Today’s discourse on artificial intelligence (AI) is dominated by the awkward exchanges between so-called AI accelerationists and AI doomers., In response to these debates, we formed a team of senior computing researchers, along with rising stars in AI (see Appendix I), to develop a nuanced take on AI’s impact. There is no shortage of studies from international bodies aimed at AI regulation or policy, including the recent final report of the United Nations High Level Advisory Body on Artificial Intelligence. Instead, this article is by researchers and primarily for researchers, taking into account recent developments relevant to AI’s impact in a variety of economic and social domains and highlighting key opportunities for upside impact. Our view is that we are still in the early days of practical AI, and focused efforts by practitioners, policymakers, and other stakeholders can still increase AI’s upsides and reduce its downsides. In particular, rather than trying to predict what it will look like in the next five to 10 years, we encourage researchers to identify concrete research milestones that, if achieved, will enhance AI’s positive impact in the near future. Effective milestones are predefined targets for high-impact research in AI, such as a platform for civic discourse, a worldwide tutor, a system for rapid workforce reskilling, or a scientific collaborator. We also introduce a new nonprofit organization that will fund a novel approach toward reaching these and other milestones.
We focused on AI impact in seven broad areas: employment; education; healthcare; information, news, and social networking; media and entertainment; governance, national security, and open source; and science. These choices were based on whether new academic research and research-based practical applications could increase their positive impact. For example, billions of dollars are already being invested in self-driving cars, so despite its large impact, it has lower priority than our other options.
To complement our expertise, we interviewed two dozen domain experts with distinctive experience relevant to our topic (see Appendix II) including Nobel laureate John Jumper on science, President Barack Obama on governance, former national security adviser Susan Rice on security, and philanthropist and former Google CEO Eric Schmidt on governance and science.
This process led to four guidelines on how to shape AI for the common good:
- Humans and AI systems working as a team can generally do more than either on their own. Applications of AI focused on human productivity produce more positive results than those focused on replacing human labor.,, AI systems should initially aim at removing the drudgery of current tasks to encourage users to later embrace more advanced AI tools. If they automate menial and unfulfilling aspects of current jobs, work can be more meaningful and enjoyable. Tools aimed at making people more productive increase their employability and empower them to act as safeguards if AI systems veer off course. In short, focusing on human productivity helps both people and AI succeed.
- To increase employment, aim for productivity improvements in fields that would create more jobs. Despite tremendous productivity gains in computing and airline travel, the U.S. in 2020 had 11 times more programmers and eight times more commercial airline pilots than it had in 1970. This growth is because programming and airline transportation were fields with what labor economists call an elastic demand. Demand for agriculture, on the other hand, is inelastic in the U.S., so productivity gains have reduced the number of agriculture jobs fourfold in one human lifetime (1940 to 2020). If policymakers and practitioners target AI systems at improving productivity in elastic fields, they will allay public fears that AI will decrease employment.
- The impact of AI varies by geography. While nations with advanced economies worry about AI displacing highly trained professionals, countries with lean economies face shortages of these same skilled experts. AI could make such expertise more widely available in these places, potentially enhancing quality of life and economic growth. AI systems could become as transformative for low- and middle-income nations as mobile phones have been. The increasing popularity of smartphones in low- and middle-income countries enables widespread access to multilingual AI models that can greatly help people in such countries gain access to information, education, media and entertainment, and more.
- Develop metrics and methods to evaluate AI innovations. To evaluate AI’s real potential, we must measure it accurately and design carefully tailored policies that take account of its benefits as well as its risks. In relatively lower-risk situations, such as the deployment of AI tools to support routine coding tasks, the marketplace and observational studies can assess the effectiveness of AI tools without needing extreme rigor. In high-stakes domains—such as critical infrastructure and national security—they cannot, because we can’t risk harming participants. Where possible, we need to use gold standards such as A/B testing and randomized controlled trials. Equally urgent is post-deployment monitoring to evaluate whether AI innovations do what they say they are doing, whether they are safe, and whether they have unintended externalities. We also need to continuously measure AI systems in the field so as to be able to incrementally improve them.
We believe researchers, policymakers, and companies can pursue these principles consistent with relevant laws and governance frameworks designed to maximize the benefits of AI and reduce its risks.
