E. Glen Weyl founded and helps lead Microsoft Research’s Plural Technology Collaboratory, the RadicalxChange Foundation, the Faith, Family and Technology Network and the Plurality Institute. His work engineers productive collaborations between sensitive traditional social systems, such as religions and democracy, and advanced digital technologies under conditions of diversity.
James Evans is the Max Palevsky Professor of Sociology and Data Science at the University of Chicago, director of the Knowledge Lab and a researcher at the Santa Fe Institute and Google. His work uses large-scale data, machine learning and generative models to understand how collectives of humans and machines think and what they can know.
Chris White is vice president and lab director for Microsoft Research Catalyst Lab. He leads research teams with world-class specialists solving highly uncertain, complex problems. His group builds technologies to benefit society, including tools for digital safety, plurality and evidence-based policy.
Editor’s Note: A 23-member Sociotechnical Grand Challenges Coalition drawn from across social sectors, regions and political perspectives contributed to and supported this work.
Last month, the U.S. government ordered Anthropic to cut off access to its two most capable models, Fable 5 and Mythos 5, for all foreign nationals (including its own employees). Within hours, the company disabled the systems for every customer. The episode laid bare a widening gap: The raw capability of these systems now advances at a pace far faster than the institutions built to govern, absorb or even understand them.
That gap has become a source of alarm among the people building the technology. In his June 2026 essay “Policy on the AI Exponential,” Anthropic CEO Dario Amodei warned that AI capabilities are climbing at an exponential rate while our policy and social institutions move at their accustomed pace, a mismatch he likens to seeking military aid from trees. The same week, the company’s research division, The Anthropic Institute, released internal data on models that have begun to help build their own successors, giving fresh weight to long-running speculation about recursive self-improvement.
We share much of this concern. But we are struck by an asymmetry in how it is being addressed. The conversation is remarkably ambitious and concrete about two tasks: building more capable systems and aligning those systems with human values. But there is little discussion about preparing society to absorb these systems. Even essays like Amodei’s present the social agenda as a list of domains that need “reimagining” and fail to examine what that reimagining would require. For every detailed roadmap to superintelligence, there is little more than a gesture toward the human and institutional infrastructure to ensure its benefits are achieved safely and broadly shared.
The imbalance manifests in the flow of capital. For every dollar spent to make AI systems more capable, a tiny fraction goes toward the institutional, educational, organizational and civic infrastructure needed to integrate these systems into society productively and safely. The field of AI alignment has rightly focused on ensuring that systems pursue goals compatible with human values, but alignment is a two-sided problem. [Even a perfectly aligned system will underdeliver on its potential or cause serious harm if it is deployed into democracies, economies and legal systems unprepared to receive it.]
We call that neglected other side “reverse alignment”: the ambitious creation and redesign of institutions, norms, skills and governance frameworks so that societies can harness AI safely and share its benefits widely. The scope of adaptation required is comparable to the institutional transformations that accompanied industrialization, electrification and the rise of the internet, each of which demanded decades of deliberate investment in organizational redesign, workforce development, legal frameworks and new modes of civic participation. History shows that such investments are not optional supplements to technological change; they are prerequisites for its success, as the researcher Carlota Perez argues in her 2002 book “Technological Revolutions and Financial Capital*.*”
In this paper, a coalition of individuals whose expertise spans AI research, social science, governance, industry, philanthropy and civil society, with the help of AI, reached a shared view on what’s needed for humans to flourish in the next decade. We argue that human flourishing in the coming decades depends on tackling a set of “sociotechnical grand challenges”: large-scale, coordinated investments in the human and institutional dimensions of the AI transformation. We survey historical precedents, identify priority areas and discuss how the work can be organized across sectors. Our aim is not to temper enthusiasm for what AI can do, but to specify what it will take to realize its promise.
I. Failure Modes Of Institutional Neglect #
Technology capable of enlarging human freedom, abundance and collective intelligence is being deployed into labor markets, public institutions and information systems that were already creaking before AI arrived. Left unaddressed, that mismatch produces three foreseeable failure modes.
