Governing AI: A Blueprint for the FutureA five-point blueprint for the public governance of AI ~ Part 1 of this paper offers a five-point blueprint to address several current and emerging AI issues through public policy, law, and regulation. We offer this recognizing that every part of this blueprint will benefit from broader discussion and require deeper development. But we hope this can contribute constructively to the work ahead.MicrosoftMS_6a084ba6-414c-11e9-9c11-5766e90f6739Brad SmithVice Chair & PresidentSatya NadellaCEO ~ These suggestions build on the lessons we've been learning based on the work we’ve been doing for several years. Microsoft CEO Satya Nadella set us on a clear course when he wrote in 2016 that "perhaps the most productive debate we can have isn't one of good versus evil: The debate should be about the values instilled in the people and institutions creating this technology."GovernmentsAcademiaCivil SocietyAI improves people’s lives_f5b1e092-fcad-11ed-b89a-14061883ea00To address current and emerging AI issues through public policy, law, and regulation_f5b1e376-fcad-11ed-b89a-14061883ea00Artificial IntelligenceWe’ve long said that advancing AI responsibly is a journey, and our own years-long effort to build a responsible AI program at Microsoft has prepared us for this AI inflection point. As we continue to unlock greater benefits from the latest AI technologies, we remain clear-eyed about risks and mindful of the important role we play in advancing the state-of-the-art, not only for AI capabilities but for responsible AI governance, mitigations, and culture-building.GovernanceOur governance approach begins with how we structure and organize responsible AI at Microsoft, with coordination from the Office of Responsible AI and essential involvement across every part of the company—core responsible AI teams in engineering, research, and policy, embedded Responsible AI Champions throughout organizations, executive leadership and accountability as embodied in the Responsible AI Council, and oversight from Microsoft’s Board. Governance extends to creating, maintaining, and implementing a shared set of rules and policies to operationalize responsible AI, which we do with our Responsible AI Standard. It also requires additional oversight and expert guidance for higher-risk or novel-use cases like the development of the new Bing, which is where our Sensitive Uses program of required reporting and deeply engaged, case-specific review is so critical.CommunityCutting across all our work is the imperative to build and sustain culture and community. In addition to investing in existing people, hiring new talent, and developing training and skills-building, we have and will continue to prioritize diversity, collaboration, and the capacity to see AI systems through a sociotechnical lens. Finally, Microsoft is committed to proactive, practical steps that institutionalize not just a culture of responsible AI within the company, but tangible tools and capabilities that make AI safer and more reliable for our customers and society. TransparencyWe will continue to be transparent and share our learnings broadly. We know that our efforts will require adjustments and course corrections, especially as we learn from those outside the company. As societal conversations and government oversight of AI evolve, we will continue to share our commitments for the responsible development and deployment of AI. We will also share our thoughts and suggestions about policy, regulation, and the role that private-public sector dialogue and partnerships can play, as we have done in our blueprint for AI policy, law, and regulation.CollaborationThe current AI moment calls for industry, governments, academia, and civil society to come together to better define the boundaries for AI in society. We welcome a robust, global, cross-sector discussion of how to build and deploy safe, secure, and transparent AI systems. We hope that by sharing more details on our responsible AI efforts, we are contributing useful information to this conversation. ^^ Together, we can build a future where AI advances the best of humanity. SafetyImplement and build upon new government-led AI safety frameworks_f5b1e538-fcad-11ed-b89a-14061883ea001U.S. National Institute of Standards and TechnologyFirst, implement and build upon new government-led AI safety frameworks. The best way to succeed is often to build on the successes and good ideas of others. Especially when one wants to move quickly. In this instance, there is an important opportunity to build on work completed just four months ago by the U.S. National Institute of Standards and Technology, or NIST. Part of the Department of Commerce, NIST has completed and launched a new AI Risk Management Framework. ^^ We offer four concrete suggestions to implement and build upon this Framework, including commitments Microsoft is making in response to a recent White House meeting with leading AI companies. We also believe the Administration and other governments can accelerate momentum through procurement rules based on this Framework. Risk ManagementImplement NIST's AI Risk Management Framework_f5b1e812-fcad-11ed-b89a-14061883ea001.1First, Microsoft is committing to the White House, in response to its recent meeting, that we will implement NIST’s AI Risk Management Framework. Microsoft’s internal Responsible AI Standard is closely aligned with the Framework already, and we will now work over the summer to implement it so that all our AI services benefit from it.Testing & EngineeringAugment Microsoft's AI testing work with new steps to further strengthen our engineering practices relating to high-risk AI systems_f5b1ea2e-fcad-11ed-b89a-14061883ea001.2Second, we are similarly committing that we will augment Microsoft’s existing AI testing work with new steps to further strengthen our engineering practices relating to high-risk AI systems. ^^ Under Microsoft’s Responsible AI Standard, our AI engineering teams already work to identify potential harms, measure their propensity to occur, and build mitigations to address them. We have further developed red teaming techniques using multidisciplinary teams, which were originally developed to identify cybersecurity vulnerabilities, to stress test AI systems with a wide range of expertise, including privacy, security, and fairness. ^^ For high-risk systems, Microsoft is committing that red teaming is conducted before deployment by qualified experts who are independent of the product teams building those systems, adopting a best practice from the financial services industry. We will rely upon these red teams, together with our product teams who are responsible for systematic evaluations of the products that they build, to help us identify, measure, and mitigate potential harms. ^^ In addition to continually monitoring, tracking, and evaluating our AI systems, we will use metrics to measure and understand systemic issues specific to generative AI experiences, such as the extent to which a model’s output is supported by information contained in input sources. (We are releasing the first of these metrics this week as part of our Azure OpenAI Service at Build, our annual developer conference.)Self-AttestationRequire vendors of critical AI systems to the U.S. Government to self-attest that they are implementing NIST's AI Risk Management Framework_f5b1ebd2-fcad-11ed-b89a-14061883ea001.3Third, we