Having set the stage by listing the four guiding principles, we now review the seven domains that we explored. We encourage readers interested in more depth or nuance to read Cuéllar et al. or visit https://shapingai.com/.
Employment
Our first topic for nearer-term AI is a major concern: the impact on jobs. Indeed, a recent Global Public Opinion Poll on AI found that the majority expect to be replaced at work by an AI system in the coming decade. Concern might have been inspired by a 2013 prediction that nearly half of U.S. jobs could be “computerized” within a decade and subject to automation.
Technological advances have long led to the decline of some jobs and the creation of new ones. For the U.S. workforce in 2018, 63% had jobs that did not exist in 1940. Using decennial U.S. census data, the figure shows examples of four job classifications where numbers changed strikingly from 1970 to 2020.
Figure. Data from decennial U.S. census job classes.
Despite the downside of job disruption, a healthy economy relies on improving worker productivity. Two-thirds of the world’s population lives in countries with below-replacement birth rates and many nations are facing labor shortages. The U.S. already lacks critical workers as varied as K–12 teachers, passenger airline pilots, physicians, registered nurses, software engineers, and school bus drivers. To supply needed services, high-income countries must either greatly expand their working population or significantly improve worker productivity.
The impact of productivity gains on jobs depends on whether the demand for goods produced by that work is elastic or inelastic. Goods with elastic demand are those where a decrease in price results in a large increase in the quantity acquired. If product demand is sufficiently elastic, productivity-enhancing technology will increase industry employment. If demand is inelastic, however, productivity gains means jobs will be lost. For example, as noted earlier, agriculture is inelastic in the U.S., so gains meant dramatic declines in both absolute numbers (fourfold) and its portion of the workforce (from 40% in 1900 to 20% in 1940 to 2% today).
As mentioned earlier, programmers today are tremendously more productive than they were in 1970—they have more powerful programming languages and tools and computers are much faster—yet there were 11 times more programmers in 2020 (see figure). Jet engines and autopilot systems boosted pilot productivity; even so, the U.S. has eight times more commercial airline pilots today than it had in 1970.
Another perspective on employment is the split between nonphysical and physical tasks. In our view, the main impact of near-term AI systems will be on nonphysical tasks. We think robots will eventually have a large effect on the way in which physical tasks are performed in the world beyond manufacturing, but this may be five or more years behind the use of AI systems for purely digital or knowledge tasks. Continued progress in robotics could have large implications for many new areas, including elder care, disaster response, and construction.
While much of the discussion here has been about employment, the main issue in the U.S. is not unemployment but rather the quality and value of available jobs. Though the examples in the figure show that lower-wage jobs decreased and higher-wage jobs increased, income disparity overall has increased since the 1970s in the U.S. because the average, rather than the median, salary tracked productivity gains. The U.S. has achieved record-low unemployment, but the college educated have had much greater economic gains while high school graduates and dropouts earned much less,thereby hollowing out the middle class.
While industrialized nations worry about AI systems displacing highly trained workers like lawyers, doctors, and programmers, low- and middle-income countries face a shortage of such highly skilled workers. Making this expertise more widely available in those regions via AI systems could enhance quality of life and accelerate economic growth.
Education
Next is education, where productivity increases and greater fairness have long been computing targets. AI systems are already affecting classrooms, and some predict a significant impact from AI on all levels of education.
From an employment perspective, we believe that education is elastic, as there is a huge demand for improving the effectiveness and efficiency of learning. Indeed, the U.S. and many other high-income nations face a shortage of K–12 teachers, as well as STEM graduates who could teach those topics in K–12 schools.
Today’s AI tutors such as CK-12 and Khanmigo already help some students, but a major educational challenge in the U.S. is the performance gap between K–12 students in high-poverty schools compared to students in other countries or to U.S. students from low-poverty schools. If the primary users are students at low-poverty schools, AI tools could inadvertently exacerbate this performance gap.
