Our task is to build a scientific enterprise that maximizes the marginal returns to intelligence, which means confronting, clearly, the factors that constrain its work. Three stand out. The first is the speed of feedback. Particle physicists have devised myriad theories about our universe, but we currently lack the experimental data to distinguish among them. No amount of intelligence can conjure observations that do not yet exist.
The second is the speed of atoms. Cells divide on their own schedule, and hardware must be physically manufactured. No cognitive power can alone make a rocket or a semiconductor fab build itself overnight.
The third is the constraint imposed by our own institutions. Publishing systems, funding structures, regulatory frameworks, clinical trial requirements, and cultural inertia govern how research happens. This will be felt first within the scientific enterprise. The coming abundance of cognitive capability demands a corresponding transformation of our scientific institutions. New technologies like AI should give us tremendous optimism, but we should harbor no illusions that AI will repair our scientific enterprise by default. Even if every scientist benefits from adopting AI, it does not follow that science as a whole will advance. This is one of the counterintuitive properties of complex systems: individual gains do not automatically aggregate into collective progress. For a century, the United States aggressively suppressed forest fires, and each intervention seemed like an obvious success. But by preventing small fires, we allowed the fuel of dead wood, dense brush, and dry undergrowth to accumulate until the inevitable fires became infernos even harder to contain. AI could do the same to science, by making each researcher seemingly more productive while allowing the conditions for catastrophic and systemic dysfunction to build.157 AI tools will make it trivial to generate more papers, more grant applications, and more submissions to peer review. But if the obstacle to scientific progress were simply the production of these artifacts, we would already be living in a scientific golden age. We are not.
As Chapter II documented, the exponential growth in publications, researchers, and funding over the past half-century has not produced a corresponding acceleration in discovery. Disruptive work represents an ever-shrinking fraction of total output, and the breakthroughs that reorient fields arrive no faster than they did generations ago.
The rise of AI in science therefore demands that we update the institutional machinery governing what gets funded, what gets published, what gets rewarded, and what gets corrected. Science works only if we can generate and verify knowledge in tandem. Models that make flawed paradigms easier to extend may entrench scientific biases; AI tools that flood peer review with unthoughtful submissions will overwhelm systems already stretched thin. In an age of abundant intelligence, everything discussed in the preceding chapters, such as the public-private partnerships that direct AI toward problems that matter, the metascience reforms that restore accountability and reward genuine exploration, the novel funding mechanisms that tolerate early failure, the rebuilding of our technical workforce and manufacturing base, becomes more urgent, not less.
BUILDING THE INFRASTRUCTURE
The reforms proposed in the preceding chapters are also essential for AIpowered science. AI has shifted basic research toward industry, made cross-disciplinary collaboration essential, and sharply increased the capital intensity of frontier science.
In doing so, AI has made the institutional adaptations described in the preceding chapters necessary to update the scientific enterprise for the modern world. The novel organizations and funding mechanisms in Chapter II matter because fully leveraging AI for science demands tight iteration between exploration and engineering, something traditional academic departments were not built to sustain. It also requires close partnerships between domain scientists who understand AI models and AI researchers who understand the science. Private research institutes are now housing machine learning researchers and life scientists in shared facilities to maximize collisions,158 while fellowship programs pair AI researchers with academic co-advisors.159 These reforms must spread across the entire scientific enterprise. Eventually, we will need AI-native scientific institutions, with rules, infrastructure, capital models, and cultural norms built around the use of powerful AI systems.
The public-private partnerships in Chapter III matter because frontier AI capabilities concentrate in private laboratories. The compute clusters required to train frontier models cost billions of dollars. The engineering teams capable of operating them mostly work at a handful of companies, on payrolls no university can currently match. Keeping private sector capabilities in conversation with the public scientific enterprise will require carefully crafted structures. Existing partnerships between federal laboratories and AI companies represent steps toward that future, with researchers applying cutting-edge AI systems to advance fusion energy, drive advancements in computational biology, and even control a rover on Mars.160
Finally, the reconnection of science and craft in Chapter IV matters because as intelligence becomes more abundant, the constraints on scientific progress shift from generating ideas to realizing them in the physical world. The world of atoms, which includes manufacturing, fabrication, and the infrastructure on which discovery depends, will increasingly determine the pace of progress. No amount of intelligence substitutes for the instruments needed to collect experimental data or the facilities needed to build and test new technologies. Reforms to rebuild apprenticeships, capture tacit knowledge, and reconnect universities with regional manufacturing ecosystems are focused precisely on this constraint.
