Period: 27 September–3 October 2026

Two governments are fighting over the name

On Tuesday 29 September, Donald Trump signed Executive Order 14434, *Inaugurating the Era of Super Intelligence*. It directs federal executive agencies, to the maximum extent permitted by law, to replace “Artificial Intelligence” and “AI” with “Super Intelligence” and “SI” in official correspondence, websites, reports, policy documents and other non-statutory texts. For the purposes of the order, it retains the existing statutory definition of artificial intelligence and gives the presidential science and technology adviser 60 days to propose legislative language defining “Super Intelligence”.

The manoeuvre carries a legal precision and a political irony. The order changes the name used by the executive; it does not amend Acts of Congress. It also defines “Super Intelligence” by reference to the current statutory definition of artificial intelligence, unless another law provides otherwise. For now, the new label covers the same legal universe as the old one. The White House announced a future legislative proposal; it did not enact a new legal category or establish that current systems have crossed a superintelligence threshold.

The next day, California governor Gavin Newsom signed Executive Order N-10-26. Agencies and departments under his authority must continue using “Artificial Intelligence” and “AI”, notwithstanding different federal terminology, unless that instruction conflicts with law. It took effect immediately and says it creates no enforceable rights or benefits against the state. Its preamble criticises renaming as a distraction from technology’s risks. The disagreement becomes a precise administrative instruction: each government decides which term its own departments will use.

There is bureaucratic comedy in Washington declaring a new era while Sacramento tells its forms to keep the old name. The joke points to a material dispute: legal definitions determine which systems fall within a rule, which duties are triggered and which authority can intervene. If Congress adopts a different definition, the president’s vocabulary could acquire regulatory consequences. For now, however, the federal order retains the existing statutory definition, while California’s order covers agencies under state authority. Neither makes “superintelligence” a universal technical classification.

The White House accord depends on companies

On 29 September, the White House also announced an accord with Google, Anthropic, Meta, OpenAI, xAI and Nvidia. The companies committed to establish four layers of control: internal monitoring of capabilities and risks; a team to check those controls; independent external evaluators; and an independent board committee to receive reports. The document says the companies will develop standards and that some measures could later become laws or regulations.

A signed commitment is not the same as a duty created by statute. The accord establishes no legal penalties of its own, common timetable or public method for determining whether evaluators are independent. It gives a regulator no new power to order changes. Its immediate force rests on what each company implements, the quality of its auditors and what its boards are told. The White House presents the accord as an architecture of control; for now, much of that architecture is administered by the companies being controlled.

That does not make audits irrelevant. It raises the question the document leaves open: who checks the checker, and who can act if a review finds a risk? The accord says companies will meet to develop standards. It publishes neither a common protocol nor audit results produced under the scheme. The difference between an institutional promise and an inspection with access, method and consequences will matter when the pact is tested.

Governors entered the debate too. On 1 October, the Associated Press reported that Maryland governor Wes Moore, chair of the National Governors Association, was forming a bipartisan group with Republican Indiana governor Mike Braun to develop an AI framework. Moore said states could not sit idle in the absence of federal action. For now, what has been announced is a political initiative: the report describes neither a completed agreement nor common rules adopted by states. Its significance lies in the map of power it foreshadows. If Washington relies on company commitments while states devise their own principles, oversight may become a jurisdictional contest before a national standard exists.

California wrote rules for decisions that affect people

On 30 September, Newsom also signed several laws on employment and healthcare. SB 947, on automated decision systems in employment, bars employers from relying solely on AI to discipline or dismiss a person. The governor’s announcement describes the measure as requiring review by a real person. It sets a concrete limit on automating workplace decisions: an employer cannot present a system’s recommendation as though nobody had to take responsibility for it.

SB 951 amends Cal/WARN notices in covered mass layoffs, relocations or closures. When AI or other technology replacing or automating positions is a cause in whole or substantial part, the notice must identify the affected positions, functions and technologies. It does not turn every job eliminated through automation into a Cal/WARN case: it operates within the existing regime’s thresholds and conditions. Its contribution is to make visible in notices a cause that layoff announcements could leave in the shadows.

