On April 29, 2024, the U.S. President's Council of Advisors on Science and Technology (PCAST) released its report "Supercharging Research: Harnessing Artificial Intelligence to Meet Global Challenges." The report presents five findings and recommendations for AI-accelerated research, aiming to drive the safe and effective application of AI, empower human researchers, explore scientific possibilities, and reduce risk in an increasingly digital world. The meta-strategy reviews the important contents of the report and provides a reference for readers to discuss the use of artificial intelligence to accelerate scientific research.
1. The future vision of AI-enabled R&D
Given the rapid development of AI technologies, especially generative AI, in scientific applications, it is difficult to predict how and when scientific workflows will evolve as AI technologies are integrated. Therefore, there are three areas that need to be continuously focused on how to use AI to accelerate scientific research:
(1) empower human scientists to make extensive use of AI tools;
(2) responsible use of AI tools;
(3) Sharing basic resources of artificial intelligence.
Figure Three areas to focus on using AI to accelerate scientific research
Shared AI infrastructure, such as increasing usage time on high-performance computing clusters, will enable scientists to use and learn about advanced AI tools. This helps to develop common standards for the responsible use of AI and promotes equity by providing access to AI to researchers in all institutions. The responsible use of AI tools will encourage the sharing of models and data in a secure manner and facilitate the careful design of AI-assisted research projects to improve the quality of scientific output. Empowering human scientists to apply AI tools widely will provide the scientific community with innovative solutions to pressing challenges, thereby justifying investments in shared resources and responsible use of AI, and enabling the creation of a community of stakeholders to continue to grow and consolidate these investments.
2. Five major discoveries and suggestions for AI accelerated research
Finding 1: Important research is hampered by the lack of access to advanced models.
Scientific research is a collaborative effort, and in order for the research of a group of researchers to be replicated, validated, and referenced by other researchers, data and models must be widely made available to researchers. Currently, the largest and most powerful AI models are often proprietary and available to only a select group of researchers, making it difficult to independently validate the study of these models. If this continues, there is a risk of creating a new "digital divide" in the AI space, as well as a potential "AI monoculture", where AI applications rely on a very small number of vendors or architectures. Research and the scientific method thrive when different hypotheses and methods deal with the same problem, and as a result, the "AI monoculture" hinders scientific progress because it may miss out on better answers. Currently, the cost of acquiring cutting-edge AI models, as well as the large disparity in returns between leading tech companies and academic or government agencies, make AI development uneven. In addition, it is important in the public interest to conduct research to explore the weaknesses and biases of AI systems, and to quantify as accurately as possible the effectiveness of various approaches to mitigating the shortcomings of AI systems, in order to understand how to deploy these powerful AI models responsibly, and to detect and respond to the effects of irresponsible use of these AI tools. Since the performance of an AI model is highly dependent on its size, it is important that all AI foundational researchers have access to both small and large models. The public and private sectors have different motivations and goals for using AI, but they can work together in value to create a culture of responsible use of AI.
Recommendation 1: Expand existing practices and share AI infrastructure widely and equitably.
In order to support scientific needs in an open and cost-effective manner, and to encourage experimentation and innovation, more support must be provided for the sharing of AI models, datasets, benchmarks, and computing resources. This benefits not only large AI companies, but also academic researchers, state and federal labs, smaller companies, and nonprofits. In the U.S., a related practice in this regard is the creation and funding of the National Artificial Intelligence Research Resource (NAIRR), which, at its current scale, cannot provide computing resources or models on the same scale as private sector (most of its models are closed-source and proprietary) or similar projects in other countries. But further expanding these shared resources would be an important first step to allow researchers to take full advantage of the state-of-the-art AI models available, while avoiding duplication of effort.
Finding 2: Cutting-edge research requires access to high-quality data.