THE GENESIS MISSION
Throughout our history, from the Manhattan Project to the Apollo Program, America’s greatest scientific advances have come when national capabilities were matched with the right institutional design. The Genesis Mission is the next chapter in that tradition, a national effort to harness AI for scientific discovery at a scale no other nation can match.
President Trump launched the Genesis Mission in November 2025 as America’s premier –“AI for science” initiative. The Executive Order establishing the Mission directs DOE to build the American Science and Security Platform, which will connect our most powerful supercomputers, AI systems, and scientific instruments and datasets into a single discovery engine designed to double the productivity and impact of American science and engineering within a decade.161 The Mission draws on an unparalleled base of national capability. DOE’s 17 national laboratories constitute the largest concentration of scientific infrastructure in the world, employing roughly 40,000 scientists, engineers, and technical staff, and receiving approximately $20 billion in annual funding.162 These institutions house our most advanced particle accelerators, synchrotron light sources, supercomputers, and experimental facilities spanning materials science, fusion energy, and nuclear security. Beyond our national laboratories, agencies including the FDA, NSF, National Oceanic and Atmospheric Administration (NOAA), and Department of Veterans Affairs steward vast quantities of scientific data accumulated over decades of federal investment, from genomic sequences to weather simulations.
The Genesis Mission will unlock this capacity, including by building foundational technologies and AI-ready datasets, to tackle the nation’s most complex science and technology challenges. Realizing the Mission’s potential requires addressing four key challenges that would otherwise constrain AI-enabled science. The first is problem selection. Not every scientific problem is well-suited to AI intervention. The strongest candidates exhibit certain characteristics, such as, for today’s AI systems, large combinatorial search spaces, substantial quantities of structured data, and clear metrics against which to benchmark progress. Protein structure prediction, for example, fit these criteria precisely.
The space of possible configurations is vast, decades of crystallographic data provided training material, and benchmarks such as the Critical Assessment of Protein Structure Prediction (CASP) allowed researchers to measure improvement. The Mission has therefore directed DOE to identify at least 20 science and technology challenges of national importance, spanning advanced manufacturing, biotechnology, critical materials, nuclear fission and fusion, quantum information science, and semiconductors. Challenges will be reviewed annually to reflect both scientific progress and national priorities. In a world where AI research is flush with capital, the Federal Government’s value-add is not funding AI in the abstract, but directing it toward problems where breakthroughs could unlock entire branches of downstream discovery and application, just as cracking the human genome did decades ago.
The second is institutional capacity. The Genesis Mission is designed to operationalize the reforms outlined throughout this report, many of which are preconditions for AI-powered science, at national scale. In December 2025, DOE announced agreements with twenty-four organizations, including leading AI companies, semiconductor manufacturers, and cloud providers.163 These partnerships, and the many that follow, will ensure the Mission’s outputs flow across the entire national research ecosystem. Furthermore, the Transformational AI Models Consortium, a cornerstone investment in the Mission, will mobilize National Laboratories to partner with industry to generate new AI-ready data and support the development of foundation models that harness DOE’s unique data, facilities, and expertise across scientific and engineering domains.164 The third is data infrastructure. Scientific data is the raw material for AI-powered discovery, but much of America’s most valuable data is currently inaccessible, uncurated, or locked behind restrictive licensing. Fixing this requires two complementary approaches. One is opening access to Federal Government data. Many valuable datasets exist because the government chose to build them, like NOAA’s weather data or the Materials Project’s mapping of inorganic crystals.165 The American Science Cloud, a cornerstone of the Mission’s infrastructure, will empower the National Labs to curate and distribute DOE’s AI-ready scientific data for the broader research community and unlock data hidden behind government bureaucracy. Approach two is creating incentives for individual researchers to curate and share their own datasets. Much valuable data arises organically, when individuals closest to the research recognize that their experimental records or failed trials could be helpful to others. This data is routinely abandoned, sometimes due to a lack of stable funding for storage and curation, and other times because there is no signal on the value of the information.166 Data on laboratory procedures and challenging experiments, for instance, may prove highly valuable in light of potential lab automation, yet such data is currently scattered. The Mission will address this directly, creating new funding opportunities for dataset curation and building new incentives to partake in these curation efforts across our science agencies.