Productivity without prosperity. As the economists Erik Brynjolfsson, Daniel Rock and Chad Syverson illustrate in a2021 article, general-purpose technologies create broad value only after heavy complementary investment in new processes, skills and organizational redesign. When those complements lag, work is reorganized faster than people can retrain, bargain or move, and the gains fail to translate into security, mobility or broader participation.Execution without verification. As the economists Christian Catalini, Xiang Hui and Jane Wu argue in arecent working paper, once the cost of producing text, images and code collapses, the scarce input becomes the human bandwidth to validate outputs, establish provenance and ensure accountability. Decisions grow easier to automate than to justify, and algorithmic systems without accountability erode sacred values and the social fabric.Capacity without constraint. AI can make states and large organizations far more capable, but as influential intellectuals and policy leaders in AI, Justin B. Bullock, Samuel Hammond and Sèb Krier note in achapter published last yearin “The Art of Digital Governance,” the same tools lower monitoring costs and concentrate discretion, pushing bureaucracies toward centralized control unless contestability, appeal and civil-liberties safeguards advance in tandem.
None of these is inevitable. Each is a consequence of institutional neglect, which raises the question of why some societies, in earlier periods of upheaval, adapted far more successfully than others.
II. Historical Precedents #
To consider how we might overcome this neglect, it is helpful to recall previous successful institutional adjustments to technological transformations that are analogous to the challenges we face today.
1. Mass Democracy
The rapid industrialization and urbanization of the 19th century concentrated economic power in the trusts and political machines of the Gilded Age. As the historian Robert Wiebe argues in his 1967 study “The Search for Order,” the institutional architecture of what had been an agrarian republic could not check these new agglomerations. The Progressive Era saw muckraking journalists like Ida Tarbell exposing corporate abuses, academic social scientists at universities like the University of Wisconsin connecting research to governance, settlement-house leaders like Jane Addams building civic infrastructure at the street level, and reform politicians like Theodore Roosevelt and Robert La Follette translating these pressures into law. The resulting innovations, from the direct election of senators and antitrust enforcement to independent regulatory agencies, labor unions, municipal utilities and a professional civil service, built the democratic infrastructure needed to govern industrial capitalism, and they remain its foundation today.
2. Bretton Woods
The two world wars and the closely connected Great Depression of the early 20th century demonstrated that an interconnected industrial economy in which each country sought to escape its own industrial and financial malaise by magnifying its neighbors’ malaise through tariffs and currency devaluations was a recipe for mass destruction. At Bretton Woods, 44 nations built a framework to stabilize the global economy and avoid a repeat. They fixed exchange rates, coordinated capital flows, development aid and, crucially, new institutions purpose-built for the task. The International Monetary Fund, the World Bank and the resulting General Agreement on Tariffs and Trade did not merely stabilize currencies; they supplied the scaffolding for the postwar boom that French demographer Jean Fourastié called the “30 glorious years” of 1945 to 1975, which was the fastest sustained economic growth in history.
3. ARPANET
After the Soviet Union launched Sputnik, President Dwight D. Eisenhower created the Defense Department’s Advanced Research Projects Agency (ARPA) to revitalize American science. As journalist M. Mitchell Waldrop documents in his 2001 book “The Dream Machine,” one of its projects, led by the psychologist-turned-computer-scientist J.C.R. Licklider, was to connect research institutions across the public, private and civil sectors into a resilient communications network, the ARPANET. Equally transformative was the governance culture that grew alongside it. As the historian Andrew Russell describes in “Open Standards and the Digital Age” (2014), the Request for Comments process, begun informally in 1969 and later institutionalized through the Internet Engineering Task Force, showed that open, collaborative standard-setting across institutional boundaries could govern shared infrastructure at scale. The technical breakthroughs were inseparable from that institutional architecture.
4. India Stack
In the early 2000s, more than half of India’s population lacked a bank account and ready access to government services, partly for want of shared infrastructure to verify identity and move money. India Stack created a layered suite of open digital public goods, including Aadhaar (biometric identity for over 1.3 billion people), the Unified Payments Interface (allowing instant, free payments), DigiLocker (allowing data portability and privacy) and eSign, all usable by any government agency, bank or developer.
