Biden signs the executive order on safe and trustworthy AI

President Biden signed Executive Order 14110, the most extensive action any government had then taken on artificial intelligence. Running to roughly 20,000 words, it directed more than fifty federal entities to act across safety, security, immigration, healthcare, housing, labour and civil rights.

Two provisions drew most attention. Using the Defense Production Act, the order required companies training models above a compute threshold — 10²⁶ floating-point operations, or 10²³ for models trained primarily on biological sequence data — to notify the federal government and share the results of red-team safety testing. It was the first time a government had asserted visibility into frontier training runs, and the first use of a compute threshold as a regulatory trigger, a device the EU and California later borrowed. The order also directed NIST to establish what became the US AI Safety Institute and to develop standards for red-teaming.

Other sections required watermarking guidance for synthetic content, tasked agencies with assessing chemical, biological, radiological and nuclear misuse risk, and moved to ease immigration for AI talent. Much of it took the form of instructions to write reports and standards rather than binding rules on industry, which critics read as the limit of what could be done without legislation. Congress passed no comprehensive AI statute in the period that followed.

The order arrived two days before the UK’s Bletchley Park summit, which several observers read as deliberate. It was revoked by President Trump on 20 January 2025 and replaced by an order framed around removing barriers to American AI leadership — making its two-year life a compact illustration of how much US AI policy rested on executive discretion rather than law.