The fourth is the integration of AI capabilities with experimental infrastructure. Where materials discovery can take around 20 years from laboratory to deployment, closed-loop autonomous experimentation could collapse that timeline by an order of magnitude.167 That makes leadership in this platform technology a strategic imperative for the United States. We have already prototyped autonomous facilities across our national laboratories, such as the A-Lab at Lawrence Berkeley, which works on solid-state synthesis of inorganic materials, and the Polybot at Argonne, a modular robotics platform for materials characterization. But other countries, including Canada and China, are racing forward. The constraint on further automation runs deeper than funding. Decades of consolidation and offshoring in the scientific instruments industry have created pathologies one would expect, including expensive products, poor software, and proprietary data formats that lock researchers into vendor ecosystems. Scientists attempting to build automated workflows spend months simply getting different instruments to communicate. Scientific instruments themselves must be redesigned for automation from the ground up, with open interfaces and standardized data formats. The national laboratories deploy the largest concentration of advanced scientific instrumentation in the world, and their purchasing power can drive that redesign. The Mission has already invested in 14 projects focused on robotics, automated laboratories, and autonomous control of large-scale experiments.169 These efforts build on a parallel push at NSF to invest an initial $380 million into programmable cloud labs across academic institutions and startups, kicking off our domestic autonomous experimentation industry, just as NSFNET played an instrumental role in forming the backbone of the modern internet in the 1980s.
America’s brightest minds and industries have always answered the call when their country needed them most. The Genesis Mission is that call to this generation of scientists and engineers to advance American scientific leadership in the era of AI. Its design reflects the core convictions of this report. The central role of the Federal Government is to shape the arena rather than direct discovery, recognizing that the private sector possesses capabilities public institutions must learn to leverage rather than replicate.
GOLD STANDARD SCIENCE
The promise of AI-enabled science rests on a foundation that is, at present, potentially unsound. We are preparing to train AI systems on scientific literature, deploy them to generate hypotheses, and trust them to design experiments, but the knowledge base they will draw on is riddled with error.
As detailed in Chapter II, a majority of researchers now acknowledge that science faces a reproducibility crisis. Between one-half and two-thirds of psychology studies failed replication attempts,171 more than one third of celebrated experimental economics studies similarly failed to replicate,172 and one study found that irreproducible findings in preclinical biomedical research alone misdirect an estimated $28 billion annually.173 Every false result can mislead subsequent researchers, creating cascading failures throughout the enterprise. Increased scientific productivity will not mean anything if the underlying findings are false. In May 2025, the President signed an executive order, Restoring Gold Standard Science, to begin addressing this dysfunction.174 The order establishes principles to govern all federally funded research: reproducibility; transparency; communication of error and uncertainty; collaboration across disciplines; skepticism of assumptions; falsifiability of hypotheses; unbiased peer review; acceptance of negative results; and freedom from conflicts of interest. Agencies are directed to apply an approach based on the weight of scientific evidence, transparently evaluating each piece of information based on quality and relevance. Replication does not happen at scale, in part, because of a market failure. Strong incentives drive researchers to publish new findings, with funding and prestige both flowing from novel claims. On the other hand, verification carries weak incentives; little glory comes from confirming someone else’s work. Previous attempts at large-scale replication have failed because they required armies of specialists to verify each study by hand. Manual verification cannot scale to the millions of papers published annually, and the problem is about to grow far more acute. As AI is introduced into the scientific process, it risks compounding these problems. False findings will multiply as it becomes easier to generate plausible-sounding scientific results than to verify them. AI research offers a preview. Leading conferences have seen submission surges of 60% in a single year, overwhelming the field’s capacity to evaluate new results. Researchers are now burdened with reviewing nonsensical AI-generated submissions while rebutting low-quality AI-generated reviews of their own work.175 Other fields will follow the same trajectory. The Genesis Mission is building a science generator with instruments capable of producing scientific discovery at an unprecedented scale. To sustain progress, we must also build its necessary counterpart: a verifier equal in rigor and scale. This is the central challenge that must be undertaken to address the reproducibility crisis and capture the full benefits of AI for science.