believe the Administration can accelerate momentum through an Executive Order that requires vendors of critical AI systems to the U.S. Government to self-attest that they are implementing NIST's AI Risk Management Framework. ^^ It's important for governments to move faster, using both carrots and sticks. In the United States, federal procurement mechanisms have repeatedly demonstrated their value in improving the quality of products and advancing industry practice more generally. Building on similar approaches used for key technology priorities like cybersecurity, the U.S. Government could insert requirements related to the AI Risk Management Framework into the federal procurement process for AI systems. ^^ As a starting point, we believe it makes sense to scope such procurement requirements to focus on critical decision systems, meaning AI systems that have the potential to meaningfully impact the public’s rights, opportunities, or access to critical resources or services. This would align with the approach set out in the Blueprint for an AI Bill of Rights, released last year by the White House’s Office of Science and Technology Policy.StandardsWork with leaders in industry leaders and government to develop standards relating to foundation models_f5b1ee5c-fcad-11ed-b89a-14061883ea001.4Finally, we are committed to working with other industry leaders and those in government to develop new and additional standards relating to highly capable foundation models. ~ We recognize that the pace of AI advances raises new questions and issues related to safety and security, and we are committed to working with others to develop actionable standards to help evaluate and address them. Already, leaders at OpenAI, Google, Anthropic, and other AI companies have advanced important ideas that will help provide a foundation for future progress. We look forward to working with them and many others as these types of efforts move forward.Critical InfrastructureRequire effective safety brakes for AI systems that control critical infrastructure_f5b1f168-fcad-11ed-b89a-14061883ea002Second, require effective safety brakes for AI systems that control critical infrastructure. ~ In some quarters, thoughtful individuals increasingly are asking whether we can satisfactorily control AI as it becomes more powerful. Concerns are sometimes posed regarding AI control of critical infrastructure like the electrical grid, water system, and city traffic flows. ^^ This is the right time to discuss this question. This blueprint proposes new safety requirements that in effect would create safety brakes for AI systems that control the operation of designated critical infrastructure. These fail-safe systems would be part of a comprehensive approach to system safety that would keep effective human oversight, resilience, and robustness top of mind. In spirit, they would be similar to the braking systems engineers have long built into other technologies such as elevators, school buses, and high-speed trains, to safely manage not just everyday scenarios, but emergencies as well. ^^ In this approach, the government would define the class of high-risk AI systems that control critical infrastructure and warrant such safety measures as part of a comprehensive approach to system management. New laws would require operators of these systems to build safety brakes into high-risk AI systems by design. The government would then ensure that operators test highrisk systems regularly to ensure that the system safety measures are effective. And AI systems that control the operation of designated critical infrastructure would be deployed only in licensed AI datacenters that would ensure a second layer of protection through the ability to apply these safety brakes, thereby ensuring effective human control. High-Risk AI SystemsDefine the class of high-risk AI systems that are being deployed to control critical infrastructure and warrant safety brakes_f5b1f33e-fcad-11ed-b89a-14061883ea002.1First, the government should define the class of high-risk AI systems that are being deployed to control critical infrastructure and warrant safety brakes as part of a comprehensive approach to system safety. ^^ In the United States, the Secretary of Homeland Security is responsible for identifying and prioritizing critical infrastructure in coordination with other government agencies. Most notably, this includes the Cybersecurity and Infrastructure Security Agency, or CISA, which has identified 16 critical infrastructure sectors, including the communications sector, the emergency services sector, and the energy sector, to name a few. ^^ For the purposes of applying the safety brake concept to AI systems, we need to focus on the AI systems that are used to control the operation of critical infrastructure. There will be many AI systems used within critical infrastructure sectors that are low risk and that do not require the same depth of safety measures -- employee productivity tools and customer service agents are two such examples. ^^ Instead, one should focus on highly capable systems, increasingly autonomous systems, and systems that cross the digital-physical divide. For the purposes of spurring further discussion, one place to start might be to focus on AI systems that: ^^ • Take decisions or actions affecting large-scale networked systems; ^ • Process or direct physical inputs and outputs; ^ • Operate autonomously or semi-autonomously; and ^ • Pose a significant potential risk of large-scale harm, including physical, economic, or environmental harm.Safety BrakesRequire system developers to ensure that safety brakes are built by design into the use of AI systems for the control of critical infrastructure_f5b1f5f0-fcad-11ed-b89a-14061883ea002.2Second, the government should require system developers to ensure that safety brakes are built by design into the use of AI systems for the control of critical infrastructure. ^^ System safety is a well-established discipline that we have put to work in the aviation, automotive, and nuclear sectors, among others, and it is one that we must bring to bear to the engineering of AI systems that control critical infrastructure. We should establish a layered approach to AI safety, with the "safety brake" concept implemented at multiple levels. ^^ While the implementation of "safety brakes" will vary across different systems, a core design principle in all cases is that the system should possess the ability to detect and avoid unintended consequences, and it must have the ability to disengage or deactivate in the event that it demonstrates unintended behavior. It should also embody best practice in human-computer interaction design.Human ControlEnsure operators test and monitor high-risk systems to ensure that AI-systems that power critical infrastructure remain within human control_f5b1f83e-fcad-11ed-b89a-14061883ea002.3Third, the government should ensure operators test and monitor high-risk systems to ensure that AI-systems that power critical infrastructure remain within human control. ^^ Specific system testing will be needed in the context of a planned deployment for critical infrastructure. In other words, the use of an advanced AI model must be reviewed in the context of how it will be used in a specific product or service. ^^ In accordance with system safety best practices, the system and each of its components should be