Cutting-edge research in the future will be increasingly data-driven, and the training of machine learning techniques in particular relies on large numbers of high-quality datasets. For example, successful protein folding models have been trained on hundreds of thousands of existing databases of protein structures. AI climate models rely on a variety of existing data, including historical records, supercomputer simulations, satellite data, and a combination of three using reanalysis methods.
The federal government has many additional data sets that are invaluable for many scientific applications as well as informed decision-making on important challenges, and more value from these data remains to be discovered. Data created by the private sector also has the potential to improve public welfare. For example, the private sector more often has data on the chemical and physical properties of compounds and materials. Many federal datasets contain sensitive personal information and therefore cannot be publicly distributed; However, while protecting the privacy of individuals, it is still possible to use this data for scientific research, for example, by restricting access to aggregated and anonymized information, or by using synthetic datasets extracted from the underlying data.
Recommendation 2: Expand secure access to federal datasets, with appropriate safeguards and safeguards in place, to meet critical research needs.
The issue of secure access to federal data is sensitive and technically complex, but the benefits of allowing limited access to such data by approved researchers, as well as publishing anonymized versions of such datasets to resource centers such as future national AI research resources or national security data services, are significant and should be encouraged to expand existing secure data access pilot programs. In addition, existing guidelines for managing federal databases could be further developed to include cutting-edge developments in privacy-preserving technologies, such as differential privacy, homomorphic encryption, federated learning, and the use of synthetic data. While raw datasets, especially those labeled with relevant "metadata," already have significant research value for both traditional data science and the latest AI data analysis, more carefully planned and maintained databases dedicated to research purposes will have a more transformative impact across multiple scientific fields. Currently, high-quality data collation is an expensive and arduous task that requires a large number of professionals to carry out the work, and there is great potential for automating the collation process using AI technology. In the future, we anticipate that it will be feasible to upgrade many existing federated datasets to this high-quality database using AI tools, and we recommend this as a long-term goal of the federated data sharing initiative.
Discovery 3: AI provides a unique resource for collaboration between academia, industry, and federal government departments.
U.S. industry has invested tens of billions of dollars in basic and applied AI research. The National Science Foundation (NSF) recently announced a $140 million investment to establish seven new national AI institutes, including the National AI Research Resource Pilot Program, across NSF departments. The AI infrastructure provided by these institutes, as well as the interdisciplinary nature of AI itself, will lead to highly interdisciplinary research. It can also bring together U.S. government and federal laboratories and user facilities, the private sector, and scholars in the basic and applied sciences for novel and potentially transformative collaborations. These platforms are developing data sharing infrastructure while incorporating AI tools as part of community building, with plans to bring in other agencies and more industry partners.
Recommendation 3: Support basic and applied AI research involving collaboration between academia, industry, national and federal laboratories, and federal agencies.
The line between federally funded academic research and private sector research is blurred, with many researchers affiliated with academic institutions, nonprofits, and/or private companies, and a large portion of research and R&D in the United States is currently supported by private companies. To take full advantage of the potential research benefits of AI, research involving a variety of promising and fruitful hypotheses and methods needs to be supported. This may require funding agencies to relax their stance on how they work with industry and which researchers can be supported, in order to promote innovative research efforts and collaboration between different sectors. While the private AI sector can bring significant and very beneficial computing resources, expertise, and funding to collaboration, there is a need to ensure that market incentives from industry partners are aligned with the public and scientific objectives of the project, while clarifying IP and licensing plans at the outset. Examples of cross-sector collaboration can include the creation of high-quality public scientific datasets from multiple sources, the creation of multimodal foundation models, or the development of next-generation technologies such as new quantum computer qubit architectures. As the field of AI rapidly evolves, federal agencies and the private sector will need to explore and reassess how emerging national AI infrastructure needs to evolve.
Finding 4: Without proper benchmarks, validation procedures, and responsible practices, AI systems may provide unreliable outputs, their quality is difficult to assess, and can be detrimental to the scientific field and its applications.
