Harvard Kennedy School

Governments around the world are developing digital sovereignty strategies to reduce their dependence on foreign technology providers and reassert control over their digital infrastructures. In the European Union, this has resulted in a number of regulations – including the Digital Markets Act, the Data Act, the AI Act, and related initiatives – that aim to reduce Europe’s reliance on US cloud hyperscalers while developing domestic alternatives. This article asks how digital sovereignty rules shape states’ bargaining with tech giants and what these rules tell us about these firms’ infrastructural power. Theoretically, it proposes an updated version of the Obsolescing Bargaining Model (OBM) for digital infrastructures, which analyses whether there are analogies between the expansion of US oil companies in the 1970s and today’s bargaining between EU governments and US hyperscalers. Empirically, it traces the expansion of four hyperscalers – Amazon Web Services, Microsoft Azure, Google Cloud, and Oracle Cloud – across Europe between 2018-2026, combining firm-level data on data centre investments with text-as-data analysis on earning calls and annual filings to capture how firms adapt and respond to increased regulatory pressure. Findings show that digital sovereignty rules have so far had limited impact on hyperscalers’ expansion. This is due to factors such as enduring power asymmetries between governments and US hyperscalers, a lack of coordination among governments themselves, and the intrinsic limits of regulatory instruments that discipline firm behaviour but do not govern market access. The article contributes to existing debates on the structural power of big tech companies, the geoeconomics of digital infrastructures, and the interaction of regulation and industrial policy to shape firm-government relations.

Intelligence organizations often fail to anticipate strategic developments even when relevant information is available because signals remain fragmented, ambiguous, overwhelming, or filtered through entrenched assumptions. The September 11 attacks, Russia’s 2022 invasion of Ukraine, and the October 7, 2023 Hamas attack illustrate how, even within some of the world’s most capable intelligence organizations, available indicators do not reliably translate into effective strategic warning, with outcomes shaped by differences in collection, interpretation, dissemination, and uptake. This article examines these failures and introduces AI-INT, or Artificial Intelligence Intelligence, as a response. AI-INT is an integrated, actor-specific synthesis-and-simulation layer within all-source analysis. It assembles and fuses relevant multi-source intelligence to build and update actor models, then uses them to run simulations at scale under systematically varied assumptions and conditions. It builds on wargaming’s value in deepening understanding, challenging assumptions, exposing blind spots, and exploring alternative trajectories by combining fused intelligence with LLM- and ML-based methods to expand the scenario space. As a bounded architectural illustration, the article presents SESBot, a Behavioral-AI framework for simulating the decision-making of selected Semi-State Terrorist Organizations (SESTOs). SESBot combines LLM-based agents with machine-learning behavioral anchors, illustrating one implementation of intelligence-grounded actor simulation within AI-INT. AI-INT’s promise lies not in predicting the future with certainty or replacing human judgment, but in helping analysts and decision-makers test assumptions, compare alternative scenarios, identify analytic blind spots, surface overlooked trajectories, connect them to monitorable indicators, and communicate uncertainty rigorously.

Evidence from V-Dem suggests cases of growing illiberalism in recent decades - but also some countries where democracy has proved stable or even strengthened during these years. What has been driving these diverse trajectories? This chapter starts to unpack the theoretical debates and considers new empirical evidence analyzed using mixed methods to explore the links potentially connecting populist governments and backsliding. Part I defines the concept and logic of populist discourse and summarizes the previous theoretical and empirical literature. Part II compares the position of a wide range of political parties towards both liberal democracy and populism. The comparison further confirms that many, but not all, political parties employing populist discourse often oppose checks and balances on executive power, as well as liberal democratic principles, norms and practices. Part III analyses longitudinal cross-national data during the first quarter of the 21st century to determine the dynamic trajectories of democratic regimes worldwide under both populist and non-populist governments. To replicate previous studies in Europe and Latin America, difference-in-difference time-series panel regression models are employed to estimate causal effects through a quasi-experimental technique using observational data from treatment (populist-led governments) and control groups (non-populist governments) in power for heterogeneous periods during the last 25 years. Part V explores the processes of backsliding in more depth, comparing six case-studies of populist governments, including longitudinal trends in the horizontal and vertical checks designed to prevent the executive abuse of power. The selected examples include presidential executives in Jair Bolsonaro’s Brazil, Recep Erdogan’s Turkey, and Donald Trump’s USA, as well as prime ministerial governments in Viktor Orban’s Hungary, Narendra Modi’s India, and Jarosław Kaczyński’s Poland. The conclusion in Part VI summarizes the key findings and considers their broader theoretical and policy implications. The chapter provides further evidence confirming the links connecting the rhetoric of populist parties with their endorsement of illiberal policies; at the same time, many exceptions to these findings can be observed, suggesting that diverse varieties of populist parties exist, even among neighboring states. The paper is drawn from Chapter 4 of my forthcoming book on The Cultural Roots of Democratic Backsliding (OUP 2027)

