Chatbots are doing Congress's work with little oversight
Two stories today on how the people making the rules are getting used to AI — one from inside the Capitol, one from the labs shipping the tools.
Across the Capitol, lawmakers and staffers are now using chatbots to write speeches and news releases, sort constituent mail, prepare hearing questions, and draft amendments — with almost no rules governing any of it. A Washington Post investigation found both chambers have cleared members and staff to use Copilot, ChatGPT, Gemini and, in the House, Claude for official work, with more tools under review. The kicker: this is happening while the same politicians debate how to regulate AI for everyone else. Congress has effectively settled its own answer — use it quickly, broadly, and with little oversight.
The story matters beyond the irony of the fox guarding the henhouse. Aides are feeding constituent correspondence into commercial chatbots, which raises real privacy and accuracy questions: whose data is in those prompts, what happens when a model hallucinates into a draft amendment, and who is accountable for an AI-written news release attributed to an elected official. It also quietly normalizes AI in the most consequential legislative body on earth. Every committee hearing on AI safety is now being prepared, in part, by AI. The gap between the technology's actual use and the pace of rulemaking has rarely been this literal.
Google is making the visible watermark on Gemini generations optional — a one-click setting that lets users strip the AI label from images, videos, and songs. The toggle, announced by Gemini VP Josh Woodward, covers the Nano Banana image model, the Omni video model, and Lyria music, and rolls out in the Gemini app and Google's Flow video editor, with Search support coming. Google's reassurance: the invisible SynthID watermark and C2PA metadata stay on, so AI origin can still be verified.
The move is a deliberate bet that provenance can survive without friction. Critics note the visible mark was the last thing forcing creators to acknowledge AI output at a glance — SynthID is robust but not immune to determined attackers, and C2PA metadata is routinely stripped by social platforms on upload. The timing sharpens the contrast: days earlier, Anthropic went the opposite direction, adding watermarks to Claude's text and files to comply with EU rules. One industry leader is baking labeling in, another is making it a preference. That divergence — regulation-driven versus market-driven — is the real story, and it will define how much of the internet's next wave of content carries a legible AI label.
What to watch: whether the Senate's AI working group publishes its own usage rules before the next election cycle — and whether Google's Search rollout of the toggle arrives before regulators ask why it exists.
If Congress ran on your company's AI tools, would you want a rulebook first? Tell us in the comments.
