Presentation (2018) on Aether Committee, with slide displaying initial set of Aether’s working groups.

Aether Committee

T

he Aether committee was created to bring together top talent in AI principles and technology, ethics, law, and policy from across Microsoft to formulate recommendations on policies, processes, and best practices for the responsible development and fielding of AI technologies. Aether works closely with Microsoft’s Office of Responsible AI, a team that the Aether Committee had helped to form in 2019.

Aether was co-founded by Technical Fellow and Microsoft Research Director, Eric Horvitz, and company President, Brad Smith, in 2016 and has been chaired by Eric Horvitz. The name is an acronym for “AI and ethics in engineering and research.” Aether and its expert working groups have played an influential role in formulating Microsoft's approach to the responsible development and deployment of AI technologies. Efforts included the development and refinement of Microsoft’s AI principles and the company’s Sensitive Uses review program.

Early Aether Committee meeting (2018).

Aether efforts have included deliberations of its main committee and its working groups, co-chaired by experts in the respective area. The working groups have continued to play a key role in investigating frontier issues, developing points of view, creating tools, defining best practices, and providing tailored implementation guidance related to their respective areas of expertise. Learnings from the working groups and main committee have played an influential role in Microsoft’s responsible AI programs and policies. Reviews and guidance by the Sensitive Uses committee has led to the company either declining or placing limits on specific customer engagements where AI-related risks were high.

In 2019, the Aether Committee helped to spawn the Office of Responsible AI (ORA), now led by former Aether Chief Counsel, Natasha Crampton. Several functions of the Aether Committee were shifted and further formalized within ORA, including the Sensitive Uses review program that Aether had established earlier. The ORA team, working closely with Aether experts, authored Microsoft’s Responsible AI Standard. The Responsible AI Standard is structured in terms of Microsoft’s AI principles and extends the top-level definitions of the principles into concrete actions and goals required to be followed when Microsoft teams develop and deploy AI applications. ORA oversees the Sensitive Uses program, the Responsible AI Standard, and efforts and engagements on policy.

Microsoft’s six AI principles.

Microsoft’s Responsible AI Standard provides concrete details on deliberation and actions required for deploying AI technologies, structured in accordance with Microsoft’s AI principles.

In the first year of the Aether Committee (2017), the Aether Sensitive Uses working group proposed a set of criteria to be used to determine if an AI application is to be considered “sensitive” and should thus be routed before deployment into a special process involving careful review, guidance, and potential gating by a special Sensitive Uses review committee. The definition of sensitive uses of AI has been remarkably stable over the years since the considerations were first proposed. Aether stood up a program and process referred to as the Sensitive Uses Review, now overseen and operated by the Office of Responsible AI. The Sensitive Uses review committee has members with expertise spanning multiple areas. The committee considers potential costly behaviors, uses, and impact of AI applications under review. Discussion includes a deep dive into the case at hand, consideration of precedent developed in discussions and decisions on earlier cases, and the development of guidance for the product team as well as the company.

When Sensitive Uses was operated as part of the Aether Committee, challenging and novel cases considered by the Sensitive Uses working group were subsequently presented to the main Aether Committee for committee-wide discussion comment and refinement. Where standing questions and new considerations were identified, the Aether committee engaged with Microsoft’s Senior Leadership team for input, sharing case summaries and results of the discussions from the Sensitive Uses committee and main Aether Committee, often with the inclusion of multiple options and associated assessed tradeoffs.

Criteria for categorizing an AI application as a sensitive and requiring special study and deliberation.

New proposed AI products or services at Microsoft are considered sensitive uses of AI if their use or misuse has a significant probability of (1) causing significant physical or psychological injury to an individual; (2) of having a consequential negative impact on an individual’s legal position or life opportunities; that is, the use or misuse of the AI system could affect an individual’s legal status, such as whether an individual is recognized as a minor, adult, parent, guardian, or person with a disability, as well as their marital, immigration, and citizenship status, and could affect their legal rights (e.g., in the context of the criminal justice system), or could influence their ability to gain access to credit, education, employment, healthcare, housing, insurance, and social welfare benefits, services, or opportunities, or the terms on which they are provided; or (3) restricting, infringing upon, or undermining the ability to realize an individual’s human rights, including the following:

· Human dignity and equality in enjoyment of rights.

· Freedom from discrimination.

· Life, liberty, and security of the person.

· Equal protection of the law and criminal justice systems.

· Protection against arbitrary interference with privacy.

· Freedom of movement.

· Freedom of thought, conscience, and religion.

· Freedom of opinion and expression.

· Peaceful assembly and association.

The three criteria have provided a strong foundation over the years and form the basis for routing projects into a Sensitive Uses review, now overseen by the Office of Responsible AI.

Today, with the Office of Responsible AI taking on the oversight and management of the Sensitive Uses review program, the Responsible AI Standard and the company’s adherence to its requirements, and external engagements on policy, Aether and its working groups have shifted to focus on its advisory role on frontier topics and rising challenges. The Aether Committee focuses on providing guidance to senior leadership, the Office of Responsible AI, and our responsible AI engineering teams on rising questions, challenges, and opportunities in the development and deployment of AI technologies. Experts and program management on Aether collaborate closely with the Office of Responsible AI and both teams coordinate with engineering efforts and sales teams to help them uphold Microsoft’s AI principles in their day-to-day work and with projects involving AI products and services.

Aether has continued to organize special deep-dive offsites and studies, bringing together teams of experts on specific challenges, opportunities, and directions with the responsible development and deployment of AI systems. These special studies include Aether strategic focus projects on emerging topics and issues in the responsible deployment of AI technologies. For example, the 2019 Aether Media Provenance (AMP) strategic focus project pursued technical methods for certifying the integrity and authenticity of media in a world of rising AI-generated abusive content. The project led to the development of core methods for media provenance. The project resulted in the co-founding of Project Origin with the BBC, CBC, and NYTimes, and later the Coalition for Content Provenance and Authenticity (C2PA) and today’s Content Credentials standard. C2PA now includes over 100 companies, and supports a broad ecosystem of over 2,000 organizations, including manufacturers of cameras, content producers, technology companies, and non-profit organizations. 

Aether Media Provenance (AMP) contributors strategic focus team on visit to BBC headquarters that resulted in the creation of Project Origin (2019), a collaboration between BBC, CBC, Microsoft, and the New York Times. Left to right: Rico Malvar, Eric Horvitz, Paul England, Cedric Fournet.

Aether also continues to also organize and lead cross-company studies, referred to as Laser studies. With Laser studies, cross-team groups of experts are assembled and work for weeks to months on deep-dive analyses and guidance on emerging, potentially disruptive AI technologies, including the rise of new fundamental capabilities or new, potentially impactful and disruptive applications. The analyses and results of Laser studies are shared with Microsoft senior leadership. Laser studies consider both short-term needs and mitigations well as recommendations for standing up longer-term R&D and processes.

As examples, influential Laser studies and reports were commissioned to study pre-release versions of GitHub Copilot and GPT-4 and provided guidance to the respective product teams months in advance of these models become generally available. Laser studies have examined capabilities, identified concerns about potentially costly behaviors, and have recommended additional research, mitigations, and follow-on workstreams for continuing review and revision.