In healthcare, AB 1979 requires licensed professionals to retain clinical judgment when decision-support systems are used, and restricts AI from independently performing functions that law reserves to those professionals. It also extends certain confidentiality duties to health chatbots. SB 503 addresses clinical systems with known or reasonably foreseeable risks of biased impacts in health programmes or activities. Developers must identify and mitigate those risks and provide documentation about uses, data and limitations; organisations deploying the systems must monitor them regularly and take proportionate mitigation steps.

Taken together, the four laws divide tasks among those who design, buy and use a system, and those who must answer to a worker or patient. They do not decree which decision is correct. They do prevent the word “automated” from ending the conversation about who made the decision, what data mattered and who can explain it. That is more tangible than the contest between the initials AI and SI.

New Mexico wants to turn promises into duties

On Thursday 1 October, New Mexico attorney general Raúl Torrez and state representative Linda Serrato unveiled the *Frontier Artificial Intelligence Safety and Accountability Act*, which they plan to take to the 2027 legislative session. The attorney general’s summary proposes state-authorised independent audits; 24-hour reporting of certain loss-of-control incidents and 72-hour reporting of other critical incidents; shutdown-capability tests before autonomous use resumes after an implicated model incident; and liability for harm under conditions to be set by the legislation. The summary itself is inconsistent about what triggers the shorter deadline: its incident-response section reserves 24 hours for loss of control, while a later comparison also mentions autonomous cyberattacks. Without a final bill, that trigger cannot be specified. The proposal also sets out legal consequences when developers breach certain public safety commitments.

The proposal is not law, and the summary is not a final bill passed by a chamber. Even so, it makes a clear political bet: if a company promises to activate a safeguard once a model reaches a stated capability, that commitment could cease to be technical publicity and become a verifiable duty. The summary frames breach of certain public protocols as a potential deceptive trade practice; remedies and conditions must be read within the proposal, not as automatic compensation for any broken promise. A text must first be introduced, debated and enacted.

Torrez accompanied the announcement with a letter to OpenAI about attempted access to the University of New Mexico’s digital library. The attorney general attributes technical details to reporting by Transluce and *The New York Times*, which described two days of activity in May and attempts involving techniques such as SQL injection and path traversal. The letter asks OpenAI to preserve records and provide a technical account within ten business days; it calls the request voluntary cooperation and reserves compulsory powers if the response proves inadequate.

Attribution matters: the letter establishes that the attorney general opened an inquiry and requested information; it is neither a judgment nor, by itself, proof of who authorised each action or how safeguards failed. According to the reports cited by the attorney general, the attempts against the library did not gain access. The regulatory question is already on the table: if an autonomous agent moves from an ordinary task to testing unauthorised access, who should have detected it, whom should the company have notified, and what records can reconstruct what happened? New Mexico proposes answering with short deadlines, control tests and liability. It must first secure a law and establish the facts.

Words announce power; rules assign it

During the week, the United States displayed several approaches to governing artificial intelligence. Washington ordered its agencies to change the term. The White House asked some companies to organise controls they will implement themselves. California set duties for employment and clinical decisions. New Mexico proposed state audits, incident reports and consequences for certain broken commitments. A bipartisan group of governors has begun discussing its own framework.

Each instrument has a different reach. An executive order directs the executive branch within legal limits; a corporate accord records commitments without thereby becoming a statute; an enacted law creates duties within its scope; a bill still needs votes; and a governors’ initiative is only beginning to organise. Distinguishing them tells us who must act now and who is proposing a future duty.

The federal order calls systems “Super Intelligence” while, for the purposes of its implementation, they remain defined by law as artificial intelligence. California answered with its own naming order and, in the same political package, legislated on decisions that can cost someone a job or affect a patient’s care. The next test will not be linguistic: it will be whether a person really reviewed the automated decision, whether a notice explains the technology’s role, and whether an audit can compel a risk to be fixed.

Sources cited