Scientific research is an interconnected ecosystem, and the research, models, and data of one research group will influence the work of other researchers in both basic and applied fields. If AI tools are used irresponsibly, they can have a negative impact on this ecosystem, as these findings are subject to algorithmic bias, cannot be replicated, lack quantification of their uncertainty, overfit training data at the expense of accuracy compared to real-world data, or fail to be validated by more transparent techniques such as numerical simulations and laboratory experiments. In addition, personal data privacy or the intellectual property rights of data creators must be protected and respected. In the long run, publishing improperly labeled synthetic data can lead to "data pollution" that negatively impacts further scientific work. Deploying a fully automated AI tool without expert oversight carries extremely high risks. Academia must develop and refine publishing standards for research and data that rely on AI inputs. Higher education institutions need to modernize the training of junior scientists so that they can take full advantage of new AI tools, be able to independently verify and measure the output of these tools, and use them ethically and responsibly. The ever-increasing pace of scientific progress also requires a higher level of two-way public engagement from researchers and government scientific institutions. Some dual-use research applications, such as gain-of-function studies of biological pathogens, will require higher levels of surveillance, regulation, and oversight.
Recommendation 4: Adopt principles for responsible, transparent and credible use of AI at all stages of the scientific research process.
Institutions such as the National Science Foundation (NSF) and the National Institute of Standards and Technology (NIST) should continue to support scientific groundwork research on responsible and trustworthy AI, including the development of benchmarks for measuring AI standards, the characteristics of AI models (such as accuracy, repeatability, fairness, resilience, and explainability), and the development of tools to assess bias in datasets and distinguish between synthetic and real-world data. Institutions should seek the input of researchers in these fields, as well as relevant disciplines in the social and behavioral sciences, to help develop policies for the responsible use of AI. These policies need to be constantly updated with our understanding of the theory and practice of these technologies. We encourage developers and users of AI technologies, policymakers, and experts in the humanities, law, and social sciences to continue their dialogue on broader topics such as AI governance, use, and impact. Broad dialogue and diverse inputs are particularly important due to increasing industry consolidation around governance tools, with a small group of technology and advisory bodies dominating most of the market. At the same time, managing the risk of inaccurate, biased, harmful, or non-replicable research results generated in the application of AI should not only be considered when a research project is underway, but needs to be planned from the initial stages of the project. Federally funded agencies should consider updating their guidelines for responsible research conduct to require researchers using such technologies to develop plans for responsible use of AI. These plans should include an assessment of potential AI-related risks, such as algorithmic bias, disclosure of sensitive information, lack of transparency or repeatability of results, or potentially harmful applications, and describe the measures taken to mitigate these risks, in particular the relevant oversight procedures for any automated processes. These requirements can be modeled after existing data management program requirements and are modeled after the AI Bill of Rights Blueprint and the AI Risk Management Framework published by NIST. To minimize the additional administrative burden on researchers and build a culture of accountability, institutions can help enumerate key risks and provide potential risk mitigation processes. In addition, data management program requirements can be updated to require disclosure and documentation of AI tools used in research projects.
Finding 5: Optimal performance requires a combination of artificial intelligence and human expertise.
The most effective and valuable application of AI is in the form of assistive tools to help science and research solve pressing problems. In addition, generative AI resources are too unreliable for fully automated continuous operation independent of human oversight or intervention at the current level of technological maturity, especially in scientific fields where accuracy, explainability, and replicability are critical. In addition, the advantages of human intelligence and artificial intelligence are largely complementary. AI can tirelessly find patterns and perform repetitive tasks from huge data sets, while humans can infer and draw conclusions from far fewer amounts of data, make systematic and strategic reasoning, and coordinate actions with other humans and AI assistants. Current AI tools are still quite weak at demonstrating true creativity, analysis, or high-level strategic thinking. Even in a futuristic world where AI-assisted tools are already very common, and with more AI technologies available in addition to the current state-of-the-art AI models, such as large language models, we expect scientific research to continue to be guided by human scientists. Scientists will harness the unique human ability to draw high-level conclusions from relatively small amounts of data, complementing AI's ability to generate recommendations and connections from large data sets and automating more routine and tedious processes in research.