AI itself could help close the generation-verification gap, but only if we invest in the necessary infrastructure. AI has already begun to automate significant parts of the scientific workflow. Meanwhile, the Gold Standard Science requirements, including reproducibility, data sharing, and methodological documentation, create precisely the conditions under which automated verification becomes possible. The combination of both could lead to low-cost, continuous AI-enabled verification. Researchers have already outlined one vision of such a system, in which specialized agents parse submitted papers, reconstruct computational environments, execute analyses in sandboxed settings, and compare outputs against claimed results.176 The same infrastructure that audits human-authored papers today could tomorrow judge which machine-generated hypotheses merit experimental resources. Rising to this moment of need, NIH has launched a new, agency-wide initiative to elevate replication and reproducibility studies, identifying critical research and infrastructure needs to advance rigorous findings that are verifiable and transparently shared.177
Looking forward, the Federal Government must continue to lay the connective tissue between verification infrastructure and our scientific enterprise. This means establishing open APIs and interoperability standards that allow verification capabilities to plug into journal submission systems, grant reporting platforms, and private-sector AI research tools; standards for replication packages that ensure computational research arrives in machine-auditable form; and prizes for successfully replicating or disproving influential papers. The result should be a verification system that is not occasional but continuous, low-cost, and commensurate with the scale of discovery we are now capable of producing.
IDEAS ON THE HORIZON
The printing press transformed what could be written, who could read, and how knowledge accumulated. The research university created entirely new apparatuses for producing knowledge. The tools emerging today will do the same, enabling new forms of collaboration, new standards for verification, and new mechanisms for allocating attention and credit. Developments in mathematics already underway offer a clear glimpse of the transformative potential of AI paired with Gold Standard Science. In mathematics, checking a proof is often far easier than discovering one, an asymmetry in favor of verification that makes it a natural proving ground for the potential of AI and Gold Standard Science. With proof assistants, the challenge of proving a novel mathematical result reduces to the formalization of a theorem statement and the construction of a chain of arguments that the proof assistant accepts. Nevertheless, formalization has historically been too laborious to matter. Translating a single theorem into machine-checkable code could take months of painstaking work. The Liquid Tensor Experiment, a project to formalize a result in condensed mathematics in 2020, consumed nearly two years of effort from expert practitioners.178 Over the past two years, large language models have begun to make it possible for mathematicians to translate ordinary mathematical writing into these formal languages in real time. The acceleration has been striking. In early 2024, an ambitious project set out to formalize the Prime Number Theorem with a proof assistant. After 18 months and the collaboration of more than twenty people around the world, it had made intermediate progress but remained stuck on core difficulties in complex analysis.179 Then, in September 2025, a startup using AI completed the project in three weeks, spanning 1,100 formally verified theorems and definitions.180 Mathematical collaboration has traditionally relied on small, trust-based networks where participation depended on reputation and proximity. Formal verification replaces that model with one grounded in mathematical certainty, allowing collaboration to scale beyond personal trust. It has been suggested that mathematicians of the future may become architects of industrialized systems rather than solo artisans.181 The profession could grow to include orchestrators who design proof strategies, domain experts who contribute specialized knowledge, and skilled practitioners who direct AI tools. The mathematics we pursue will change as well. When AI handles computational drudgery, entire classes of problems become tractable, opening new scientific frontiers. Proof assistants and AI-enabled verification in mathematics represent a prototype of the Gold Standard Science tools that could propagate across disciplines. Wherever checking an answer is easier than finding one, AI stands to reorganize not just scientific discovery, but the social structures that govern who does it and how.
RETHINKING SCIENTIFIC PUBLICATION
The journal system was designed for a different era. When scientific journals emerged in the 17th century, they served perhaps hundreds of active researchers who corresponded by post. Today, there are nine million full-time researchers worldwide, publishing millions of articles across tens of thousands of journals.
The infrastructure of scientific communication has not kept pace with the scale of science itself, and AI will only widen the gap.182 The publication system’s structural problems go beyond scale. Journals create artificial scarcity, rewarding secrecy rather than open collaboration.