Before many schools will deploy AI tools for their students, the tools will require careful evaluation, including randomized control trials (RCTs) to establish in what circumstances they help or hurt and, if so, by how much. At least in the U.S., it will be difficult to gain access to a sufficient number of K–12 students with the desired heterogeneity, as well as their test-score results and details about their backgrounds, schools, and teachers to assess the AI tools properly. A separate challenge would be to convince parents and teachers that their students should be the subject of educational experiments where potentially valuable opportunities would be unavailable to control groups. Another obstacle to deploying AI tools in K–12 schools in the U.S. is that there are many people beyond the teachers involved in decisions about what tools to use, including administrators, school boards, parents, and students.
Colleges may offer an easier initial target for assessing the benefits of AI systems in education, as the content is more up to the instructor and the courses are much larger. Also, community colleges play a major role in adult education, including retraining. If AI systems could enhance retraining, this improvement might partially compensate for the downside of job disruption from AI deployment in inelastic fields.
If some K–12 school districts in other countries or in some adventuresome U.S. school districts decided to deploy AI systems for students without waiting for RCT data, their results could be used as evidence in a natural experiment. The theory is that individuals are exposed to sufficiently different conditions naturally, rather than through a researcher’s design. Statisticians or econometricians then try afterward to find demographically matched groups and draw inferences. This approach essentially acts as if random assignment occurred, allowing researchers to observe and analyze the effects without actively manipulating variables. These studies are used widely in both healthcare and other policy areas.
For AI systems to eventually help most students, however, we recommend first improving the lives of teachers, as teachers largely determine the success of technology that gets deployed. Helpful AI for teachers could involve offering assistance with lesson plans, progress reports, homework assignments, and grading. To succeed, these initial AI solutions must be driven by the real, day-to-day challenges teachers face, and they must be aligned with teachers’ and students’ needs rather than based on the opinions of school boards or administrators. The targets should be set for and by teachers.
Healthcare
Healthcare, responsible for 16% of U.S. GDP, is next. As with education, many believe that society should offer high-quality healthcare regardless of individual wealth.
Evidence suggests that healthcare is also elastic: Demand for healthcare will increase more than proportionately as the cost and quality of healthcare improves. Indeed, the U.S. and many other countries are facing a shortage of healthcare professionals. Beyond improving the employment prospects of healthcare workers via productivity gains, AI tools must also keep healthcare professionals in the decision path for actual patient-therapy recommendations, as AI systems are not guaranteed to make the best recommendation 100% of the time. As people and AI systems tend to make different mistakes, the collaboration of experts with AI systems could help improve healthcare quality.
Healthcare decisions are made with life-or-death stakes on short timeframes based on complex data, requiring years of specialized training for the best human clinicians. But even then, human specialists have limitations: They are experts in only narrow subdomains, informed by firsthand experience with only tens of thousands of patients, bound by the preconceptions and imperfections of past medical knowledge, available only to the best-resourced healthcare systems, and unable to extract every pattern from the vast sea of medical data.
In addition, there are billions of genetic variants, terabytes of medical images, years of lab results, and nontraditional clinical data sources such as smartwatch readings, nutritional logs, and environmental risk factors—the complexity of this information inherently exceeds human understanding. Perhaps this is why approximately 15% of all diagnoses in the U.S. are incorrect and why most Americans will be misdiagnosed at least once in their lifetime. Such errors contribute to about 10% of all U.S. deaths. We envision a world in which all relevant health-related data and every past healthcare decision can be used to inform every future healthcare decision and benefit everyone.
We are currently far from that world. AI practitioners should be humble about the enormous challenges that must be overcome to provide advanced tools, especially given unrealistic past claims that AI will soon obviate clinicians. Some challenges involve open research questions that arise in the deployment of real-world systems, including those around fairness, usability, robustness, and interpretability, among others. Another important barrier is infrastructure and regulation: Few health systems have the infrastructure to easily deploy, update, and monitor algorithms, and health systems are cautious in light of strict regulations.
Though the path is long between the current state of medical AI and the world we envision, the rapid pace of progress in developing the underlying technology is cause for optimism. However, progress relies critically on data availability—large, diverse, multisite cohorts to ensure models perform robustly and fairly across many populations and conditions, as well as techniques such as federated learning, which allows AI models to be trained on many distinct pools of data without centralizing any of the raw data.
Policymakers and stakeholders should permit and encourage healthcare organizations to participate in multi-institution collaborations that use de-identified data to train machine learning (ML) models for the benefit of their patients and others worldwide. Policymakers and stakeholders should also insist on open standards for the interchange of health data and on including AI-based predictions and guidance in clinical workflows.
