What about the loyalty of AI agents in government function?

by Sebastian Benthall

For the private sector, there is a well-developed theory of loyalty that shows up in fiduciary duties. I and others have argued that these duties of loyalty are the right tool to bring artificial intelligence systems (and the companies that produce them) into “alignment” with those that employ them.

At the time of this writing (the early days of the second Trump administration), there appears to be a movement within the federal government to replace many human bureaucratic functions with AI.

Oddly, it doesn’t seem that government workers have as well-defined (or well-enforced) a sense of who or what they are loyal to as fiduciaries in the private sector. This makes aligning these AI systems even more challenging.

Whereas a doctor or lawyer or trustee can be expected to act in the best interest of their principal, a government worker might be loyal to the state broadly speaking, or to their party, or to their boss, or to their own career. These nuances have long been chalked up to “politics”. But the fissures in the U.S. federal government currently are largely about these divisions of loyalty, and the controversies are largely about conflicts of interest that would be forbidden in a private fiduciary context.

So, while we might wish that a democratically elected government have an affirmative obligation to loyalty and care towards, perhaps, the electorate, with subsidiary duties of confidentiality, disclosure, and so on, that is not legally the case. Instead, there’s a much more complex patchwork of duties and powers, and a set of checks and balances which is increasingly resembling the anarchic state of international relations.

The “oath of office” is perhaps an expression or commitment of loyalty that federal government workers could be held to. As far as I know, this is never directly legally enforced.

Between this inherent ambiguity of loyalty, and further complications brought on by the fact that government AI will in most cases be produced by third parties and procured, not hired, make the AI alignment especially fraught.