tested, verified, and validated rigorously. It should be provable that the system operates in a way that allows humans to remain in control at all times. In practice, we anticipate that this will require close and regular coordination between a system operator, their AI infrastructure provider, and their regulatory oversight bodies.Critical InfrastructureEnsure that AI systems controlling critical infrastructure are deployed in licensed AI infrastructure_f5b1fb2c-fcad-11ed-b89a-14061883ea002.4Fourth, AI systems that control the operation of designated critical infrastructure should be deployed only in licensed AI infrastructure. ^^ We believe it would be wise to require that AI systems that control the operations of higher-risk critical infrastructure systems be deployed on licensed AI infrastructure. This is not to suggest that the AI infrastructure needs to be a hyperscale cloud provider such as Microsoft. Critical infrastructure operators might build AI infrastructure and qualify for such a license in their own right. But to obtain such a license, the AI infrastructure operator should be required to design and operate their system to allow another intervention point -- in effect a second and separate layer of protection -- for ensuring human control in the event that application-level measures fail. ^^ These proposals might leave some wondering how realistic or futureproof "safety brakes" are if we are on a path to developing AI systems that are more capable than humans. They might ask: couldn't the AI system itself work around safety brakes and override them? Won't the AI system know how humans will respond at every step of the way and simply work around those responses? ^^ In posing those questions, it’s important to be clear about the facts as they stand today. Today's cutting-edge AI systems like GPT-4 from OpenAI and Claude from Anthropic have been specifically tested -- by qualified third-party experts from the Alignment Research Center -- for dangerous capabilities, such as the ability to evade human oversight and become hard to shut down. Those tests concluded that GPT-4 and Claude do not have sufficient capabilities to do those things today. ^^ This rigorous testing and the conclusions drawn provide us with clarity as to the capabilities of today’s cutting-edge AI models. But we should also heed the Alignment Research Center’s call for ongoing research on these topics and recognize the need for industry-wide commitment to AI capability evaluations. Put simply, we need to ensure that we have the right structures in place not only to understand the status quo, but to get ahead of the future. That is precisely why we need action with respect to the small but important class of highly capable AI models that are on the frontier -- a topic that our next section addresses.Laws & RegulationsDevelop a broad legal and regulatory framework based on the technology architecture for AI_f5b1fdfc-fcad-11ed-b89a-14061883ea003Third, develop a broad legal and regulatory framework based on the technology architecture for AI. ~ We believe there will need to be a legal and regulatory architecture for AI that reflects the technology architecture for AI itself. In short, the law will need to place various regulatory responsibilities upon different actors based upon their role in managing different aspects of AI technology. ^^ For this reason, this blueprint includes information about some of the critical pieces that go into building and using new generative AI models. Using this as context, it proposes that different laws place specific regulatory responsibilities on the organizations exercising certain responsibilities at three layers of the technology stack: the applications layer, the model layer, and the infrastructure layer. ^^ This should first apply existing legal protections at the applications layer to the use of AI. This is the layer where the safety and rights of people will most be impacted, especially because the impact of AI can vary markedly in different technology scenarios. In many areas, we don’t need new laws and regulations. We instead need to apply and enforce existing laws and regulations, helping agencies and courts develop the expertise needed to adapt to new AI scenarios. ^^ There will then be a need to develop new law and regulations for highly capable AI foundation models, best implemented by a new government agency. This will impact two layers of the technology stack. The first will require new regulations and licensing for these models themselves. And the second will involve obligations for the AI infrastructure operators on which these models are developed and deployed. The blueprint that follows offers suggested goals and approaches for each of these layers. In doing so, this blueprint builds in part on a principle developed in recent decades in banking to protect against money laundering and criminal or terrorist use of financial services. The "Know Your Customer" -- or KYC -- principle requires that financial institutions verify customer identities, establish risk profiles, and monitor transactions to help detect suspicious activity. It would make sense to take this principle and apply a KY3C approach that creates in the AI context certain obligations to know one’s cloud, one’s customers, and one’s content. ^^ In the first instance, the developers of designated, powerful AI models first “know the cloud” on which their models are developed and deployed. In addition, such as for scenarios that involve sensitive uses, the company that has a direct relationship with a customer -- whether it be the model developer, application provider, or cloud operator on which the model is operating -- should "know the customers" that are accessing it. ^^ Also, the public should be empowered to "know the content" that AI is creating through the use of a label or other mark informing people when something like a video or audio file has been produced by an AI model rather than a human being. This labeling obligation should also protect the public from the alteration of original content and the creation of "deep fakes." This will require the development of new laws, and there will be many important questions and details to address. But the health of democracy and future of civic discourse will benefit from thoughtful measures to deter the use of new technology to deceive or defraud the public.Best PracticesHelp our customers apply state-of-the-art best practices to deploy AI lawfully and responsibly_f5b20072-fcad-11ed-b89a-14061883ea003.1First, we will work with our customers to help them apply state-of-the-art best practices to deploy AI lawfully and responsibly. ^^ One of the critical characteristics of AI is that the real-world impact on specific groups and issues is defined not just by the developer of an AI model or system, but also in its implementation in a specific service or application. In fact, in many circumstances it is only at the application level that it’s possible to specifically identify and test for these real-world impacts before AI is deployed. As a result, responsibilities are often shared or even distributed, with different organizations needing to play different roles. ^^ This helps explain why it's so important for customers that use AI in their services to develop their own capabilities to do so responsibly. This also explains why it is so important for a leading tech company to share