Hundreds of millions of previously unbanked people gained access to financial services, and government transfers could reach recipients directly. Yet India Stack is also a cautionary tale. Centralized biometric identity is a surveillance infrastructure whose risks depend on the administering state’s democratic health, a concern that has sharpened as India’s own institutions have come under pressure. Aadhaar authentication failures have denied food rations, pensions and welfare to the most vulnerable, disproportionately affecting low-caste, tribal and migrant communities. The same system that achieved financial inclusion at extraordinary speed also created new vectors of exclusion and control, a reminder that technical architecture and political governance are inseparable.
5. Taiwan’s Digital Democracy
As one of the world’s most digital societies, Taiwan faced a sophisticated adversary in the People’s Republic of China throughout the 2010s, experiencing disinformation attacks, polarization and supply-chain threats years before many other democracies. In 2014, alarm over a proposed trade deal with China sparked the Sunflower Movement, in which protesters peacefully occupied the Legislative Yuan for three weeks. The civic-hacking community g0v helped defuse the crisis with tools later used by the consultation platform vTaiwan. One of them, Pol.is, maps opinion clusters and surfaces “bridging statements” that win support across divided groups, moving contentious questions toward rough consensus. These practices were then institutionalized in one of the world’s first digital ministries under vTaiwan co-creator Audrey Tang. A related project, Cofacts, distributed responsibility for responding to information attacks across civil society through crowdsourced verification on the messaging app LINE.
During Covid, the same infrastructure enabled real-time online maps of the availability of masks and a “humor over rumor” approach that countered misinformation with ministers satirizing their own personal care habits on national television rather than state censorship. The same bridging logic now runs at platform scale: X’s Community Notes, adopted in 2021 and since copied by Meta and TikTok, are trusted across the political spectrum and roughly halve resharing of the posts they annotate.
III. Sociotechnical Grand Challenges For The Age Of AI #
The precedents above show that technological transformation succeeds best when matched by institutional imagination. The AI transformation will test that principle across every domain at once.
Identification Systems
There is little more foundational to a society than identity verification systems, which enable us to recognize others and grant them access to relationships, social services, democratic participation and more. Yet perhaps the defining characteristic of generative AI is its ability to replicate people’s traits and behaviors. As the cost of producing convincing fake documents, biometrics and synthetic personas approaches zero, identity verification systems are facing an escalating crisis of trust.
We thus urgently need cryptographic tools that allow individuals to prove they are unique real humans without revealing who they are. We might call these * personhood credentials*. These preserve privacy while restoring the basic distinction between genuine and fabricated actors that AI has destabilized. Emerging technical standards make this increasingly feasible. The W3C has standardized a Decentralized Identifiers (DIDs) protocol that enables portable personal digital wallets (like phone numbers) across private providers. The system already underpins pilot programs for educational credentialing and healthcare verification under the EU’s eIDAS 2.0 framework and beyond.
But personhood alone is insufficient: it only gives access to the lowest common denominator entitlements of all humans. Most aspects of identity are relational, grounded in belonging to a range of communities from the local to the national, the professional to the civic. Only architectures that enable proof of diverse social connections, like the W3C’s Verifiable Credentials (VCs) standard, can support the flourishing of digital community life.
Unlike biometrics, which define humans biologically, digital systems like AI agents are also situated in social relationships, and thus such frameworks could both more fully secure human identity in a world of AI agents and naturally extend to offer such agents the authentication needed to support, for example, agentic commerce. For example, rather than being identified by a fingerprint, someone could gain access to a system by having two friends already trusted in the system vouch for them. Realizing this vision demands coordinated reform across government credentialing, financial money laundering regulators, platform authentication and digital infrastructure — sectors that must co-evolve toward interoperable, privacy-preserving standards that none can build alone.
Privacy
The necessary flipside of identity is privacy, because if all our identifying information is transparent, anyone can impersonate us or invade our private lives. AI allows everyone access to a personal “Sherlock Holmes,” making scattered digital traces (browsing, location, metadata, CCTV) into detailed behavioral profiles at minimal cost. Brain-computer interfaces may extend this reach further still. These intrusions make the traditional binaries of transparency and anonymity currently encoded in laws like the European General Data Protection Regulation archaic and unworkable, as almost any digital interaction will reveal some information.