Recommendation 5: Encourage innovative approaches to AI-assisted integration into scientific workflows.
Scientific research is a great sandbox where researchers can practice, explore, and evaluate new models of collaboration between humans and AI assistants. Such examples could include: scientific consulting AI agents, which act as complex natural language interfaces for the operation of complex software or laboratory equipment; Combine generative AI algorithms with human expert feedback and validation methods (or other types of AI systems) to improve the accuracy of generative AI; or using AI tools to carry out interdisdiscentric, decentralized or crowdsourced research projects on a scale that would otherwise be unattainable. Our goal should not be to maximize automation, but rather to leverage complementarity and allow human researchers to conduct high-quality scientific research with the responsible use of AI. Funding agencies should recognize the emergence of new workflows and design flexible procedures, metrics, funding models, and challenge questions to encourage (but not oblige) experimentation in new AI-assisted ways to organize scientific projects. More broadly, incentive structures may need to be updated to be able to support different types of scientific contributions, for example, using feedback from human experts to train AI systems, developing new AI software tools specifically for scientific applications, or curating high-quality and widely available datasets. These contributions may not be as recognized as they should be through traditional research productivity metrics. The new metrics will complement traditional metrics and provide a fuller understanding of the various types of contributions that researchers make to scientific progress.
III. Conclusion
The world today is in dire need of scientific progress to address a variety of global and societal challenges. AI tools, if properly and responsibly supervised by human experts, trained on the basis of high-quality data, and validated using sound science and technology, can be an engine of innovation and fuel the ability of scientists and policymakers to address these challenges. Achieving these goals requires broad access to advanced models, datasets, and benchmarks to the scientific community, the integration of responsible AI practices into scientific workflows, and continued investment in the foundations of AI and its various scientific applications. While building such an AI infrastructure and scientific culture would require enormous resources and effort, it would also yield huge rewards: a broad and equitable R&D ecosystem that is both open and secure. In this ecosystem, scientific ideas can be quickly transformed into experiments, prototypes, and successful solutions to pressing problems. AI tools can complement traditional research work and rely on AI assistance, but human-directed collaboration and research paradigms still meet the highest scientific standards of validation, replicability, and objectivity. AI technologies can make significant contributions to science, but we must remain vigilant and rigorously identify and manage their weaknesses and risks. While AI will be the means and means of this transformation, it will ultimately be humans who will be empowered. Whether scientists or the general public, tackle grand challenges by increasing their ability to collaborate and gain a deeper understanding of the complexities of the world around us. We should embrace this opportunity with enthusiasm, fully aware of the risks that may arise if these AI tools are used irresponsibly, and ensure that AI is used to accelerate scientific research while staying at the center of the world to address global challenges.
Disclaimer: This article is transferred from Meta Strategy. The content of the article is the original author's personal point of view, and this official account is compiled/reprinted only to share and convey different views, if you have any objections, please contact us!
Transferred from 丨 Yuan Strategy
About the Institute
Founded in November 1985, the International Institute of Technology and Economics (IITE) is a non-profit research institute affiliated to the Development Research Center of the State Council, whose main functions are to study major policy, strategic and forward-looking issues in the economic, scientific and technological and social development of the mainland, track and analyze the development trend of the world's science and technology and economic development, and provide decision-making consulting services for the central government and relevant ministries and commissions. The "Global Technology Map" is the official WeChat account of the International Institute of Technology and Economics, which is dedicated to conveying cutting-edge technology information and technological innovation insights to the public.
Address: Block A, Building 20, Xiaonanzhuang, Haidian District, Beijing
Phone: 010-82635522
WeChat: iite_er