information and lend their expertise on state-of-the-art best practices and tooling for responsible AI deployment. ^^ We have been doing this type of work for two decades on other issues involving digital technology, including to implement legal compliance systems, advance cybersecurity, and protect privacy. We began five years ago to do similar work relating to artificial intelligence, and we will expand this initiative to work more broadly and deeply with our customers in the year ahead.RegulationIncrease the AI expertise and capabilities of regulatory agencies_f5b2025c-fcad-11ed-b89a-14061883ea003.2Regulatory AgenciesFood and Drug AdministrationFederal Aviation AdministrationSecond, we believe that regulatory agencies will need to add new AI expertise and capabilities. ^^ Very quickly, this need will reach virtually every agency in most governments in the world. For example, an agency like the Food and Drug Administration will need more AI experts who can help evaluate the use of cutting-edge AI systems by companies in something like the clinical trials for new drugs. Similarly, agencies like the Federal Aviation Administration will need additional AI experts to help evaluate the new uses of AI by aircraft manufacturers in developing new planes. ^^ Generative AI itself will be a powerful tool that will better enable regulatory agencies to evaluate the use of AI. This is because models like GPT-4 and services like ChatGPT, GitHub Copilot, and Microsoft M365 Copilot make it far easier for people to harness the power of AI to access data and evaluate it more quickly. As Google rightly recommended in a new white paper just last week, it will be important for governments to "direct sectoral regulators to update existing oversight and enforcement regimes to apply to AI systems, including on how existing authorities apply to the use of AI." Agencies will need the funding, staff, and commitment to put these new tools to work.EducationMake information about AI technologies and responsible AI practices available to legislators, judges, and lawyers_f5b20536-fcad-11ed-b89a-14061883ea003.3LegislatorsJudgesLawyersThird, we will support broad educational initiatives to make information about AI technologies and responsible AI practices available to legislators, judges, and lawyers. ^^ Finally, rapid AI advances are creating new pressures on those who make or help enforce the law to learn about new AI technologies and how they work. We witnessed a similar need when the personal computer first became popular in the 1980s. For example, judges needed to decide cases that started to turn in part on evidence about or involving PC software and hardware. ^^ Beginning in the 1990s, Microsoft supported broad initiatives to share information about how this new technology worked. We continue to do this today in selected areas such as electronic discovery. The accelerating use of AI means that new such efforts will be needed. We will support this work, including by supporting bar associations and other public interest and civic groups and activities.Laws & RegulationsDevelop new laws and regulations for highly capable AI foundation models_f5b20798-fcad-11ed-b89a-14061883ea003.4Developing new laws and regulations for highly capable AI foundation models ^^ While existing laws and regulations can be applied and built upon for the application layer of the tech stack, we believe that new approaches will be needed for the two additional layers beneath that reflect the new and more powerful AI models that are emerging. The first of these is for the development of the most powerful new AI models, and the second is for the deployment and use of these models in advanced datacenters. ^^ From our work on the frontiers of AI, we have seen a new class of model emerge. Highly capable foundation models are trained on internet-scale datasets and are effective outof-the-box at new tasks—a model like GPT-4 allows you to create a never-seen-before image using words in one prompt, and a speech in the style of Franklin Roosevelt in the very next. ^^ At the cutting-edge, the capabilities of these foundation models are at once very impressive and can be harder to predict. As the models have been scaled up, we have seen anticipated advances in capabilities, as well as surprising ones that we and others did not predict ahead of time and could not observe on a smaller scale. Despite rigorous prerelease testing and engineering, we’ve sometimes only learned about the outer bounds of model capabilities through controlled releases with users. And the work needed to harness the power of these models and align them to the law and societal values is complex and evolving. ^^ These characteristics of highly capable models present risk surfaces that need to be addressed. To date, we have benefited from the high safety standards self-imposed by the U.S. developers who have been working at the frontiers of AI model development. But we shouldn’t leave these issues of societal importance to good judgment and self-restraint alone. We need regulatory frameworks that anticipate and get ahead of the risks. And we need to acknowledge the simple truth that not all actors are wellintentioned or well-equipped to address the challenges that highly capable models present. Some actors will use AI as a weapon, not a tool, and others will underestimate the safety challenges that lie ahead. ^^ Last week, Sam Altman, the CEO of OpenAI, testified before Congress and called for the establishment of a licensing regime for this small but important class of highly capable models at the frontiers of research and development. As Microsoft, we endorse that call and support the establishment of a new regulator to bring this licensing regime to life and oversee its implementation.KnowledgeShare specialized knowledge about advanced AI models to help governments define the regulatory threshold_f5b20996-fcad-11ed-b89a-14061883ea003.4.1First, we and other leading AI developers will need to share our specialized knowledge about advanced AI models to help governments define the regulatory threshold. ^^ One of the initial challenges will be to define which AI models should be subject to this level of regulation. The objective is not to regulate the rich ecosystem of AI models that exists today and should be supported into the future, but rather the small number of AI models that are very advanced in their capabilities and in some cases, redefining the frontier. We refer to this small subset of models as highly capable AI models in this white paper. ^^ Defining the appropriate threshold for what constitutes a highly capable AI model will require substantial thought, discussion, and work in the months ahead. The amount of compute used to train a model is one tractable proxy for model capabilities, but we know today that it is imperfect in several ways and unlikely to be durable into the future, especially as algorithmic improvements lead to compute efficiencies or new architectures altogether. ^^ A more durable but unquestionably more complex proposition would be to define the capabilities that are indicative of high ability in areas that are consequential to safety and security, or that represent new breakthroughs that we need to better understand before proceeding further. Further research and discussion are needed to set such a capability-based threshold, and early efforts to define such capabilities