Information/News/Social Networking
While education and healthcare offer targets that could potentially increase the upsides of AI, the goal of this section is to discuss the risks associated with a widely feared downside of AI. In most of the scenarios discussed so far, AI systems act to provide information, such as tutoring in education or helpful diagnostic information in medicine. However, AI may accidentally generate incorrect information (misinformation) or be used to maliciously generate incorrect information (disinformation), such as in the case of false news presented as fact (especially an issue in election tampering), or generate visual imagery or spoken audio presented as real (deepfakes).
To achieve AI’s potential, we must maximize the benefits of AI-provided information but mitigate the effects of misinformation, disinformation, and confirmation bias. Though the threats of disinformation to personal well-being and international security are clear, the threat of partial misinformation or favoritism and preconceptions are more nuanced: AI systems are imperfect, yet people want their tools to be dependable and fair.
Often, people tend to overtrust AI systems, not only in cases when they are clearly wrong (for example, recommending an unsafe drug dosage) but also in cases when they are only partially correct (for example, radiology support that finds some tumors but not all) or unfair (for example, favoring some job candidates over others on the basis of protected characteristics). Market forces may further distort the response of an AI system if, for example, information providers allow sponsors to pay in exchange for AI-generated answers that favor certain products or viewpoints.
Solutions to overcoming AI misinformation, disinformation, favoritism, and preconception challenges will require not only high-quality AI systems, but also effective human+AI (and AI+AI) interaction—an area that has received significantly less regulatory and research attention. We must develop methods that provide users, developers, and regulators with control over and understanding of AI systems.
One piece of good news is that AI systems show early evidence of identifying disinformation. OpenAI created a tool for its DALL-E LLM that correctly identified 98% of the images it generated and misidentified only 0.5% of non-AI generated images as ones it created. Alas, it did not work as well with generative AI programs from other sources. Other studies demonstrated that watermarks could be deployed to identify machine-generated text.
AI systems can also help with civic discourse. One study compared discussions between people on opposing sides of an issue. Conversations in which the AI system would make suggestions on how to rephrase comments and questions more diplomatically before being communicated led to much greater understanding between the two sides than conversations without the help of AI. Another study used an AI system to hold discussions with conspiracy theorists. Conspiracists often changed their minds when presented with compelling evidence. A recent paper reports on two promising current platforms for civic discourse—the School of Possibilities and Polis—and proposes required features for success.
Figure. AI can find exits from conspiratorial rabbit holes.
Media/Entertainment
Unlike education and healthcare, many areas of entertainment are inelastic. It is not obvious that if AI improved the productivity of fine artists and graphic designers the market would grow to accommodate many more paintings and designs. And there is no shortage of novelists; publishers have inboxes full of unsolicited manuscripts. Even an author as successful as Stephen King had to adopt a pen name because his publishers feared he would write more books than the market would bear.
Journalism is very different from writing fiction. Journalists must write to a tight deadline while under tremendous pressure to avoid mistakes in their reporting. While there are some tedious news chores that are better left to AI systems—turning quarterly financial reports from companies into text or reporting on high school sports—investigative journalism and many other tasks are not low-hanging fruit for AI. CNET secretly (and now infamously) tried using AI to write dozens of feature stories, but then had to write long correction notices about “very dumb errors.” Given the importance of journalism despite its precarious economic state, AI tools that reduce the stress and burnout of journalists could significantly contribute to civil discourse.
The impact of AI systems on the movie industry is much more difficult to predict. The special effects using toy models and stop-action animation prior to the 1980s have been replaced by computer-generated images, the creation of which employs many more people, though the skill sets are very different.
Figure. Entertainment before movies.
Neal Stephenson provides an analogy for the potential impact of AI systems on entertainment by highlighting the impact of movies on stage actors in 1900. Back then they memorized plays, performed every night in front of a live audience with other actors, and had to project their voices to reach the back of the theaters. Imagine if these actors were told that the future includes: individual performances in a warehouse, no live audience, no need to project their voices, sometimes no other actors, and a single performance recorded and repeated in thousands of theaters for months. Actors would likely fear for their future and for their profession. Instead, live theater is still healthy on Broadway and the West End alongside cinema because audiences get different experiences from these performances.