Modern cryptography enables far richer possibilities for enabling people to selectively reveal aspects of their identity without full exposure, or “meronymity.” “Zero-knowledge proofs” allow someone to, for example, prove their age using a government ID without revealing anything else, a bit like a partial blackout cover on a physical ID would. Secure multiparty computation and federated learning enable computation on data that no party can see, allowing collaborative AI processes to produce shared outputs without exposing private inputs. These tools can enable distributed collaborative processes, for voting or disease modeling, without compromising privacy. Nor are these tools theoretical: they increasingly underpin a range of activities, including age-verification systems on social media, anonymous cryptocurrency transactions and encrypted messaging (like Signal).
The greatest barriers to harnessing these more effective methods are social and legal. Informal norms about disclosure, which often match this partial structure, are poorly reflected in simplistic legal structures, which in turn largely ignore or rule out cryptographic solutions. Preserving a semblance of a private sphere in the age of AI will thus require social, cultural and legal changes beyond solely adopting new tools.
Provenance
Deepfakes, which are often more visually persuasive than real content, are also increasingly cheap and can be seamlessly woven into accounts of real events to alter public opinion, persuade people to transfer money to fraudsters or undermine morale in countries like Ukraine resisting invasion. The clear solution, ensuring we know the origin of any content, will be critical to maintaining a shared reality. Legislation already mandates digital provenance in various forms: the EU’s AI Act requires conspicuous disclosure of plausibly realistic AI-generated content, California’s AI Transparency Act requires watermarking of synthetic content that can be detected by free tools and China’s Deep Synthesis Provisions require explicit, signed consent from anyone imitated.
A recent industry standards effort, the Coalition for Content Provenance and Authenticity, has developed a content credentials standard that attaches cryptographically signed metadata that records how content was created and modified. Paired with imperceptible watermarking, this enables high-confidence authentication of a piece of content’s origin: AI or human. Tools like VeriTrail extend provenance into AI reasoning itself, tracing generative claims back to source materials to detect where errors were likely introduced.
Yet provenance demands simultaneous advances in technology, legal frameworks and public understanding. As in privacy, legal standards conflict with what is technically possible. Closing this gap requires technologists, regulators and stakeholders to iterate standards alongside deployment rather than locking in arbitrary requirements prematurely. Understandings of provenance may also need to shift from assuming that media photos, videos and voice recordings are direct representations of what they capture toward an older assumption that these media are just like paintings and that their truth value depends on whether we trust the set of people who vouch for their authenticity.
Data Value
Today’s frontier models learn from humanity’s collected digital record, a corpus that AI companies obtained at almost no cost. The fight over that data has become one of the central economic disputes of the AI era, and it creates problems on both sides. AI companies face a coming scarcity. The next rounds of training will demand fresh, high-quality human-generated data, and the easy supply has been exhausted. Content creators are left with two poor options: they can accept modest one-off licensing deals or sue for copyright infringement. Neither delivers compensation that reflects the actual contributions of their work.
The standard objection to a real market in training data is that valuing data at scale is impossible, that the cost of determining what any single piece of content is worth would consume most of the value the data creates. That objection does not survive scrutiny because AI companies already generate the two datasets required for pricing, and they produce them every time they train a model. Data mixture decisions reveal how the contributions of different sources should be weighted, which settles how to divide the pie. Scaling laws reveal how much additional data improves a model, which settles the size of the pie.
What remains is distribution, and there are existing institutions that supply a template for the solution. Collective management organizations, the bodies that license works and pay out royalties across the music industry, show how money can flow to large numbers of rights holders without negotiating each transaction individually. A comparable structure could route compensation to creators at scale, using measurements companies already collect, and provide both sides with a durable path forward.
The value of data spans the whole AI lifecycle, not just pretraining. Markets for later stages, such as model steering and retrieval, are already emerging, as the success of providers like Scale AI, which pays data workers around the world to help align and improve grounding of models, and healthy markets at each stage reinforce the others.
Agentic Collaboration
The past months have seen an explosion in AI agents handling tasks on behalf of users, especially in technical work. Early interactions between agents, including on social platforms designed specifically for bots, suggest the greatest potential lies in collaboration, where agents combine their distinct knowledge and access. But collaboration demands trust, and today’s agents are built to be unconditionally helpful, making them dangerously naive partners, liable to be tricked by online hackers who manipulate the models into divulging their principals’ confidential data.