must continue apace. In the meantime, it may be that as with many complex problems in life, we start with the best option on offer today—a compute-based threshold—and commit to a program of work to evolve it into a capability-based threshold in short order. LicensingDefine the requirements to obtain a license to develop or deploy highly capable AI models_f5b20cc0-fcad-11ed-b89a-14061883ea003.4.2Second, we will support governments in their efforts to define the requirements that must be met in order to obtain a license to develop or deploy a highly capable AI model. ^^ A licensing regime for highly capable AI models should be designed to fulfill three key goals. First and foremost, it must ensure that safety and security objectives are achieved in the development and deployment of highly capable AI models. Second, it must establish a framework for close coordination and information flows between licensees and their regulator, to ensure that developments material to the achievement of safety and security objectives are shared and acted on in a timely fashion. Third, it must provide a footing for international cooperation between countries with shared safety and security goals, as domestic initiatives alone will not be sufficient to secure the beneficial uses of highly capable AI models and guard against their misuse. We need to proceed with an understanding that it is currently trivial to move model weights across borders, allowing those with access to the “crown jewels” of highly capable AI models to move those models from country to country with ease. ^^ To achieve safety and security objectives, we envision licensing requirements such as advance notification of large training runs, comprehensive risk assessments focused on identifying dangerous or breakthrough capabilities, extensive prerelease testing by internal and external experts, and multiple checkpoints along the way. Deployments of models will need to be controlled based on the assessed level of risk and evaluations of how wellplaced users, regulators, and other stakeholders are to manage residual risks. Ongoing monitoring post-release will be essential to ensuring that guardrails are functioning as intended and that deployed models remain under human control at all times. ^^ In practice, we believe that the effective enforcement of such a regime will require us to go one layer deeper in the tech stack to the AI datacenters on which highly capable AI models are developed and deployed. DatacentersImpose licensing requirements on the operators of AI datacenters that are used for the testing or deployment of these models_f5b20f54-fcad-11ed-b89a-14061883ea003.4.3AI DatacentersThird, we will support government efforts to ensure the effective enforcement of a licensing regime for highly capable AI models by also imposing licensing requirements on the operators of AI datacenters that are used for the testing or deployment of these models. ^^ Today’s highly capable AI models are built on advanced AI datacenters. They require huge amounts of computing power, specialized AI chips, and sophisticated infrastructure engineering, like Microsoft’s facilities in Iowa, described above. Such AI datacenters are therefore critical enablers of today’s highly capable AI models and an effective control point in a comprehensive regulatory regime. ^^ Much like the regulatory model for telecommunications network operators and critical infrastructure providers, we see a role for licensing providers of AI datacenters to ensure that they play their role responsibly and effectively to ensure the safe and secure development and deployment of highly capable AI models. To obtain a license, an AI datacenter operator would need to satisfy certain technical capabilities around cybersecurity, physical security, safety architecture, and potentially export control compliance. ^^ In effect, this would start to apply for AI a principle developed for banking to protect against money laundering and criminal or terrorist use of financial services. The “Know Your Customer”—or KYC— principle requires that financial institutions verify customer identities, establish risk profiles, and monitor transaction to help detect suspicious activity. ^^ In a similar way, it would make sense for a similar KYC principle to require that the developers of powerful AI models first “know the cloud” on which their models are deployed. The use of authorized and licensed AI datacenters would ensure that those who develop advanced models would have several vendors from which to choose. And it would enable the developer of an advanced model to build or operate their own cloud infrastructure as well, based on meeting the requisite technical standards and obligations. ^^ The licensed AI datacenter operator would then need to meet ongoing regulatory requirements, several of which are worth considering. ^^ First, operators of AI datacenters have a special role to play in securing highly capable AI models to protect them from malicious attacks and adversarial actors. This likely involves not just technical and organizational measures, but also an ongoing exchange of threat intelligence between the operator of the AI datacenter, the model developer, and a regulator. ^^ Second, in certain instances, such as for scenarios that involve sensitive uses, the cloud operator on which the model is operating should apply the second aspect of the KYC principle – knowing the customers who are accessing the model. More thought and discussion will be needed to work through the details, especially when it comes to determining who should be responsible for collecting and maintaining specific customer data in different scenarios. ^^ The operators of AI datacenters that have implemented know-your-customer procedures can help regulators get comfortable that all appropriate licenses for model development and deployment have been obtained. One possible approach is that substantial uses of compute that are consistent with large training runs should be reported to a regulator for further investigation. ^^ Third, as export control measures evolve, operators of AI datacenters could assist with the effective enforcement of those measures, including those that attach at the infrastructure and model layers of the tech stack. ^^ Fourth, as discussed above, the AI infrastructure operator will have a critical role and obligation in applying safety protocols and ensuring that effective AI safety brakes are in place for AI systems that manage or control critical infrastructure. It will be important for the infrastructure operator to have the capability to intervene as a second and separate layer of protection, ensuring the public that these AI systems remain under human control. ^^ These early ideas naturally will all need to be developed further, and we know that our colleagues at OpenAI have important forthcoming contributions on these topics too. What is clear to us now is that this multitiered licensing regime will only become more important as AI models on the frontiers become more capable, more autonomous, and more likely to bridge the digital-physical divide. As we discussed earlier, we believe there is good reason to plan and implement an effective licensing regime that will, among other things, help to ensure that we maintain control over our electricity grid and other safety-critical infrastructure when highly capable AI models are playing a