Today’s image- and video-editing tools allow amazing control over every pixel on the screen, but they are infamously difficult to use given the need to set hundreds of parameters. If the tools themselves make it much easier for the director to maintain control of the thousands of microdecisions needed to create art—or if an AI aide can step in and remove the drudgery of such tools—then we can imagine a Cambrian explosion of feature-length films where the movie studios are no longer the gatekeepers of what can be made since filmmakers would no longer need to raise millions of dollars beforehand. If entertainment is a storytelling industry, one potential outcome is that AI systems could help more individuals tell more stories.
If such advances enable a thousand future Martin Scorseses or Steven Spielbergs to make movies with dozens of their friends, it is not clear whether the movie industry would be smaller, even if the job makeup for portions of the industry would be very different. The future of cinema could be more like what is happening today, with television and video-sharing platforms today for younger audiences. One report found that Gen Z watched only 20 minutes of live TV daily but spent more than 90 minutes on platforms like TikTok and YouTube. These new platforms have, in turn, created new, well-paid roles, such as influencers and content creators. Similarly, digitization of the music industry disrupted the prior business model, but it enabled a much longer tail of artists to reach listeners and eliminated the manufacturing of billions of plastic disks.
Governance/National Security/Open Source
As AI systems permeate workplaces and daily life, policymakers and the public will face choices—some relatively familiar and others that create potential policy and regulatory challenges. Like aviation, television, and the Internet before it, AI promises both bountiful rewards and potential perils. Yet its nature as a rapidly evolving, general-purpose technology that can be used to outsource human decision making sets it apart. As people come to use AI systems more often in their workplaces and daily lives, policymakers are confronted with decisions about not only whether or how to design new policies for particular AI systems or uses, but also how to interpret existing laws that already govern what people do. It is important not to ignore the risks of overly rigid legal arrangements.
These decisions will take place in multiple legal systems, and will be influenced by ongoing deliberations in a few key international fora that have already produced relevant statements on how to govern AI, including the G7’s Hiroshima AI Process, the International Telecommunications Union, undefinedand the Council of Europe. There are, of course, differences of perspective among different jurisdictions and groups of experts. That said, the American experience is especially relevant because most AI law and policy decisions will be implemented through national legal systems, and the U.S. has both unique global influence and a pivotal role in the AI ecosystem—hence the specifically American examples. Our broader point is about the balance needed in AI law and policy everywhere.
Existing laws already govern many AI applications. Tort law, for instance, holds entities liable for unreasonable risks when deploying AI systems in services like accounting. Sector-specific regulations, such as FDA oversight of AI-enabled medical devices or provisions of international humanitarian law governing military decisions, remain applicable. The challenge lies in interpreting these laws for novel AI use cases and supplementing them with targeted new laws addressing specific problems, such as the security of systems used to support critical infrastructure. These endeavors will often demand fact-specific and context-informed judgments, along with greater knowledge from policymakers about the technical attributes of AI systems.
Other challenges include addressing gaps in existing laws with carefully targeted policies that take into account the unique capabilities and benefits of advanced AI systems, and creating legislation that recognizes how quickly the technology is changing. Safety and security testing for AI models—including appropriate backups and fail-safes—for managing critical infrastructure, such as power grids or air traffic control, are crucial. Countries need strategies to determine how the most advanced AI models enable adversaries to engage in cyberattacks or to design specialized weapons, and to reduce those risks through international collaboration.
The debate over whether to open source AI models exemplifies the nuanced approach required. While sharing model weights and technical details can spur innovation, it may also aid adversaries. The devil is in the details: Legal terms of sharing, built-in safeguards, and the extent of disclosed information all factor into the equation. Accordingly, policies designed to limit the risks of open-weight model releases must be carefully designed to, as much as possible, retain the benefits of openness while limiting the ease with which openly available models can be reconfigured for malign use.
To address these issues, policymakers should consider some key principles:
- Balance benefits and risks: AI systems can exacerbate certain risks, but focusing solely on perils can stymie beneficial outcomes and innovation.
- Leverage existing legal frameworks: Rather than crafting entirely new regulatory schemes, adapt and apply current rules where possible.
