I am writing regarding the request for comment to the Office of Management and Budget draft guidance on Advancing Governance, Innovation, and Risk Management for Agency Use of Artificial Intelligence. Broadly speaking, I believe that the OMB and other federal agencies must use all of their powers and resources to ensure that the harmful effects of AI are limited when developed, procured, or used by the federal government—and that the government provides meaningful transparency and publicly available information about the ways our government uses AI so that we have more power to regulate, push back on harmful uses, and limit AI use in the future.

There are some strong pieces of the OMB guidance that I urge you not to rollback going forward. The requirement that agencies meet minimum practices for rights-impacting and safety-impacting AI is very important in addressing my concerns around bias and discrimination in AI, and I urge you to keep these strong requirements and ignore pressure to weaken this guidance.

There are a number of areas where this guidance, and future AI policy documents, must be improved:

The Executive Order on AI did not do enough to address law enforcement agencies and the use of AI on immigrant communities, which are some of the spaces where AI has already been shown to have serious negative impacts. I was glad to see that the OMB guidance did include law enforcement under the covered agencies, and strongly urge you not to change this or allow these agencies special exceptions. We must establish strong policies to address law enforcement use of AI, and to limit use to mitigate harms.
Some AI is too dangerous to be used, and any possible benefits are outweighed by known harms. Facial recognition is an AI-enabled biometric technology that infringes on the right to privacy to such a degree that it should not be used at all by the government, and this guidance should ban its use by government agencies.
The guidance does not establish enough data privacy protections. Because AI is reliant on data the opportunity for data abuses is high, and ensuring strong privacy protections must be a top priority. While there is a need for Congress to pass comprehensive data privacy protections, the OMB should use this guidance to establish data privacy policies when government agencies develop, procure, or use AI.
The minimum practices for safety-impacting and rights-impacting AI require agencies to stop using an AI when the “risks to rights or safety exceed an acceptable level,” but the guidance both gives agencies an option to skip determining an AI system’s risk, and also doesn’t specify what an “acceptable level” is. In these safety- and rights-impacting cases the consequences of the use of AI are so high that the OMB must set a very limited acceptable level of risk and ensure that all agencies using AI comply with this requirement.
The minimum practices should also apply in cases where federal funds are used to purchase AI systems, including government grants. This would better protect civil rights in even more situations, including at the state and local levels.
All information reported by agencies in their annual use case inventory should be made publicly available, to allow for as much public scrutiny and transparency as possible. This is critical for ensuring that the public, and experts in civil rights, have the opportunity to critically assess the use of AI by the government.

While there may be uses of AI that can benefit the federal government, I call on the OMB to focus on protecting against the harms of AI your top priority in this guidance.

Signed,