central role in their operation.Transparency & AccessPromote transparency and ensure academic and nonprofit access to AI_f5b21166-fcad-11ed-b89a-14061883ea004AacademiaNonprofitsFourth, promote transparency and ensure academic and nonprofit access to AI. We believe a critical public goal is to advance transparency and broaden access to AI resources. While there are some important tensions between transparency and the need for security, there exist many opportunities to make AI systems more transparent in a responsible way. That’s why Microsoft is committing to an annual AI transparency report and other steps to expand transparency for our AI services. ^^ We also believe it is critical to expand access to AI resources for academic research and the nonprofit community. Basic research, especially at universities, has been of fundamental importance to the economic and strategic success of the United States since the 1940s. But unless academic researchers can obtain access to substantially more computing resources, there is a real risk that scientific and technological inquiry will suffer, including relating to AI itself. Our blueprint calls for new steps, including steps we will take across Microsoft, to address these priorities.ReportsRelease an annual transparency report to inform the public about our policies, systems, progress, and performance in managing AI responsibly and safely_f5b21486-fcad-11ed-b89a-14061883ea004.1First, Microsoft will release an annual transparency report to inform the public about its policies, systems, progress, and performance in managing AI responsibly and safely. ^^ Transparency reports have proven to be an effective measure to drive corporate accountability and help the public better understand the state-of-the-art and progress toward goals. Microsoft believes transparency reports have a role to play in the responsible AI context too, and so we will release an annual transparency report to inform the public about our policies, systems, progress, and performance in managing AI responsibly and safely. If adopted across the industry, transparency reports would be a helpful mechanism for recording the maturing practice of responsible AI and charting cross-industry progress.RegistryDevelop a national registry of high-risk AI systems_f5b21814-fcad-11ed-b89a-14061883ea004.2Second, Microsoft will support the development of a national registry of high-risk AI systems that is open for inspection so that members of the public can learn where and how those systems are in use. ^^ Public trust in AI systems can be enhanced by demystifying where and how they are in use. For high-risk AI systems, Microsoft supports the development of a national registry that would allow members of the public to review an overview of the system as deployed and the measures taken to ensure the safe and rights-respecting performance of the system. ^^ For this information to be useful to the public, it should be expressed at the system level, providing details about the context of use, and be written for nontechnical audiences. To achieve this, the United States could implement the approach of several European cities in adopting the Algorithmic Transparency Standard and developing accessible explanations of how it uses AI (see, for example, the City of Amsterdam’s Algorithm Register).NoticeEnsure AI systems inform the public when they are interacting with them_f5b21a58-fcad-11ed-b89a-14061883ea004.3Third, Microsoft will commit that it will continue to ensure that our AI systems are designed to inform the public when they are interacting with an AI system and that the system's capabilities and limitations are communicated clearly. ^^ We believe that transparency is important not only through broad reports and registries, but in specific scenarios and for the users of specific AI systems. Microsoft will continue to build AI systems designed to support informed decision making by the people who use them. We take a holistic approach to transparency, which includes not only user interface features that inform people that they are interacting with an AI system, but also educational materials, such as the new Bing primer, and detailed documentation of a system’s capabilities and limitations, such as the Azure OpenAI Service Transparency Note. This documentation and experience design elements are meant to help people understand an AI system’s intended uses and make informed decisions about their own use.LabellingRequire AI-generated content to be labeled_f5b21d6e-fcad-11ed-b89a-14061883ea004.4Fourth, we believe there is benefit in requiring AI-generated content to be labeled in important scenarios so that the public “knows the content” it is receiving. ^^ This is the third part of the KY3C approach we believe is worth considering. As we are committing above for Microsoft’s services Bing Image Creator and Designer, we believe the public deserves to “know the content” that AI is creating, informing people when something like a video or audio has been originally produced by an AI model rather than a human being. This labeling obligation should also inform people when certain categories of original content have been altered using AI, helping protect against the development and distribution of “deep fakes.” This will require the development of new laws, and there will be many important questions and details to address. But the health of democracy and future of civic discourse will benefit from thoughtful measures to deter the use of new technology to deceive or defraud the public.AI ResourcesProvide access to AI resources for academic research and the nonprofit community_f5b2202a-fcad-11ed-b89a-14061883ea004.5Academic Research CommunityNonprofit CommunityAccess to AI resources for academic research and the nonprofit community ^^ We believe there is another element that adds to transparency and that deserves more prominent attention. This is the need to provide broad access to AI resources for academic research and the nonprofit community. ^^ The high cost of computational resources for the training of large-scale AI models, as well as other AI projects, is understandably raising concerns in the higher education and nonprofit communities. We understand this issue well because Microsoft’s large technology investment in OpenAI in 2019 originated from precisely this need for OpenAI itself, due in part to its nonprofit status. ^^ Basic research, perhaps especially at universities, has been of fundamental importance to the economic and strategic success of the United States since the 1940s. Much of the tech sector itself owes both its birth and ongoing innovation to critical basic research pursued in colleges and universities across the country. It’s a success story that has been studied and emulated in many other countries around the world. The past few decades have seen huge swaths of basic research in almost every field propelled by growing computing resources and data science. Unless academic researchers can obtain access to substantially more computing resources, there is a real risk that scientific inquiry and technological innovation will suffer. ^^ Another dimension of this problem is also important. Academic researchers help ensure accountability to the public by advancing our understanding of AI. The public needs academics to pursue research in this area, including research that advances AI accountability by analyzing the