OpenAI released a report titled "Democratic Governance of Frontier AI" on the 3rd local time, proposing that a standard federal bill be drawn up that combines artificial intelligence regulations written separately by each state. The idea is to pull out the core provisions that appear in common across the AI safety bills prepared individually by California, New York, Illinois and others, and raise them to federal law. The same day, Chief Executive Sam Altman visited Washington and met in succession with House Speaker Mike Johnson and Democratic House Minority Leader Hakeem Jeffries.
The logic the report sets out is simple. For AGI, meaning AI with broad human-level intelligence, to benefit humanity, voluntary corporate commitments, company-by-company internal policies and partial regulations attached case by case are not enough. As an alternative, OpenAI presented three mandatory items. Regular publication of transparency reports, annual evaluations and audits carried out by independent outside bodies, and mandatory reporting when a serious safety incident occurs.
It also named an enforcement body. The proposal is to elevate the Center for AI Standards and Innovation (CAISI) under the U.S. Department of Commerce into a permanent body and put it in charge of evaluating companies' highest-performing models before they are released. The picture is one in which the federal government transcribes rules that come up from the states, and a federal agency keeps a pre-release threshold. OpenAI described this approach as building a national standard through "reverse federalism."
A day earlier, on the 2nd, President Trump signed an executive order requiring AI companies to submit their latest high-performance models to the government for security verification before release. The period during which the government examines a model in advance and checks for security flaws is up to 30 days, sharply reduced from 90 days in the draft. President Trump had been set to sign the original version on the 21st of last month but pulled it back at the last moment, reportedly out of concern that 90 days was too long.
The working details of the executive order are divided among several departments. The Treasury, Defense and Homeland Security departments and others will create classified benchmarking procedures that are not disclosed externally, and developers will set up a voluntary cooperation framework with the government. The AI cybersecurity partnership to be formed here will scan for software vulnerabilities on a large scale and coordinate when and how security patches are distributed. The White House said the measure is meant to cement U.S. AI superiority and protect infrastructure and national security systems from cyber threats, and drew a line that it is not the introduction of a licensing or pre-approval system.
Altman said the executive order struck an appropriate balance between innovation and safety. Speaker Johnson said after the meeting that they had discussed a "light-touch regulation" framework to prevent AI harms.
For companies, meeting 50 states' differing rules one by one is a cost, and the predictability of knowing in advance what line they need to meet governs the development schedule itself. Seen from the other side, consolidation may amount to leveling down. Whether a federal standard would replace state law or apply alongside it has not been settled, and until that point is sorted out it is also hard to define what character the consolidation would take.
Political money is part of the backdrop as well. OpenAI stressed its neutrality on the 1st, saying it does not donate to particular candidates, parties or political action committees and does not operate a PAC. President Greg Brockman and his wife Anna Brockman, however, donated in a personal capacity to the pro-Trump super PAC "MAGA Inc." and to pro-AI political groups.
What changes immediately is the release rhythm that developers and users feel. Up to a month of government review will be inserted before a new model reaches the world, and depending on the form in which those results are disclosed, users will be able to check for the first time what kind of examination the tool they use has passed. If outside audits and incident reporting become mandatory every year, it will also start to become possible to distinguish whether a chatbot's wrong answer is an isolated occurrence or something subject to reporting.
Neither the threshold for what counts as a "high-performance model" subject to verification nor what happens when the 30 days are not observed has been fixed in a document. Making CAISI permanent requires a budget and legislation, and what kind of bill the "light-touch regulation" mentioned by Speaker Johnson will turn into remains to be seen. Until then, what users can hold in their hands is roughly the transparency reports each company puts out.
