wind energy research. Howland describes a massive gap between idealized theoretical and simulation models on one side and full-scale testing in extremely complicated field environments on the other, a gap that he says nothing else currently fills except these pressurized experiments. By enabling high-throughput, controlled experiments that achieve the right physics in a time-efficient and low-cost manner, the technique could allow researchers to investigate a wide range of previously unanswered questions about turbine aerodynamics, wake behavior, and farm-level control. The work, supported in part by the Natural Sciences and Engineering Research Council of Canada, the National Science Foundation, and the MIT-GE Vernova Alliance, suggests that the path to more wind power may run not only through bigger turbines and new installations, but through smarter operation of the machines already turning in the wind.

Subject of Research: Pressurized wind tunnel experiments validating wind turbine aerodynamic models and control strategies for increased wind farm power output

Article Title: Pressurized experiments could help wind farms generate more power

Article References: Pressurized experiments could help wind farms generate more power. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: wind energy, wind turbines, pressurized wind tunnel, dynamic similarity, aerodynamics, wake modeling, turbine control, tip speed ratio, blade pitch, PNAS Nexus, MIT, renewable energy

The first component of the model concerns the emergence of potential perpetrators. Using demographic data drawn from the U.S. Census, the researchers estimated a per-capita rate at which motivated individuals arise in a given population, and this rate was estimated jointly with the other model parameters rather than fixed in advance. This joint estimation is a key strength of the Bayesian approach: uncertainty in one parameter, such as perpetrator emergence, can be constrained by observations that also bear on other parameters, such as firearm availability or population structure. By allowing all parameters to be inferred simultaneously from the historical record of events, the framework avoids the pitfalls of chaining together independently derived estimates and produces a coherent probabilistic picture of shooting risk across the United States.

The second component, access to firearms, proved to be the dominant predictor of risk in the model. Because comprehensive data on private gun ownership are difficult to obtain, the researchers used the local density of Federal Firearms License (FFL) holders as a proxy for firearm availability. This proxy emerged from the calibration as the single most influential factor shaping the geographic distribution of expected shootings, outweighing the contributions of population density alone or the estimated rate of perpetrator emergence. In practical terms, the model implies that the availability of a firearm may function as an important limiting factor in mass shooting risk: where guns are more accessible, the probability that a motivated individual can translate intent into a high-fatality event rises substantially, holding other conditions constant.

The third component captures the presence of suitable target populations, reflecting the observation that mass shootings typically occur in places where potential victims gather, such as schools, workplaces, retail spaces, and public venues. By combining these three ingredients within an agent-based architecture, the simulation generates stochastic realizations of shooting risk across counties, producing not just a single point estimate but a distribution of plausible outcomes. This probabilistic character is central to the study’s forecasting ambitions. Just as epidemic models produce ranges of possible outbreak trajectories rather than deterministic predictions, the shooting model yields expected event counts and their uncertainty, allowing researchers and policymakers to reason about risk in a statistically disciplined way.

To test the model, the authors compared its predictions against mass shooting events recorded in the Mother Jones database covering 1982 through 2024. This widely used database includes approximately 160 events with three or more fatalities, excluding the perpetrator, and deliberately excludes incidents categorized as terrorism, gang-related violence, or domestic disputes. The exclusion criteria matter: they define a specific and arguably more preventable class of public mass shootings, distinct from forms of gun violence with different underlying dynamics. The researchers found that the models assigned higher risks of shootings precisely to areas where shootings had historically occurred, demonstrating that the framework can discriminate between geographic areas with genuinely different levels of risk rather than merely reproducing population counts.

Beyond retrospective validation, the calibrated model identifies the areas facing the greatest expected burden in the coming years. The ten highest-risk areas by expected events per decade are Maricopa County, Arizona; Harris County, Texas; Dallas County, Texas; Los Angeles County, California; Kings County, New York; Clark County, Nevada; Orange County, California; Queens County, New York; Cook County, Illinois; and Philadelphia County, Pennsylvania. The list spans large metropolitan counties across the Sun Belt, the coasts, and the industrial Midwest, and reflects the combined influence of substantial populations, dense networks of firearm dealers, and the urban characteristics that concentrate suitable targets. Notably, several of these counties also rank among the most populous in the nation, but the model’s ranking is not a simple population map; the dominance of FFL density as a predictor reshapes the geography of expected risk in ways that pure demographic extrapolation would not.

According to the authors, agent-based models of this kind can outperform standard statistical baselines in identifying areas at elevated risk. The improvement arises from the models’ structural realism: instead of fitting a regression to correlate inputs with outcomes, the simulation encodes a causal narrative, a motivated perpetrator meeting a firearm and a target population, and lets the data determine how strongly each element drives observed events. When the fitted mechanism reproduces historical patterns better than atheoretical alternatives, confidence grows that the model captures something real about the process, not merely a statistical coincidence. That realism also makes the framework adaptable; as new data arrive or as policy environments change, parameters can be re-inferred and projections updated within the same Bayesian machinery.

The potential applications extend to the allocation of prevention resources. If the geographic distribution of risk can be estimated with quantified uncertainty, interventions such as threat assessment programs, community-based violence prevention, extreme risk protection orders, and law enforcement preparedness could, in principle, be targeted toward the communities where expected benefits are largest. The model’s emphasis on firearm access as the dominant predictor also speaks directly to ongoing policy debates. If access to a firearm is an important limiting factor in whether mass shootings occur, then measures that reduce availability in high-risk areas, or that introduce friction into the acquisition process, may carry disproportionate preventive value relative to interventions focused solely on identifying perpetrators in advance, a task that remains notoriously difficult given the rarity and unpredictability of individual events.

The study also carries important caveats and a broader scientific significance. An agent-based model calibrated on roughly 160 events over four decades cannot predict when or where any specific attack will happen, and the use of FFL density as a proxy for gun availability introduces measurement uncertainty that the Bayesian framework represents but cannot eliminate. Mass shootings are rare, socially complex events shaped by factors, including mental health, social contagion, media dynamics, and economic conditions, that no single model fully captures. Yet the value of the work lies in demonstrating that even deeply troubling social phenomena can be brought within the reach of formal, mechanistic, data-constrained science. By combining agent-based simulation with Bayesian calibration, the researchers offer a template for rigorously reasoning about rare violent events, and a tool that, its authors argue, may help direct scarce prevention resources toward the places where they can do the most good.

Subject of Research: Bayesian agent-based modeling of the geographic risk and potential severity of mass shootings in the United States

Article Title: Forecasting mass shootings

Article References: Forecasting mass shootings. (n.d.). Original publication

Image Credits: AI Generated

DOI: Not provided

Keywords: mass shootings, agent-based modeling, Bayesian inference, firearm availability, geographic risk, PNAS Nexus, gun violence prevention, Federal Firearms License, Mother Jones database, risk forecasting, public health, United States