behavior of the models the commercial sector is creating. While new and smaller open-source AI models are emerging and clearly are important, other basic research projects involving AI will almost certainly require more computational power than in the past. And unless new funding sources come together to provide a more centralized resource for the academic community, academic research will be at risk. This has led us to offer two focused commitments:ComputingProvide computing resources for academic research_f5b2226e-fcad-11ed-b89a-14061883ea004.5.1National AI Research Resource (NAIRR)First, Microsoft will support the establishment of the newly proposed National AI Research Resource (NAIRR) to provide computing resources for academic research and would welcome and support an extension to accommodate access by academic institutions in allied nations abroad, including Japan, the United Kingdom, the European Union, and other like-minded countries. ^^ The National AI Research Resource has its origins in the National Initiative AI Act of 2020, passed by Congress. The Act called on the National Science Foundation, in consultation with the White House Office of Science and Technology Policy, to create a task force to create a roadmap for “a shared research infrastructure that would provide AI researchers and students with significantly expanded access to computational resources, high-quality data, educational tools, and user support.” This January, the Task Force completed its work, publishing a final report calling for the creation and funding of a federated mix of computational and data resources, testbeds, software, and testing tools, based on a platform that can reduce the barriers to participation in the AI research ecosystem and increase the diversity of AI researchers. ^^ Microsoft supports the establishment of the National AI Research Resource and believes it to be of fundamental importance to the United States’ leadership in AI innovation and risk mitigation. We will collaborate with the National Science Foundation to explore participation in a pilot project to inform efforts to stand up the National AI Research Resource, including by facilitating independent academic research relating to the safety of AI systems. ^^ We also would welcome and support an extension of the NAIRR to provide access by academic institutions in like-minded nations. Already we’re seeing similar and substantial interest in these other countries. For example, Japan’s recent “National Strategy in the New Era of AI” calls for work to expand the computing resources for public and private use. We believe that a multilateral AI research resource would accelerate existing efforts to establish global norms and interoperable approaches to risk mitigation, including those underway in the U.S.-EU Trade and Technology Council and the G7.InvestmentIncrease investment in academic research programs_f5b22606-fcad-11ed-b89a-14061883ea004.5.2Academic Research ProgramsSecond, we will increase investment in academic research programs to ensure researchers outside Microsoft can access the company's foundation models and the Azure OpenAI Service to undertake research and validate findings. ^^ This expanded commitment builds on the success of our Turing Academic Program and Accelerating Foundation Models Research Program. It is designed to help the academic community gain API-based access to cuttingedge foundation models from Microsoft, as well as OpenAI models via Microsoft’s Azure OpenAI Service. This will ensure that researchers can study frontier applications and the sociotechnical implications of these models. Microsoft will ensure that its program design accommodates API-based access by a diverse community of academic researchers, including researchers at Minority Serving Institutions across the United States. ^^ An important complement to providing such access is the development of governance best practices for the academic community engaged in frontier research on applications and the safety and security implications of highly capable models. Microsoft would welcome the opportunity to develop such practices by supporting and collaborating with a multistakeholder group, including representatives across the academic community.NonfprofitsCreate free and low-cost AI resources for use by the nonprofit community_f5b228ea-fcad-11ed-b89a-14061883ea004.5.3Nonprofit CommunityThird, Microsoft will create free and low-cost AI resources for use by the nonprofit community. ^^ Finally, we deeply appreciate the critical role that nonprofit organizations play in addressing societal needs around the world. Given their role as great incubators of innovative solutions, we believe it is critical for nonprofits to have broad, easy, and inexpensive access to new AI models and features for their work. Microsoft Philanthropies, including its Tech for Social Responsibility arm, supports 350,000 nonprofits in the Microsoft Cloud. It provides more than $4 billion annually in cash and technology donations and discounts to nonprofits worldwide, a figure comparable to one of the 10 largest government foreign aid budgets. ^^ Last week we expanded this support by announcing AI solutions to Microsoft Cloud for Nonprofit. These AI solutions are designed to improve the ability of nonprofit organizations to optimize operations, engage with donors, and manage campaigns. This is the first of several steps we will take to reduce technical and cost barriers and enable nonprofits to harness the latest advances in AI. PartnershipsPursue public-private partnerships to use AI as an effective tool to address the inevitable societal challenges that come with new technology_f5b22b4c-fcad-11ed-b89a-14061883ea005Public SectorFinally, we believe there is enormous opportunity to bring the public and private sectors together to use AI as a tool to improve the world, including by countering the challenges that technological change inevitably creates. We are cleareyed about the future and realize that some will seek to use AI as a weapon rather than a tool. And even when people of goodwill do their best, technological change inevitably creates unforeseen bumps in the road ahead.Private SectorDemocratic SocietiesBut we've also learned from numerous efforts over the years what democratic societies can accomplish when they harness the power of technology and bring the public and private sectors together. Two examples are perhaps the most profound.Christchurch, New ZealandThe first is the Christchurch Call, born from the tragic terrorist tragedy that took place in Christchurch, New Zealand, on March 15, 2019. The attack claimed the lives of 51 innocent Muslims at two mosques and was livestreamed worldwide. The internet provided a stage not only to broadcast the attack but perhaps provided an incentive to pursue the assault in the first place.Jacinda ArdernNew Zealand Prime Minister Jacinda Ardern vowed that the world would learn from the attack and take steps to prevent technology from being used this way again.Emmanuel MacronPartnering with French President Emmanuel Macron, she brought leading tech companies together to pursue concrete steps to prevent the livestreaming and internet distribution of similar violent attacks in the future. Exactly two months after the attack, on May 15, 2019, government and tech leaders met at the Elysée in Paris to sign the Christchurch Call and commited to collective action