- Fill in the gaps: Even if we rely on careful interpretations of existing laws to handle a large and underappreciated proportion of AI policy challenges, carefully designed new policies will help society manage a technology that can turn some forms of intelligence into a manufactured commodity.
- Holistic and transparent impact assessment: Impact assessments of evolving AI systems play a key role in ensuring people, organizations, and governments better understand the potential contributions of AI systems as well as their risks and limitations.
- Mitigate unfairness: There are myriad past examples of unfair AI systems that make unwarranted inferences or behave in ways that unfairly contravene what users have been told to expect. To mitigate these concerns, regulators should encourage best practices following lessons learned from these past failures.
- Invest in public interest and national security: Some promising policies involve investment and infrastructure rather than regulation.
- Embrace iterative policymaking: The rapid pace and general-purpose nature of AI development demand the continuous evaluation and refinement of policies. AI-enabled platforms can facilitate both evaluation as well as public consultation and deliberation.
Science
The poster child for AI in science is protein folding. Remarkably, the scientists involved in the AlphaFold project received a Nobel Prize just six years after the first version was released. AlphaFold addressed a 50-year-old puzzle: how to predict protein structures from their amino acid sequences. Michael Levitt, who received the Nobel Prize in a related field, said AlphaFold advanced the field by 10 to 20 years. More than two million scientists in 190 countries have used it.
Nobelist John Jumper told us that an enabling artifact was the Protein Data Bank (PDB). Started in 1971, this peer-reviewed repository holds information about the 3D structures of proteins, nucleic acids, and complex assemblies. Containing more than 200,000 examples, it is one of the best databases in biology, making it an attractive target for AI systems. More curated datasets would create more opportunities for scientific advancement via AI systems, as progress can be limited by the availability of high-quality, ground-truth data.
In many science fields—chemistry, materials science, biology—researchers are now using “self-driving labs” that combine robotics and AI systems to reduce the time to make a new scientific discovery.
It was only a dozen years ago that machine learning’s neural networks started outcompeting AI alternatives. It is difficult to overestimate the current excitement about the promise of AI within the broader scientific community. Here are three more examples:
- Controlling plasma for nuclear fusion. Researchers used an AI system to autonomously discover how to stabilize and shape the plasma within an operational tokamak fusion reactor. Stabilizing plasma is a critical step on the path toward stable fusion. Practical fusion would be revolutionary for climate change.
- Flood forecasting. Researchers were able to develop an AI model that achieves reliability in predicting extreme river-related events in ungauged watersheds at up to a five-day lead time, with reliability matching or exceeding those of instantaneous predictions (zero-day lead time). It now covers hundreds of millions of people in more than 80 countries. Similar progress has been made in AI-based weather forecasting.
- Contrail reduction. An AI model identifies areas where airplane contrails are likely to form, allowing for flight rerouting to reduce the climate impact of air travel. Initial tests by American Airlines showed, with minimal fuel increase, a 54% reduction in contrails, which are 35% of the global warming impact of the aviation industry. Aviation produced 950 megatonnes (mt) of carbon dioxide equivalent emissions in 2023. If the entire industry had similarly reduced contrails, the decrease would be 19% (54% of 35%), or 180 mt. For perspective, that amount is five to 10 times larger than the 2023 data center emissions of Amazon, Google, Meta, and Microsoft combined.
Figure. Reducing contrails reduces climate impact.
Dario Amodei, co-founder and CEO of Anthropic, argues that with more powerful AI a system could perform, direct, and improve upon nearly everything biologists do. He projects that such an AI system might enable biologists and neuroscientists to make 50 to 100 years of progress in five to 10 years. undefinedScientific breakthroughs just on chronic neurological disorders could be life changing for the more than 10 million people worldwide—including two million Americans—who have been diagnosed with multiple sclerosis or Parkinson’s disease plus the 40 million Alzheimer’s patients (with seven million Americans). Each year, another 10 million people will be told they have a chronic neurological disorder.