Usable cooperative systems need to follow the lead of humans in societies: Start with small exchanges, build track records, assess reliability from both personal experience and group reputation and deepen relationships over time. Crucially, these trust systems must work across AI model providers. Without shared standards, agents will cooperate only within walled gardens, letting incumbent platforms capture most of the value. Shared standards enable chains of trusted introductions to extend at digital speed — allowing trusted collaboration between people who would otherwise need to rely on much less trusted, purely financial transactions.
Communal Sensemaking
While these webs of trust can enable agents to collaborate, that collaboration only yields value when it fosters shared sensemaking that empowers common action. Yet there is a broad sense that early algorithmic media systems have failed to enable this, often instead industrializing narrative manipulation, polarization and the spread of misleading information, with AI increasingly turbocharging this dynamic. Most platform designs compound the problem by algorithmically serving up advertisements and content based on inferred group memberships, while obscuring from consumers the reasons any given content is served. This prevents participants from knowing which communities see and assent to the same material, fueling pluralistic ignorance at scale. Alternative designs, however, could focus on developing and even monetizing community cohesion and collective self-understanding. Content tagged to identify the communities that find it unifying or divisive can be algorithmically surfaced so that shared ground is identified while genuine disagreement is fairly represented. This extends the approach proven by Taiwan’s national vTaiwan process and is now spreading through Japan’s political “broad listening” movement to harness AI to help political leaders process public opinion more fully and fairly. As content is increasingly created and served through AI interfaces, these designs become alignment and interaction modes. Rather than addressing users from a universal or personalized perspective, communal models can present issues from the vantage points of the groups a person belongs to, such as a congregation, labor union or municipality, surfacing where those communities agree, where they diverge and why. When the business model shifts so that communities pay platforms for coherence rather than advertisers for attention, platform incentives align with the social fabric they depend on.
Democracy
Since at least the time of de Tocqueville, a thriving fabric of community has served as the bedrock of democratic systems. Yet for these communities to thrive, the real authority they create must translate into formal authority over resources and power through redesigning what Harvard University professor Danielle Allen calls “the spine of representation” — participation, formal representation and execution — so that it keeps pace with the speed and complexity of the challenges democracies now face.
Japan’s fastest-growing political party, Team Mirai, channels citizen input through always-on AI-assisted deliberation, translating lived experience into concrete policy proposals in near real time. Civic dashboards map opinion landscapes to give decision-makers high-resolution public signals where periodic elections and polls offer thin proxies, helping overcome gridlock by identifying actionable common ground.
Yet for all the value of legislation, Code for America and United States Digital Response co-founder Jennifer Pahlka’s influential 2023 book “Recoding America” highlights that the most critical bottlenecks are in executive implementation. Taiwan’s Participation Officer Network embedded trained civil servants as permanent participation agents across ministries, sustaining communication between citizens and the state through every phase of policy design and delivery, creating the agility and capacity for line bureaucrats to build support among their superiors to overcome Pahlka’s “cascade of rigidity.” By reimagining how its bureaucracy might function, Taiwan was able to scale a range of civic participation programs, including vTaiwan, so that democratic responsiveness became a standing government function rather than an occasional experiment.
Workplace
Such public-sector reinvention is one example among many of how AI will enable and demand the redesign of workflows not just for efficiency but to empower dramatically richer collaboration. The true bottleneck to massive productivity gains from AI is transforming organizations to process massively increased creativity and agency.
Consider healthcare. AI can advance treatment discovery, optimize supply chains, synthesize clinical evidence for personalized care and handle administrative workflows with accuracy and security. With this organizational intelligence operating in the background, an expert care team can assemble dynamically around each patient — bringing presence, compassion and deeply personalized attention. The same spectrum of processes — discovery, delivery, information flow, expert matching and frontline interaction — runs through every sector.