A small number of anonymous, unpaid reviewers who may have vested interests, limited expertise, or simply not enough time, determine what counts as legitimate science. The format rewards polished narratives over honest accounts of the research process. Null findings, failed experiments, methodological details, and the true rationale behind research choices rarely reach publication. These challenges will only sharpen with AI-enabled science. When anyone can generate plausible-looking research at industrial scale, the current metrics for evaluating scientific productivity, like papers published, citations accumulated, and impact factors achieved, will all fall to Goodhart’s Law as gameable targets. As information technology evolves, select research organizations backed by private funding have stopped supporting traditional journal publications. Their researchers release findings through alternative channels, including preprints, data repositories, and dynamic notebooks, which get reviewed and replicated rapidly within their community. They find that when researchers stop optimizing for publishable units, they design experiments differently. They become more creative, more collaborative. They care about whether results are useful rather than whether they make a compelling story.
The future of scientific communication may look very different from the present. Researchers might release shorter outputs more frequently, including datasets, code, preliminary findings, and methodological notes. Dynamic papers could update automatically as underlying data changes. Public peer review, conducted in the open rather than behind closed doors, could offer faster feedback loops. This is clearly seen by reference to machine learning communities, which already rapidly replicate papers posted to online repositories and turn social media platforms into forums for debate.
NEW FORMS OF COLLABORATION AND CREDIT
Today’s frontier advances in AI-for-science, such as AI models and autonomous laboratories, remain largely reflective of the existing structure of science. But combined with emerging decentralized technologies, they point toward the possibility of a more profound transformation in AI agents. Those agents would not merely assist human researchers, but participate as autonomous actors in a scientific economy.
One key building block of this transformation will be more granular credit attribution. Blockchain-based systems can create immutable records of scientific contributions, timestamping every dataset uploaded, every analysis run, and every hypothesis proposed, and linking each to its creator.183 When the record is fully traceable and captures every contribution comprehensively, credit attribution need not be zero-sum. Contributions to shared resources, such as datasets, code libraries, and protocols, become properly visible and rewardable. Another building block will be new modes of financial transaction for scientific knowledge. Decentralized Autonomous Organizations, communities that pool resources and allocate them through collective governance, are beginning to fund scientific research directly without going through traditional institutional gatekeepers.184 Prediction polls, augmented with proper scoring feedback and statistical aggregation, have also been shown to forecast scientific developments better than prediction markets, based on technological trends already underway.185 Together, these mechanisms can direct resources toward problems based on the wisdom of crowds rather than committee review, potentially faster and more effectively.
These pieces lay the foundation for a continuous, market-mediated, agentbased scientific economy. Imagine a funder posting a million-dollar bounty for the first validated therapeutic target for a rare disease. An agent working on adjacent problems notices a promising lead and posts a smaller bounty for replicating the finding. Other agents assess whether the problem falls within their competence, bid for the work, and contract an autonomous laboratory accessible through the internet, which runs the experiment and returns cryptographically signed results. The agent evaluates the evidence, updates its models, and publishes conclusions to a distributed ledger. When results prove ambiguous, human experts provide the judgment that automated systems lack. Smart contracts release funds automatically as milestones are verified. In such a world, experimental information becomes a tradeable commodity, and price mechanisms replace slow institutional coordination. Markets could form to support the scientific enterprise, such as prediction markets informing grantmakers about technologies on the horizon, bounty markets directing resources toward unsolved problems, and reputation markets tracking which agents produce reliable results. Agents would interact directly, exchanging data, hypotheses, and compute time through microtransactions. The whole system runs continuously, at speeds no human institution could match, but is guided by human judgment about which breakthroughs merit large bounties, and which questions require framing that machines cannot yet provide.