Shaping the Goals of AI Researchers
To make the goal of shaping AI’s impact to help billions of people more concrete, in each of the seven domains above we proposed example milestones that, if achieved, could increase AI’s upsides or decrease its downsides (for more details, see Appendix III and Cuéllar et al.). To encourage progress toward these and other concrete research milestones, we propose a few mechanisms:
- The creation of $1 million+ inducement prizes for each milestone, which work well in many fields. Rather than recognize past achievements, inducement prizes try to stimulate research on focused targets (Appendix IV provides an example). Probably the best known inducement prize is the X-prize. While we believe progress in AI makes ambitious prizes plausible, participants can use any technology to win the prize.
- A novel way for governments and corporations to shape AI’s impact is to create a dvance market commitments that create economic demand for a product that does not yet exist; one can think of them as “inducement purchase orders.” This approach was successful in U.S. development of a COVID vaccine and is how NASA works with private rocket companies.
- For topics without ample research, create ad hoc, high-impact, multidisciplinary, five-year research labs following the Berkeley Lab model. (The table in Appendix V lists the five most successful of the 10 labs that cover one author’s academic career (Patterson from 1976 to 2016); Appendix VI explains their philosophy.) In addition to pursuing research that enables the milestones, labs can define success for the related prize, evolving the metrics and benchmarks that show whether the milestone has been achieved so as to award the prize.
We see these three paths as synergistic. Seekers of inducement prizes or market commitments can set research challenges, track progress, and provide feedback to research labs working toward key milestones. They can also offer targets for technology transfer that research labs need to demonstrate success.
AI Moonshots and Funding Them
As we wrapped up this study, we spent more than a little time attempting to imagine the AI equivalent of President Kennedy’s 1961 call for going to the moon. It was surprisingly frustrating, since it was hard to pick a single moonshot. For example, we could aim to create an AI mediator that orchestrates conversations across political chasms to pull us out of divisiveness and back into pluralism. Or a system that helps rapidly reskill those who lost their jobs due to AI’s impact on inelastic fields. Or we might enable biologists and neuroscientists to make a century of progress in a single decade. We finally realized that if we create the right innovation blueprint, we don’t have to pick one moon: We can use AI to launch a thousand moonshots.
At this point, readers might expect that without more and reliable government funding we cannot achieve progress on research milestones. Indeed, some call for a CERN for AI, employing thousands of scientists and spending billions of dollars per year. While we need government collaboration on the AI blueprint, the current disruption in research funding by the U.S. government makes the raising of billions of dollars for a big science approach to AI more difficult today than when it was first proposed in 2017. As an alternative, we suggest that new money for inducement prizes and research labs can come from the philanthropy of the technologists who have prospered in the computer industry.
An example of AI-directed philanthropy is Laude Institute, a recently launched nonprofit organization committed to enabling computer science research aimed at benefiting society. Donations come primarily from individuals who have benefited financially from computer science research. Its focus is on facilitating research breakthroughs and translating research into impact via open source and startups. One of its first projects was the AI Moonshots Grant Program that started in fall 2025. The initial grant program has four targets:
- Workforce reskilling to counter AI displacement. Many current workforce development programs have a modest impact on wages. The first moonshot is a computer system that could reduce the time for workers who are unemployed or in low-income jobs to gain an in-demand skill that is a ticket to the middle class. In the U.S., for example, a goal might be to empower a worker making 15,000, and to do so in three to six months (see Appendix IV for more details).
- A functional civic discourse by 2030. This moonshot is for a platform that mediates conversations or attitudes to enhance public understanding and civic discourse. Features measured could include the breadth of the topics covered, the effectiveness of the tool versus the difficulty of each challenge, how widely it is used and by how many parties, and so on.
- Broad AI for healthcare practitioners. A broad medical AI system would learn from many data modalities—images, laboratory results, health records, genomics, medical research—to carry out a diverse set of tasks and be able to explain its recommendations using written or spoken text and images. Such tasks might include bedside decision support, interacting with patients after they leave the hospital, and automatically drafting radiology reports that describe both abnormalities and relevant normal findings while taking into account the patient’s history.
- A century of scientific progress in a single decade. As mentioned earlier, Dario Amodei argues that with more powerful AI a system could perform, direct, and improve upon nearly everything biologists do. He projects that such an AI system might enable biologists and neuroscientists to make 50 to 100 years of progress in five to 10 years. Keeping a human scientist involved with an AI collaborator would help ensure safety and affordability. Breakthroughs on chronic neurological disorders alone would be life changing for millions.
