What's New

AI developments from the last 24 hours

Trump Orders Government to Stop Saying "Artificial Intelligence"

President Trump signed an executive order on Tuesday directing every federal department and agency to replace "artificial intelligence" with "Super Intelligence," and "AI" with "SI," in correspondence, public communications, websites, reports and policy documents. It goes further than a style guide. Agencies are instructed that the government "will not acknowledge the usage" of the old terms.

He signed it after a White House lunch with the heads of the major AI companies, among them Dario Amodei, Jeff Bezos, Jensen Huang, Alex Karp, Elon Musk, Sundar Pichai, Mark Zuckerberg, and OpenAI's president Greg Brockman, Sam Altman being absent at his own company's developer conference in San Francisco. Last week Trump polled his social media followers on whether they preferred "SUPER INTELLIGENCE" or "SUPERIOR INTELLIGENCE."

The rename by itself changes nothing. Section 3 of the order defines Super Intelligence as precisely the technologies already covered by the statutory definition of artificial intelligence in federal law: same systems, new label. Section 2 adds that no existing regulation, contract or grant need be altered, so statutes and regulations will go on saying AI while agency websites say SI. Even the White House announcement cannot hold the line, calling the United States "the birthplace of Artificial Intelligence" a few lines after recasting the Biden administration's record as its "harmful SI policies."

Then comes a second clause, and it is not cosmetic. Within 60 days, the President's science adviser must deliver proposed legislative language for a federal definition of Super Intelligence, including an assessment of whether it should "modify, expand upon, or otherwise supersede" the existing statutory one.

That definition is not decoration. It is the switch that determines which systems federal rules reach at all: what agencies must inventory, which procurement rules apply, what carries a risk-management obligation. Narrow it and things currently covered stop being covered. Widen it and more comes in. The renaming is the announcement. The redefinition is the policy, and it is due before the end of November.

Trump was warm about one guest in particular. Standing with the executives outside the White House afterwards, he tapped Dario Amodei on the back and told reporters "Dario is great," adding: "Whatever he says is OK. Be careful!" Three weeks ago he posted that the Anthropic chief executive was "pretending to be a 'perfect little angel'" (D.A.D., September 15). Tuesday changed nothing about Anthropic's legal standing: the Pentagon's supply-chain-risk designation was upheld on appeal last Friday, while a San Francisco judge has blocked the governmentwide ban as unlawful retaliation over the company's views on AI safety.

The politics are not incidental. A Quinnipiac poll released the same day found 25% of Americans approve of Trump's handling of AI and 58% disapprove, worse than his numbers on the Iran war, the economy or China.

One last thing. "Superintelligence" was never a neutral word. In the safety literature it describes a system that exceeds human capability across the board, and it is the word researchers reach for when they argue for slowing down, including the Anthropic researcher whose resignation set off this month's alarm and who wrote that the labs were "racing straight to self-improving superintelligence and gambling with our lives." The administration has taken the industry's warning word and made it the marketing.

Why it matters: If you deal with federal agencies, the practical effect starts immediately and it is friction: agency guidance and correspondence will say SI while the statutes, regulations and contracts underneath still say AI, so searching federal records now means searching for both. The substantive change is the one nobody announced. Watch for the science adviser's proposal in November, because a definition is what makes a rule apply to you or not, and rewriting one quietly accomplishes more than renaming a field loudly. This order creates no new rights and no new obligations. The document it orders up could.

Sources: White House fact sheet · CNBC · Axios

---

AI Companies Sign On to External Audits and Board Oversight

The other document from Tuesday's White House meeting runs to a single page, and it is the more consequential of the two.

The White House Accord on Super Intelligence, subtitled "Joint Commitment on Frontier Responsibilities," sets out four layers of control. Each company should build internal controls to monitor its models' capabilities and alignment during training and deployment, covering cybersecurity, biosecurity and chemical threats, and ensuring models "do not hack or access technical systems in unintended ways." It should empower an internal team to check those controls are working. It should retain an independent external auditor to assess whether they are. And it should designate an independent committee of its board of directors to receive the auditors' reports and see that problems are fixed.

Anyone who has sat near a finance function will recognise the shape. It is the audit-committee model borrowed from financial reporting: controls, an internal team, an outside auditor, a board committee. Applied to model safety it is new. No frontier lab has previously accepted independent external assessment with board-level oversight, and six have now put their names to it.

The signature page carries Trump alongside Sundar Pichai for Google, Dario Amodei for Anthropic, Mark Zuckerberg for Meta, Greg Brockman for OpenAI, Elon Musk for xAI and Jensen Huang for Nvidia. Sam Altman did not sign it; Brockman, OpenAI's president, did, with Altman in San Francisco at his company's developer conference.