Private-sector development finance institutions (DFIs) have been asked over the past decade to scale climate finance dramatically while maintaining their traditional development mandate. The two objectives are linked but not the same, and little is known about how the institutions delivering both manage the trade-offs between them. This paper asks whether the two mandates are compatible in practice, what DFIs are doing to manage the tension, and what governance would make it more likely that they reinforce rather than displace each other. It draws on analysis of climate and development finance flows and on interviews with senior executives at seven bilateral DFIs and EBRD, alongside development finance experts.

The central finding is that the balance between the two mandates is unmanaged, struck by political instinct and institutional habit, and that the resulting incentives favor climate deals in richer markets over development deals in poorer ones. Climate mandates are set by shareholders without analysis and loosely specified, and a third mandate, the national interest, is now emerging through informal shareholder pressure. Rapid growth in climate investment has concentrated portfolios in energy and mitigation, while emitting but development-critical industries such as cement and steel have become harder to finance. Upper-middle-income countries take a larger share of DFIs' climate finance than of their other investment, and the private capital DFIs mobilize is more skewed still, because they typically mobilize less than 0.50 per dollar invested in low-income countries against more than 1 dollar in middle-income ones. The drive to mobilize is also misallocating concessional capital and turning DFIs into merchant banks, originating assets to sell them on.

DFIs are responding with portfolio floors, carbon-intensity budgets and structural separation, largely without direction from their shareholders. For DFI boards, the paper recommends deal-level tools to weigh climate against development, transition-finance frameworks for emitting sectors, more disciplined use of concessional capital, resilience screening across the whole portfolio, and structural separation of the two mandates where shareholders are aligned. For shareholders, it recommends setting the climate-development balance through analysis, not political negotiation, floors for low-income and least developed countries, new capital behind new climate commitments, and mobilization measured for quality, not volume alone. Taken together, these would put the balance between climate and development under active management.

Digital tools now enable crowds to assemble without formal organization, but their long-term impact is understudied. Drawing on 397 events and 15 months of fieldwork, this study compares the two poles of an organizational continuum during Brazil's 2013--2016 protest cycle: leaderless online-to-offline protests, and those convoked by unions and other formal movement associations. I code the main challenger groups on organizational vitality, leadership trajectories, electoral articulation, and policy durability (OLEP) five and ten years past the peak, a longitudinal complement to WUNC, Tilly's contemporaneous measures of protest strength. The two poles scored similarly on worthiness and numbers, differed on commitment and unity, and digital origins shaped but did not determine outcomes. Across left- and right-wing formations, protest groups that invested in membership structures converted mobilization into political influence. The labor federations atrophied, and many once-prominent groups dissolved. I explain these dissolutions as a consequence of means--ends decoupling---activity that proceeds without producing the outcomes it is meant to serve. Digital decoupling specifies the mechanism under technological mediation: action multiplies, while the capacities that produce intended outcomes do not. Mistaking protest activity for the movements it once indexed helps explain a decade of mass mobilization that produced little durable change.