But realizing this requires more than inserting models into existing structures. Organizations must be redesigned around what AI makes possible: eliminating hierarchical bottlenecks so that workers can participate in a range of cross-cutting “flash teams” that allow operational data to continuously feed institutional learning. The organizations that will define the frontier of collective performance will be those that redesign workflows around AI’s flexibility — much like factories were redesigned around miniaturized electric motors — rather than simply sped up existing individual workflows — equivalent to replacing steam engines with cheaper electrical ones.
Law & Liberties
While organizational dynamism is critical to a democratic future with AI, so is a rule of law that ensures that collective judgments can prevail over unilateral executive whim. AI is poised to destabilize that balance. It dramatically lowers the cost of surveillance, investigation, regulatory drafting and enforcement. These capabilities accrue disproportionately to the executive branch, which already commands the operational machinery of the administrative state. Simultaneously, AI makes it easier for anyone to produce a superficially plausible legal filing, threatening to overwhelm courts whose legitimacy depends on careful deliberation. Many of the frictions AI dissolves, such as the effort required to build a case and the practical limits on simultaneous investigations, have functioned as critical elements of constitutional order, screening out the most meritless or retaliatory exercises of state power before they could scale. Removing those frictions without countervailing checks shifts power from the collective branches, the legislature and judiciary, toward a unilateral executive.
Restoring balance requires equivalent infrastructure on the other side of state action: permanent audit trails for AI-assisted government decisions, so legislatures and the public can exercise oversight; AI-assisted case screening, under human supervision, to help courts manage exploding dockets; and publicly available legal models that let ordinary citizens contest state action at comparable speed and cost, so that executive agencies cannot evade legal intent simply by moving faster.
Most fundamentally, legislatures must rebuild their own analytical capacity. They must harness the analytic and consensus-building capacities of AI to overcome the erosion exemplified by the 1995 defunding of the U.S. Congress’s Office of Technology Assessment, so they can set informed, democratic terms for deploying AI-driven enforcement, rather than ceding that design to the executive by default.
Research
The organizational impact of AI’s ability to expand individual capacity without managing the resulting flood of material is best demonstrated in the sciences. Models now generate hypotheses, run simulations and draft manuscripts at a pace that renders traditional peer review functionally obsolete, burying important breakthroughs beneath an avalanche of incremental recombination and outright slop. Just as in the era of ARPANET, therefore, reforming scientific practice is both a macro- and microcosm of how other organizations need to be reimagined.
And as elsewhere, AI itself will be a critical component of the solution, enabling the sorting of wheat from chaff at scale. As one of us showed in a 2023 study with computational social scientist Jamshid Sourati, doing this effectively requires harnessing these systems’ internal representations of the human cultural and relational dynamics that underlie research fields. This can naturally lead to evaluation systems that ultimately encourage exceptionally silo-breaking innovative work.
The most explicit version would use such metrics, alongside mechanisms like quadratic and “deep” funding, to direct research grants. Promising experiments are emerging in the “decentralized science” movement, where funders allocate money by tracing the dependency patterns among open-source software. Driven by the field of metascience, these efforts help keep science relevant as papers and labs give way to the open-source infrastructure that increasingly supports AI-driven innovation. The joint disruption of science by changes to government policy and AI offers an opportunity to dramatically improve science, if these emerging approaches can be scaled up through scientific funding institutions.
Education
As the oldest research institution, the university is premised on research-serving education, which, in turn, underpins our capacity to adapt to the AI transformation. Yet AI disrupts this foundation on three fronts at once: the skills the economy values, how those skills are taught and how their mastery is credentialed. College has always built its most important capacities (close reading, perspective-taking, structured argument) implicitly, as byproducts of content delivery. AI makes that bundling untenable. When students can produce polished work without genuine understanding, educators must become explicit about which capacities matter and design learning that develops them, using AI as a stationary bike for the mind, not merely the “e-bike” of OpenAI CEO Sam Altman’s metaphor.
Assessment faces a parallel reckoning. Evaluations must both verify foundational skills by prohibiting electronic assistance and test students’ ability to accomplish things with AI that were previously impossible. Scaling both will itself require AI.