An agent-based scientific economy will reshape what science gets done. Agents might specialize in replication, profiting by verifying or falsifying claims that humans find too tedious to check. Others might focus on negative results, which journals refuse to publish but which hold real value for anyone exploring the same territory. Unconstrained by disciplinary boundaries, career incentives, or the limits of human attention, agents could pursue the questions that matter most, rather than the ones that yield publishable results. Cloud laboratories become the factories of this economy. Robotic facilities already exist that can synthesize molecules, run assays, and return results without human intervention. As these facilities proliferate and standardize interfaces, they become nodes in a network that any agent can access. An AI pursuing a hypothesis about protein folding could contract with a lab in Colorado, run crystallography experiments, receive results within hours, and integrate them into its next round of reasoning as it collaborates with humans in Boston. Physical experimentation, long the bottleneck of empirical science, becomes as accessible as computation. None of this exists today in a mature form, but the pieces are emerging separately. Whether they will combine into something like the system sketched here, or into something we cannot yet imagine, remains unknown. But the vision belongs in the same tradition as Bush’s original argument, that the frontier of scientific knowledge is open, expansive, and worth pushing into. The duty to keep pushing falls squarely on us.
As We May Build
For millennia, scientific knowledge and technological progress were bounded by the cognitive faculties of the human mind. Knowledge, however collective in its making, had to fit inside the heads of individual thinkers, flow through human patterns of communication, and conform to the social technologies we invented to guide inquiry. That era is ending.
Our civilization has been built on bronze and steel, substances we discovered and exploited, but did not design. The 21st century will be built on materials we engineer from first principles, metamaterials that bend light in ways nature never attempted, programmable matter that reconfigures on command, selfassembling structures that grow like living things but serve engineered purposes. The progression from the forge to the semiconductor fab took centuries; the progression from semiconductor fab to molecular assembler may take only decades.
We may begin to engineer cells as precisely as we now engineer circuits, programming immune systems to hunt malignancies with complete specificity, shaping cell differentiation and tissue growth to repair damaged organs, and designing therapeutics atom by atom rather than discovering them by trial and error. If we get all this right, within a generation, the diseases that today kill millions—like cardiovascular failures, neurodegenerations, and cancers—may yield one by one to instruments we are now starting to build.
The technological transformation is already underway. In the first year of the Trump Administration, more than a trillion dollars of investment commitments have been secured for advanced manufacturing infrastructure and for technology companies building in the physical world. The best minds of a generation are bent on breakthroughs in machine intelligence and its applications to science. New companies are created every day to discover new materials, design revolutionary drugs, build fusion power, and explore unsolved conjectures in mathematics. In parallel, a revival in the crafts has made advanced technology possible. Americans are grinding precision bearings to tolerances measured in millionths of an inch, polishing optics for surgical lasers and microscopes, spinning carbon nanofibers for spacecraft and medical implants, and growing semiconductor crystals of inhuman purity. The nation is rediscovering its capacity to build, grounded in the recognition that the frontier advances on two kinds of knowledge: the explicit, which can be written down and taught, and the tacit, which can only be learned through practice. America’s strength has always come from a culture that honors both science and craft, and keeps both open to all with the aptitude and interest to learn.
To sustain this progress, we must invent new ways of doing science. Science is the pool of knowledge that underlies our technological pursuit. The science of the coming decades could produce knowledge that no single person fully grasps, verified by systems that no single person fully audits, yet more reliable than anything we have built before. Future infrastructure for discovery may harness trillions of AI agents running experiments, testing conjectures, and surfacing insights across every scientific domain, with human researchers setting directions, posing questions, integrating findings, and making the judgments that require wisdom rather than computation. We urgently need to begin preparing for this AI-enabled future, by building the institutions, incentive structures, and information systems that let us trust what we cannot individually comprehend and steer what we cannot fully predict. When Vannevar Bush wrote to President Roosevelt, the nation faced a choice, whether to continue the wartime mobilization of science, or let the momentum dissipate. We chose to build. The institutions that emerged gave America a half-century of scientific dominance that translated into security and prosperity. But they are no longer sufficient for the new frontier we face today.
This report has described what must replace them: new partnerships that bridge discovery and production, new mechanisms that reward boldness over consensus, new infrastructure that reunites science with manufacturing and craft, and preparations for an AI-transformed era of scientific discovery. Our competitors understand this; they are building their own systems to capture this next era of science and technology, and to shape what it will be used for. The task, then, falls to our generation to design the institutions, standards, and capabilities that can guide a scientific enterprise larger, faster, and less individually comprehensible than any in history. In doing so, we will determine not only the future of American technological prowess, but also the trajectory of human knowledge itself. Rising to this challenge is vital if America is to continue to deliver prosperity and security to its people.
Leave a Comment
Thank you for your comment!
It will appear after review.