Laude Institute is offering two one-year 10 million-plus research lab and define a related $1 million-plus inducement prize. Depending on proposal quality and fundraising, Laude Institute will award prizes to one to four labs. An independent program committee of research leaders will select the awards—just as ACM and IEEE appoint committees to select recipients of the Thacker, Turing, and von Neumann Awards sponsored by industry—which will help allay fears that funds coming indirectly from industry could have undue influence on academia. Depending on how well this initial offering goes, Laude’s AI Moonshop Grant program could soon become a regular event, offering new moons to shoot for annually.
Measuring and Encouraging Impact
One obstacle to academic AI experts working on high-impact research that could benefit the public good is the fear their work will not be rewarded by their administrations. Citation counts and the h-index are easy-to-understand numbers—bigger is better—a form of recognition that is easily documented as part of a tenure or promotion case.
Alas, traditional scholarly measures do not always faithfully reflect impact. For example, the most popular paper from the plausibly most impactful project in the table (Appendix V) has only 5% of the citations of the highest one. We also note that papers that win test-of-time awards 10 to 20 years later rarely win the best paper awards at the time of publication.
To help document such impact and to encourage researchers to work on high-impact problems, Laude Institute is developing an impact index. Inputs to the index include, for example, awards from professional societies, measures of popularity of open source projects, founding of successful startup companies, technical recognition in large corporations, and so on (see table in Appendix V). While the gap between research contributions and award recognition can be long, the popularity of open source projects and the success of startups can be evident even more quickly than large citation counts, so they can help early in careers. Another benefit of the impact index is that it spotlights high-impact leaders for more junior researchers to emulate.
Conclusion
Several reports have surveyed the state of the art of AI and considered its potential rewards and risks.,,, Like some of these earlier studies, our approach has been to envision what the impact could be if researchers directed their efforts toward AI systems that benefited society. The computer industry has at times deployed technology quickly without fully considering societal impact beforehand, with social networking as a prime example. Concrete research milestones offer predefined targets that, if achieved, could help the public and advance the state of the art. We also propose a novel approach to funding and evaluation methods that measure and encourage high-impact research.
AI moves quickly, and governments must keep pace with—or even better, stay ahead of—developments. Decisions made today will shape the AI landscape of tomorrow, influencing everything from economic competitiveness to social stability. In the dawn of practical AI, thoughtful governance is not just desirable—it is essential.
Similar to how the government collaborated with industry on the successful development and deployment of cars and integrated circuits, establishing a coordinated, public-private partnership including governments, universities, and companies would be beneficial. The goals of this partnership would be to remove bureaucratic roadblocks where possible, to ensure the safety of the technology, and to educate and provide a transparent view of the development of the technology to both policymakers and the broader public. The proposed research guidelines and moonshots might be complemented by plans for low-income nations to help harness the benefits of AI for everyone. However, if new technology is sufficiently compelling and affordable, it has in the past had widespread penetration with good effect without planning. This year, 90% of people will have access to smartphones, which previously led to a 0.6% rise in the rate of GDP growth for low-income countries for every 10% increase in mobile penetration.
Though AI has risks, there are also many known and unknown opportunities. For example, some estimate that AI could plausibly raise the growth rate of the U.S. GDP from its current 1.4% to 3%. Doubling the growth rate could lead to “poverty reduction, better health care, improved environment, stronger national defense, and a reduced budget deficit.” It could also potentially contribute to rebuilding the middle class of high-income nations and make needed expertise more widely available in low- and middle-income countries, potentially enhancing quality of life and economic growth.
We conclude that it can be as big a mistake to ignore the potential gains of AI as it is to ignore its risks. The future of AI depends on choices we make now—funding high-impact research, enabling real-world deployment, and building with clarity and purpose to steer AI toward outcomes that benefit all of humanity.
To view the appendices mentioned in this article, please click here.
DOI
10.1145/3746132
January 2026 Issue
Vol. 69 No. 1
Pages: 54-65
Strategies for Crossing the GenAI Divide
AI Is Being Kicked Off the Therapist’s Couch
A Seven-Step Guide to IT Risk Management in Healthcare Environments