that has continued in the four years that have followed.UkraineThis work provided inspiration for the larger assault that began when the Russian military unleashed waves of cyberattacks on Ukraine on February 23, 2022. As we noted last year, this reflected an age-old lesson from history: countries wage wars using the latest technology, and the wars themselves accelerate technological change. ^^ But the role of technology in the war in Ukraine has brought a new dimension to the defense not only of Ukraine, but of democracy itself. The war has required a new form of collective defense. It has pitted Russia, a major cyberpower, not just against an alliance of countries, but also against a coalition of tech companies and NGOs.Tech SectorAcross the tech sector, companies have stepped up to support Ukraine's remarkable tenacity and innovation. Individual and collective technology measures have sustained Ukraine's digital operations, defeated cyberattacks, documented war crimes, and enabled students to stay in school even when their schools are damaged or destroyed. Microsoft has now provided $450 million of financial and technology assistance to Ukraine, an amount that is unprecedented in the company’s history.The lessons from the Christchurch Call and the war in Ukraine should guide us on the role of AI in the future. One key is to focus on specific problems that can benefit from new initiatives and concrete action. Another is to bring governments, companies, and NGOs together on an international basis not only to move faster, but to accomplish more than any single organization or even country can achieve on its own. Microsoft is committed to pursuing and supporting similar initiatives in the months ahead.Fifth, pursue new public-private partnerships to use AI as an effective tool to address the inevitable societal challenges that come with new technology. ~ One lesson from recent years is what democratic societies can accomplish when they harness the power of technology and bring the public and private sectors together. It’s a lesson we need to build upon to address the impact of AI on society. ^^ We will all benefit from a strong dose of clear-eyed optimism. AI is an extraordinary tool. But like other technologies, it too can become a powerful weapon, and there will be some around the world who will seek to use it that way. But we should take some heart from the cyber front and last year and a half in the war in Ukraine. What we found is that when the public and private sectors work together, when like-minded allies come together, and when we develop technology and use it as a shield, it’s more powerful than any sword on the planet. ^^ Important work is needed now to use AI to protect democracy and fundamental rights, provide broad access to the AI skills that will promote inclusive growth, and use the power of AI to advance the planet’s sustainability needs. Perhaps more than anything, a wave of new AI technology provides an occasion for thinking big and acting boldly. In each area, the key to success will be to develop concrete initiatives and bring governments, respected companies, and energetic NGOs together to advance them. We offer some initial ideas in this report, and we look forward to doing much more in the months and years ahead. Problems & ActionsFocus on specific problems that can benefit from new initiatives and concrete action_f5b22eee-fcad-11ed-b89a-14061883ea005.1The lessons from the Christchurch Call and the war in Ukraine should guide us on the role of AI in the future. One key is to focus on specific problems that can benefit from new initiatives and concrete action.International CollaborationBring governments, companies, and NGOs together on an international basis_f5b231dc-fcad-11ed-b89a-14061883ea005.2Another is to bring governments, companies, and NGOs together on an international basis not only to move faster, but to accomplish more than any single organization or even country can achieve on its own. Microsoft is committed to pursuing and supporting similar initiatives in the months ahead.Democracy & FakeryAddress risks to democracy and the public from the potential weaponization of AI to alter content and create "deep fakes"_f5b23452-fcad-11ed-b89a-14061883ea005.3Coalition for Content Provenance and AuthenticityAs we do so, it will be important to start with important building blocks that exist already. One of the most important is the Coalition for Content Provenance and Authenticity, or C2PA. Co-founded by companies such as Adobe, the BBC, Intel, Microsoft, Sony, and Truepic, C2PA unifies the efforts of the Adobe-led Content Authenticity Initiative (CAI), which focuses on systems to provide context and history for digital media, and Project Origin, a Microsoft- and BBC-led initiative that tackles disinformation in the digital news ecosystem.Eric HorvitzAs Microsoft's Chief Scientific Officer, Eric Horvitz, said last year, success will require "a multi-pronged approach, including education aimed at media literacy, awareness, and vigilance, [with] investments in quality journalism." There will be opportunities in the coming months to take important steps together.In recent years, there has been a growing focus on addressing the new risks to democracy and the public from the potential weaponization of AI to alter content and create "deep fakes," including videos. The concern about future technology is wellplaced (although we are concerned that countries are doing too little to address foreign cyber influence operations that are prolific and impactful already). In short, we will all need to do more collectively to combat this type of threat.ProvenanceDeploy new state-of-the-art provenance tools to help the public identify AI-generated audio-visual content and understand its origin_f5b237f4-fcad-11ed-b89a-14061883ea005.3.1This week Microsoft will deploy new state-of-the-art provenance tools to help the public identify AI-generated audio-visual content and understand its origin. At Build, our annual developer conference, we are announcing the development of a new media provenance service. The service will mark and sign AI-generated videos and images with metadata about their origin, enabling users to verify that a piece of content was generated by AI. The service implements the C2PA specification. Microsoft will initially support major image and video formats and release the service for use with two of Microsoft’s new AI products, Microsoft Designer and Bing Image Creator. ^^ This is an important step, but just a single one. Fortunately, many others are moving forward with similar and critical measures. We will need the right combination of focused steps and broader initiatives.Activity & ScopeThink big and act boldly_f5b23ace-fcad-11ed-b89a-14061883ea005.3.2Perhaps more than anything, a wave of new AI technology provides an occasion for thinking big and acting boldly. Important work is needed to use AI to protect democracy and fundamental rights, provide broad access to the AI skills that will promote inclusive growth, and use the power of AI to advance the planet’s sustainability needs. In each area, the key to success will be to bring governments, respected companies, and energetic NGOs together. There will be no shortage of opportunities or challenges. We need to seize the moment.2023-05-27https://query.prod.cms.rt.microsoft.com/cms/api/am/binary/RW14GtwOwenAmburOwen.Ambur@verizon.net