Biden Signs Most Sweeping US Executive Order on AI

On October 30, 2023, President Joe Biden signed an Executive Order on artificial intelligence, marking the most ambitious intervention by a Western government on the technology to date. The order requires developers of the most powerful AI systems to notify the federal government and share the results of their safety tests before releasing them to the public.

The order lands two days before the first international AI Safety Summit kicks off at Bletchley Park in the UK, where Vice President Kamala Harris will attend. The White House clearly wanted to walk into that meeting from a position of strength: while Europe is still negotiating its AI Act and the UK is hosting diplomatic talks, Washington is already imposing concrete obligations on its own companies.

What the order requires

The central mechanism relies on the Defense Production Act, a Korean War-era statute that lets the executive branch demand information from private companies on national security grounds. Going forward, any company training a foundation model above a set compute threshold — 10 to the 26th floating-point operations, or 10 to the 23rd for models focused on biological sequences — must notify the government of the training run and share the results of its "red-teaming" tests, the exercise in which experts deliberately try to provoke dangerous behavior in a system before launch. Source.

Today, that threshold is only crossed by the largest models from labs like OpenAI, Google DeepMind, Anthropic and Meta. The order also directs the National Institute of Standards and Technology (NIST) to develop rigorous standards for those safety tests before models reach the public.

Content labeling, biotech and cybersecurity

The order directs the Department of Commerce to develop guidance for labeling and watermarking AI-generated content, so citizens and authorities can distinguish synthetic material from the real thing. It also orders various agencies to assess how AI could facilitate the design of biological, chemical or nuclear weapons, and establishes new safeguards for the use of these systems in critical infrastructure such as power grids or water supply systems.

The order also addresses AI's use in cybersecurity, both as a defensive and offensive tool, and creates programs for federal agencies to test AI systems before adopting them internally.

Equity, jobs and talent

Beyond technical safety, the text includes provisions on consumer and worker protection: it directs federal agencies to develop principles to prevent hiring, housing or criminal-justice algorithms from discriminating on the basis of race or gender — a problem well documented for years in credit-scoring systems and facial recognition.

The order also aims to make it easier for foreign AI talent to come to the United States, streamlining visas for researchers and experts, a move that responds to the global race to attract the scientists who are currently in short supply in the sector's labor market.

From voluntary pledges to law

This order didn't come out of nowhere. Back in July, the White House had already secured voluntary safety, transparency and external-testing commitments from seven major companies — including Google, Microsoft, Meta, Amazon, Inflection, and OpenAI and Anthropic themselves. Those agreements lacked any enforcement mechanism: they were promises, not legal obligations. Today's Executive Order converts much of that promise into formal requirements backed by executive authority.

Still, it's a tool with limits. An executive order is not a law passed by Congress: it can be revoked by a future president as easily as it was signed, and its reach depends on the ability of agencies like NIST and the Department of Commerce to translate its mandates into concrete, verifiable technical standards. The White House is nonetheless presenting it as the most comprehensive government action ever taken on artificial intelligence — and as proof that the United States wants to set the rules before anyone else does.

What happens in the coming months at NIST and the Department of Commerce will determine whether these mandates turn into standards with real teeth or remain a framework of good intentions. The Bletchley Park summit, opening the day after tomorrow, will be the first venue where Washington tries to convince its allies to follow the same path.

An executive order combines rules, assignments and timetables

The signed text does not turn every objective into an immediate duty for every company. Some clauses direct an agency to act; others define a term, request a report or make a measure conditional on existing legal authority. To learn what changes, identify the verb, addressee, deadline and later instrument. “Develop guidance” does not mean “prohibited from today”.

Executive Order 14110 in the Federal Register runs for dozens of pages and assigns work to Commerce, NIST, Homeland Security, Labor, Health and other agencies. The compute threshold triggers reporting duties for certain models, while many sections begin processes that will require standards, consultations or later rules. The political announcement and administrative implementation are different stages.

A practical reading can become a five-column table: exact section, responsible actor, action, deadline and proof of completion. When a deadline passes, look for the resulting report, standard or proposal. Record a delay if it is absent; if it appears as a draft, do not describe it as a final rule. This method prevents a press release from being mistaken for implementation.

Thresholds need context too

A figure such as 10^26 operations looks like a clean technical boundary, but it depends on definitions: which training counts, how runs are aggregated and which records the company must retain. The document itself gives models involving biological sequence data specific treatment. Comparing systems requires the same axis and the definitions, not turning one number in the order into a universal measure of danger. Data source.

The order does not replace legislation from Congress. It relies on existing executive powers, directs the administration and can be changed by another president. It nevertheless has practical effects when an agency publishes a standard, adds a procurement condition or requires a report under current authority.

The transferable skill is to follow technology policy from its verb to its evidence. For any large regulatory announcement, find who must do what, under which authority, by when and in which document completion can be checked. That chain separates an intention, an administrative instruction and an enforceable obligation.

The exact page prevents overstatement

In a document this broad, linking only its landing page is insufficient for a specific assertion. A reader should be able to locate the section, subsection and, where possible, page. The training-report duty is not in the same place as instructions on work or privacy. Citing the section preserves scope and exposes conditions that paraphrase can lose.

Tracking also needs a cutoff date. An agency may comply after the story is published; that progress belongs in another update, not in the account of October 30. Preserving the editorial clock makes implementation measurable without crediting the announcement with outcomes that did not yet exist.

This procedure lets a reader return months later and verify progress. It also reveals which part depended on the president, which on an agency and which required Congress. Responsibility stops being a general word and becomes attached to an actor and document.