Technology & public policy · News English

California’s AI order: what changes now?

Separate immediate instructions from proposed safeguards, then make a qualified argument.

C1 · Advanced5–8 minutes18 September 2026

Created by Denis M., British English teacher.

Read the story, then try the activities.

An order today, further decisions ahead

Which actions are definite, and which measures still depend on further decisions?

1California Governor Gavin Newsom issued an executive order on 18 September seeking stronger oversight of advanced artificial intelligence. CNBC reported the announcement that morning. Although the language surrounding the announcement emphasises urgency, the signed document distinguishes immediate instructions to state agencies from measures that remain under consideration.

2The order accelerates implementation of two recently enacted laws concerning independent verification organisations and AI auditors. It also instructs the Government Operations Agency, consulting national experts and the state’s emergency services office, to submit recommendations by 16 November. These must address the technical feasibility and potential efficacy of changes to existing safety legislation.

3Options to be examined include placing independent evaluators inside major AI laboratories, independently verifying safety documents and requiring an emergency shutdown mechanism for frontier models. The order also seeks recommendations on widening the definition of reportable incidents to cover situations in which control is lost. These are subjects for assessment, rather than a declaration that every proposed safeguard is already compulsory.

4Newsom argues that federal oversight is inadequate and presents California’s approach as a possible national model. That is his political assessment. The practical question remains whether the proposed safeguards would work as intended, and how their effectiveness would be independently established.

An original teaching summary of the linked reporting and signed order. The discussion and model responses are learning activities, not reported reactions or quotations.

Word help · Three useful words

Read the meanings and examples. Choose one word to use in your own sentence.

oversight

Supervision intended to ensure that an organisation acts responsibly and follows applicable rules.

Independent oversight can challenge an organisation’s own assessment of its performance.

efficacy

The ability of a measure to produce its intended result; a formal word often used about interventions or policies.

A convincing proposal still needs evidence of its efficacy.

safeguard

A measure designed to reduce a risk or prevent harm.

A safeguard may reduce one risk without eliminating every possible problem.

Check your reading · Three questions

Choose one answer for every question, then check. You can change an answer and check again.

1. What distinction does the opening paragraph prepare the reader to make?

Political urgency means every proposal is immediately enforceable.Agency instructions are definite, while some proposed rules need further assessment.The announcement cancels the state’s existing AI legislation.

2. Why does the reading describe the shutdown mechanism as an option to be examined?

Its technical feasibility and potential efficacy are part of the requested assessment.The order prohibits experts from discussing it.Independent tests have already established that it always works.

3. How does the writer handle Newsom’s criticism of federal oversight?

As an independently proven explanation of every AI incident.As a prediction that Congress will adopt the proposals.As an attributed political judgement, separate from the order’s instructions.

Useful English: concede, then qualify

While / although + clause acknowledges a reasonable point. Follow it with a precise limitation: “While independent checks may improve accountability, their value would depend on the evaluators’ powers.”

May / could + base verb expresses possibility. Would depend on + noun phrase makes a condition explicit. Qualification should explain uncertainty, not simply make every sentence vague.

Try it: make this claim more defensible: “A shutdown mechanism will make AI safe.”

Compare with a model

“Although a shutdown mechanism could reduce some risks, its effectiveness would depend on reliable testing and the circumstances in which it is used.”

The revision limits the claim and names conditions. Other evidence-based qualifications are possible.

These are illustrative arguments for language practice, not expert conclusions reported in the sources.

Say it aloud · 30 seconds

Explain the difference between “The state will request recommendations” and “The proposals will make AI safe”. Which claim is supported, and which needs further evidence?

See a model and follow-up

Model: The request for recommendations is specified in the order. The claim about safety predicts an outcome that the document has not established.

Solo: repeat your answer using “whereas”. Tutor: ask “What evidence would justify the stronger claim?”

Speak about the story · Your assessment

Urgency and evidence

Speak for 45 seconds: should governments introduce safeguards before their effectiveness is fully established? Refer to one action in this story, acknowledge a competing concern and state one condition for your support.

Useful frame: “While I accept that…, I would support… provided that… The order’s request for… matters because…”

Open a model response

I would support interim safeguards where there is credible evidence of serious risk. While waiting for perfect evidence could delay protection, an untested measure could create false confidence. The order’s request for an assessment of technical feasibility therefore matters. I would support a shutdown requirement provided that independent testing demonstrated what it could and could not do, and that the rule was reviewed as evidence developed.

This is an illustrative learner opinion, not a quotation, reported public reaction or the only acceptable position.

Solo: record your answer, then identify one factual claim and one judgement. Tutor: ask “Who should decide whether the evidence is sufficient, and why?”

Mastery check · Three questions

Choose one answer for every question, then check. You can change an answer and check again.

1. Which summary stays closest to the evidence?

California has demonstrated that all frontier AI can be shut down safely.California has commissioned recommendations on additional AI safeguards.All proposed safeguards became binding when the order was signed.

2. Which sentence combines a concession with a clear limit to the evidence?

Although oversight may improve accountability, its effectiveness would depend on the powers and independence of evaluators.Although oversight may improve accountability, it therefore guarantees safety.Oversight might perhaps possibly do something useful somehow.

3. What evidence would best address a proposed safeguard’s efficacy?

How urgently its advocates describe the problem.How often its name appears in a headline.Independent testing of whether it achieves its intended result under relevant conditions.

Sources and reporting date

Reporting checked on 18 September 2026. CNBC published at 11:02 EDT (10:02 Bogotá); original publication metadata confirms 15:02:50 UTC. The official announcement and signed order are also dated 18 September. This summary uses the signed order to distinguish immediate directions from proposals for possible changes to the law.

CNBC: Newsom’s executive order on AI safety
Justin Papp, 18 September 2026, 11:02 EDT. The reporting selected for this lesson.
California Governor: announcement of the AI oversight order
18 September 2026. Primary source for the announcement and the governor’s stated position.
Signed Executive Order N-9-26 (PDF)
Signed 18 September 2026. Sections 1–3 set agency instructions, deadlines and subjects for recommendations.