Modern political campaigns decide who to contact using turnout-propensity scores. Practitioners have long argued that these scores “redline” low-scored communities, which campaigns rarely attempt to reach. I examine this claim using person-level records of Democratic campaigns and community-based organizations’ 414 million contact attempts (aimed at 14 million voters in ten battleground states in 2024), matched to validated turnout. Three of every five attempts—and two of every three successful conversations—went to voters already scored likely to vote. Nearly a third of the registered voters in this universe (3.8 million) did not cast a ballot, and 1.43 million of them were never reached by any program. These patterns suggest a self-fulfilling prophecy: low-scored voters go uncontacted, and the abstention that follows becomes the record from which their next score is estimated. Local membership-based community organizations offer one promising response to this dynamic. In 2024, these groups’ contact attempts succeeded at roughly twice Democratic campaigns’ rate among Black, Hispanic, Native American, young, new-registrant, and low-income voters. Each additional conversation is associated with a turnout gain, and the association is larger among voters scored less likely to vote.

Protests constitute a primary way that members of the public express their preferences. Focusing on policy protests in the United States, we use a survey experiment to study how the race of the protesters affects public reactions. We find that relative to white nonviolent protests, Black non-violent protests are perceived as more violent, but that does not dampen individuals’ likelihood of joining the protests. Moreover, Black nonviolent protests, but not white nonviolent protests, significantly move opinion in the direction advocated by the protest. All of the race effects, though, cease to exist when protests are violent. We replicate the results with samples of citizens with connections to civic organizations and elected officials. The findings suggest that the effect of race on protest reactions is conditional, depending on the protest’s goals (e.g., alter perceptions, impact policy) and tactics (nonviolent, violent).

Mass protest has reached record levels, yet the movements producing it have rarely been weaker. Why does so much participation leave so little behind? Most explanations center on recruitment, while retention has received far less attention. Drawing on a year of participation records from 20 organizations across 11 states, paired with prospective interviews and fieldwork, I observe 91,648 attendances by 38,740 people who took civic action beyond voting. I find that 71 percent of these participants act once and do not return within the observation window. Among those whose first participation was a mass mobilization, attrition approaches 90 percent. Across all groups, however, similarly sized events resulted in very different retention rates. I call the capacity behind these differences catchment and trace it to organizational practice. A first event where most of those present are established members raises the probability of return within 90 days by 12.8 percentage points, net of event size and robust to organization-clustered inference. The pattern holds across constituencies and regions, and in 13 years of records from one organization. The strongest organizational correlate of catchment is the share of members embedded in teams.

Many states possess substantial capacity but experience gaps between potential and performance. This paper argues that the gaps reflect differences in state capability, defined as the empowered ability of state agents to transform capacity into performance. Drawing on nearly two decades of experience with Problem Driven Iterative Adaptation (PDIA), the paper examines how state capabilities develop and what role outsiders can play in supporting their development. PDIA combines action learning, through which insiders build capability by working on real problems, and action research, through which outsiders learn about capability by accompanying these processes. Thus, capabilities emerge through repeated cycles of action, reflection, experimentation, and adaptation. The paper also uses episodes from the television series Ted Lasso as a metaphor for this process, illustrating how a coach from the outside contributes not by providing solutions but by creating conditions in which individuals and teams learn, adapt, and discover their own capabilities in and across multiple moments. The findings challenge conventional approaches to external assistance. Capability is not built through products, plans, or technical advice alone, but through the relationships, experiences, and learning processes that enable state agents to transform capacity into performance.

Policy success is widely recognized as multidimensional, yet relatively little attention has been paid to a fundamental question: whose definition of success matters? This paper argues that different policy actors prioritize different dimensions of success because they occupy different positions in the policy process. Drawing on applied literature, the paper identifies four generic actor groups—political authorizers, designers, implementers, and civic advocates—and develops expectations about the success criteria each is likely to prioritize. These expectations are explored through a survey of 121 experienced practitioners, each drawn from one of the four actor groups and asked how frequently they consider thirty policy success criteria spanning six dimensions of success. The results show clear and consistent differences. Designers concentrate primarily on program design and goal achievement. Political authorizers balance programmatic, political, and impact concerns. Implementers exhibit the broadest conception of success, placing particular emphasis on processes, stakeholders, capability, and enduring impact. Civic advocates prioritize fairness, participation, trust, and beneficiary outcomes. The findings suggest that policy disagreements arise not because actors disagree about whether success matters, but because they evaluate success from different vantage points.