According to a Washington Post examination, members of the US Congress and their aides are using AI tools in their daily work with little oversight. The uses range from writing speeches and preparing news releases to sorting mail from constituents.

Why it stands out

Some of this is ordinary office work. Having a model draft a news release is a question of efficiency in Congress just as it is in any organisation.

But the list does not stop there. Sorting constituent mail determines which citizen's request reaches a lawmaker's desk. That is a filtering decision, and who built the filter and on what criteria it runs touches directly on how representative democracy functions.

What the lack of oversight means

The key phrase in the reporting is "little oversight". That does not mean the use is prohibited and happening covertly; it means no shared rule exists at all.

In practice each office makes its own call: which tool to use, for which task, how much to check the output, and whether to disclose any of it. The result is wildly different practices sitting side by side within the same institution.

Where the real risk lies

Congress using AI is not in itself a problem. The problem is that the use is not traceable:

  • Unclear provenance — where a claim in a speech came from, and whether a model produced it, is unknown.
  • Error propagation — when a model states a fact wrongly, that error enters an official text.
  • Data security — constituent mail is personal data; which tool it was sent to, and what happened there, is unclear.

The regulator's dilemma

There is a particular irony here: the institution charged with regulating AI has not regulated its own use of it. Experience accumulating inside the legislature could improve the resulting law — someone who does not use a technology does not know what needs regulating either.

But that same experience creates a conflict of interest. Constraining a tool that eases your own workload is easier for a lawmaker who never uses it.

Nor is this argument specific to the US. AI tools are spreading rapidly through public administration, and in most countries that spread is happening before any framework defines their use. When a rule arrives late, it usually arrives after an incident and on top of established habits.

Not what to ban, but what to write down

The answer is unlikely to be prohibition. A legislative office that does not use AI cannot keep pace with one that does, and that gap eventually shows up in quality as well.

What is missing is a set of simple rules that make the use visible:

  • Recording which parts of an official text were prepared with model assistance.
  • Defining which tools personal data from constituents may be sent to.
  • Requiring that a factual claim produced by a model is verified before publication.

None of these prevents the use of AI; they simply leave behind an answer to the question that gets asked later — where did this come from?