○On April 14, the Trump administration quietly acknowledged that AI is being widely used to automate government processes. The Office of Management and Budget (OMB) has identified 3,611 ongoing or planned AI use cases across the federal government. The list is 70% larger than the one released in the final year of the Biden administration and includes many ominous-looking plans to transfer sensitive government functions to AI.
After skimming this list, many readers may find many reasons for feeling anxious. This means a massive shift in decision-making processes from humans to machines on issues such as personal freedom, public health and welfare, and nuclear reactor safety.
Consider these examples. The Department of Health and Human Services’ (HHS) Office for Children and Families hired Palantir, the world’s “most feared AI company” notorious for working on behalf of the military, CIA, and ICE, to scan all grant applications and issue warnings to those who ideologically do not align with the administration’s lead. The Federal Bureau of Prisons is developing an AI system that assesses “the potential for misconduct in newly admitted inmates” and sends people to close confinement before they actually commit misconduct while in custody. These are the kind of programs that would be perfect for a Philip K. Dick or George Orwell novel.
Other use cases include incorporating AI into life and death decisions. The Department of Veterans Affairs is developing AI that will intercept calls to the Veterans Crisis Line and gather information from external databases to assess the caller’s mental state and risk of suicide.
The Department of Energy is testing the use of AI to control nuclear reactors, with the goal of autonomously responding to potential nuclear safety incidents. The point of concern is not the deployment, but the retirement. The State Department has ended a program that used AI to predict mass killings of civilians, which was aimed at conflict prevention.
While it is easy to question these and similar uses of AI, the reality is that any of these programs can be implemented responsibly. In some cases, like the HHS system, AI may force adjustments to policy prescriptions that opponents dislike. But the concern is more about the policy itself than the idea that government agencies should follow executive orders.
Or there may be bipartisan agreement on a goal, such as taking urgent steps to help veterans at risk of self-harm. Much work and validation is needed to prove that AI is safe and effective for these use cases, and to convince the public that it is appropriate, but the idea is plausible.
In some cases, the use of AI may not be new, even though it sounds scary. The use of predictive techniques and statistics to assign prisoner security classifications dates back several decades, even though such systems are often biased and inefficient.
The use of autonomous systems for model predictive control (MPC) of nuclear reactors is a well-studied and widely applied aspect of nuclear power plant management. And the recently revealed AI additions began under the Biden administration.
But anyone looking at the stock picture for 2025 could be forgiven for jumping to harsh conclusions. What matters is the details of how the AI system will be used, and this is where inventory is severely lacking.
The disclosure contains minimal information and lacks the necessary context to understand its purpose and approach. Descriptions are usually just sentences and rarely exceed a paragraph.
And while this process would theoretically include some form of public consultation, it generally does not in practice. It would take a sharp-eyed citizen to notice this information disclosure. Unless you regularly read FedScoop or monitor OMB’s Federal Chief Information Officer GitHub account, you’re probably missing out.
Of the examples cited above, only one, the Department of Justice, proposes public participation. Government policy does not require the remaining use cases to be classified as “high-impact” use cases, a label that is applied inconsistently across government agencies.
We wrote a book exploring the application of AI to democratic processes around the world, including executive agencies, courts, legislatures, and politics. Our conclusion was that while there is a need to resist the misapplication of AI in governance, an urgent need to reform the economics of AI, and an urgent need to revamp the democratic systems in which AI is being unleashed, there are also valuable and beneficial use cases for AI in government.
Machine translation is a good example. Customs and Border Protection (CBP) has deployed an AI translation system to assist officers in the absence of human interpreters. The idea that the CBP agency, which has reported human rights violations and is under intense scrutiny, would instruct people to talk to machines rather than humans may seem inhumane to many.
Indeed, human interpreters have a huge advantage when it comes to picking up nuances from physical cues and social context. But a police officer with a capable AI interpreter at his fingertips is better than one who can’t communicate with the person in front of him.
The Trump administration’s AI use case inventory lists 70 such translation use cases, up from 58 in the Biden administration’s 2024 disclosure.
Disclosure of AI use cases can be a means of building public trust and confidence, but only when combined with consistent and meaningful public consultation. Washington, DC and California are actively engaging with the public to determine where and how AI should be used in government processes, or when it is appropriate for governments to regulate AI use in society.
Both companies have used their AI platforms to host large-scale public debates on this topic. These examples demonstrate the potential of gathering broad public input to shape AI policy.
The international gold standard was most likely set up by France in 2016 under the Digital Republic Act. The law itself, enacted through an online citizen consultation, requires all algorithms used to automate government administrative decisions to be subject to public records requests, to be able to challenge human reviewers, and to notify those affected by the decision about the use of automation.
Canada provides another example of what more rigorous and participatory disclosure can look like. In 2025, we launched an AI Use Case Registry, similar to our U.S. inventory. However, Canada also has a federal directive that requires transparent risk scoring and impact assessment processes for automated systems that make administrative decisions about citizens.
This long-standing directive requires a detailed description of risks and benefits and consultation with specific stakeholders from the conception of AI use cases. Canada’s system could be improved. A public comment period and an obligation for agencies to respond substantively to feedback may be required before addressing the prudent use of AI.
AI offers real potential to improve government effectiveness, efficiency, and accessibility. But equally, there are legitimate concerns and mistrust among the public that can only be addressed through transparency and dialogue. The United States should implement algorithmic impact risk assessment procedures and public comment processes at the federal and state levels to facilitate safe, reliable, and fair transformation of government agencies to take advantage of emerging technologies.
-
Nathan E. Sanders is a data scientist at Harvard University’s Berkman Klein Center and co-author with Bruce Schneier of Rewiring Democracy: How AI Will Transform Our Politics, Government, and Citizenship. Bruce Schneier is a security engineer who teaches at the Harvard Kennedy School at Harvard University and the Munk School at the University of Toronto.