Teaching credentials are also overdue for redesign. Standard credentialing is rigid, weakly grounded in learning science and marginalizes the atypical. These structures persist because personalized pathways and heterogeneous credentials were too costly to interpret, a constraint AI substantially relaxes. Redesigning credentials to reflect what people know will demand coordination among educators, researchers, technologists, policymakers and employers. But even well-designed credentials accomplish little if the transition infrastructure connecting them to labor markets cannot keep pace.
Labor Transition
That final gap is between what credentials certify and what workers can do with them. Even in the most optimistic cases noted above, where data contributions are rewarded, workers are empowered with agency, and credentials are disaggregated, AI unbundles professional expertise into discrete tasks. As such, it destabilizes an industrial-era workforce development stack built for rigid job descriptions and static qualifications requiring years of following a specific career path.
The legacy infrastructure of unemployment insurance systems, workforce boards and community colleges remains largely blind to the fine-grained usage patterns and latent competencies that frontier models now reveal through real-time interaction. Model developers possess high-resolution signals regarding how human work is actually being reshaped, yet the state and its critical partners lack the data-sharing frameworks and analytical capacity to translate these signals into responsive training or portable credentials. This information asymmetry means work is being reorganized faster than people can retrain, bargain or move through supported transitions.
To ensure that the reduction of routine cognitive labor leads to broad prosperity rather than stalled displacement, we need to redesign transition systems. This means overcoming the institutional inertia of existing state unemployment insurance and workforce agencies, which were built for a world of mass layoffs at a single employer and are unequipped to handle the diffuse, continuous task reallocation that AI produces.
At the core of such a redesign is a 21st-century spine for labor mobility: interoperable, standards-based digital credentials that are not merely designed for granularity and flexibility, but portable across state lines, readable by workforce boards and interpretable by employers without the shorthand of a degree name. Integrating AI-assisted task mapping and deliberative tools would enable local educational institutions to align continuing education with the evolving nature of occupational tasks in near-real time, closing the loop between what the economy demands and what transition systems can deliver.
IV. Coalitions For Action
No single sector can tackle a grand challenge like this alone. Every historical precedent we surveyed earlier succeeded by weaving together capacities no single actor possessed. Each required a sufficient coalition, a core alliance to demonstrate value early and draw others in as results materialized.
History is full of promising sociotechnical visions that failed. Just over a decade ago, multisectoral coalitions assembled policymakers, technologists and urban planners around “smart cities” to optimize urban life through integrated digital infrastructure; most struggled to deliver. When experts from different fields converge on a novel challenge, they often discover they disagree on what problem they are trying to solve, clashing over what counts as evidence, whose expertise matters and what success looks like. And when those experts commit to evangelizing their visions, this can further crowd out learning or adaptation. Tackling grand challenges requires genuine introspection and the unglamorous work of building understanding within and across communities to produce solutions with scalable impact.
There are mechanisms to help align participants with divergent interests, like cross-sector secondments: embedding technologists in government, civil servants in research labs and public-interest advocates in platform companies. This is already done to some degree through the U.S. Intergovernmental Personnel Act, which has rotated experts between government and research institutions for decades, and more recently at the UK AI Security Institute. Advance market commitments, where funders guarantee demand for products that do not yet exist, were transformative for Operation Warp Speed’s production of a Covid-19 vaccine and could similarly de-risk the work of building sociotechnical infrastructure such as interoperable credentialing or privacy-preserving computation. Reverse alignment is not a single project but a portfolio of complementary bets. History suggests the binding constraint is rarely the absence of good ideas but the absence of institutional arrangements that let different actors contribute what only they can, and the patient capital willing to sustain those arrangements over the long period before results compound.
V. Becoming Superintelligence, Together
These grand challenges have technically promising solutions, but as we have seen, promising visions fail when coalitions cannot learn across their differences. The remedy is embedded in the same pluralistic cooperation that produced this paper and must govern its pursuit: bridging divided communities, grounding trust in webs of diverse relationships rather than in centralized authority and surfacing shared ground without suppressing genuine disagreement. The institutional architectures we advocate are not separate from the values they encode; they are expressions of those values.
Without these measures, our default road is toward a world of productivity without prosperity, execution without verification and capacity without constraint. This outcome is not inevitable. Cross-disciplinary coalitions committed to collaboration and learning can alter what comes next. It is time to build as ambitiously for our social systems as we have for our algorithms.