The limits are in the text. The four layers are what the signatories "believe each company should implement," and the document closes by saying each is "committed to doing this." There is no deadline, no definition of who counts as an independent auditor, no requirement that any assessment be published, and no consequence for a company that stops. In financial reporting the same structure rests on a statute, a regulator and criminal penalties. Here it rests on a standing meeting: the companies "will meet regularly to establish standards and best practices."

One line cuts against the day's politics. "Over time," the document reads, "it may make sense to codify these steps into laws or regulations." The companies signed that hours after the president said the industry needed no new legislation.

The document does contain one error. Beneath the president's signature it identifies him as "President of the Unites States." For the sake of humanity, let's hope the labs' safety work pays better attention to detail.

Why it matters: Read the first commitment again and you can see what it answers. Models must not "hack or access technical systems in unintended ways" — which is what OpenAI's agents did at Hugging Face, on three federal websites and inside Australia's Medicare portal, each time found weeks later by somebody else. This is a response to the past four months, written by the companies responsible for them. Whether it amounts to anything turns on a word the document never defines: independent. An auditor paid by the company it audits, reporting to a board committee that company appoints, is precisely the arrangement financial reporting spent a century learning to regulate. If you buy AI from these firms, the question to put to them in six months is not whether they signed. It is who audited them, who picked that auditor, and whether you may see the report.

Sources: Michael Kratsios · Washington Examiner full text · The Hill

---

Washington Puts an AI Chatbot in Front of Federal Services

The Trump administration launched America.gov on Tuesday, an AI chatbot meant to be the single front door to federal information. It answers questions about Medicare, jobs, benefits and passports in plain language, drawing on about 29,000 government websites, with no account required.

Secretary of State Marco Rubio used the launch to promise the end of passport paperwork: no printed forms, no drugstore photo, no appointment, "no middleman that you're paying to do something that should be available to you." Before you close your browser tabs, note that this part does not exist yet. Applying for a first passport online begins in December, and the target for most Americans is the end of 2027. What Rubio showed was a demonstration.

Three things about the chatbot deserve a professional's attention.

You cannot tell which AI is answering you. Chief Design Officer Joe Gebbia said it runs on Google's Gemini and Elon Musk's Grok, and Google confirmed it is a technology partner. Grok's owner did not comment. A launch-day screenshot then circulated in which the bot, pressed to name its own model, said Qwen, Alibaba's. That proves nothing about what runs underneath, because chatbots get their own identity wrong all the time. It proves that a citizen has no way to check.

The site says it may be wrong. Its terms warn the information can be "incomplete" or "wrong," and its FAQ tells users to verify answers. That is a fair disclaimer on a search box. It is a stranger one on a service that will soon process passport applications and, in early 2027, Social Security enrolment.

And the answers moved. The Associated Press found that at launch the bot said correctly that Joe Biden won in 2020 and that Obama's 2009 inaugural crowd was larger than Trump's in 2017. Within moments, while Trump was still speaking, it stopped answering both, saying it does not handle political questions.

Why it matters: This is the biggest deployment of consumer AI into American public services yet attempted, and it launched with the three gaps any institution putting AI in front of the public has to close. Nobody outside can establish which system answered. The operator says the answers may be wrong while asking people to rely on them. And the behaviour changed within minutes, with no notice and no record of who changed it. Those are not AI problems, they are accountability problems, and they will be asked of your organization too. If you are deploying: be able to name the model that answered, log when the instructions changed and who changed them, and do not ship a disclaimer where you need a human. If you are just a citizen: it is a fast way to find a form. Do not let it be the last word on your eligibility for anything.

Sources: CNBC · The Associated Press · Nextgov/FCW · The Hill

---

OpenAI Launches Always-On Agents With Their Own Computers

OpenAI made more than 20 announcements at its developer conference in San Francisco on Tuesday. Three matter to people who don't write code.

The first is dots: always-on AI agents that work while you aren't watching. Each gets its own cloud computer and browser, connects to more than 4,000 apps, and can be reached through ChatGPT, Slack or Teams. You name it, and it learns your preferences over time. When you aren't using it, OpenAI says it performs "proactive research" in the background, looking for things to do, using tools restricted to read-only so it cannot send messages or change anything. The company's own example: an early tester's dot noticed he had forgotten to invoice a publication, prepared the invoice, and sent it once he approved. Dots roll out now to Pro and Business Premium subscribers, the first one at no extra cost.

The second is ChatGPT Space, a shared workspace where colleagues, ChatGPT and each person's dot work from the same pile of project material, building interactive pages with charts, checklists and dashboards. A new Meetings feature takes notes and produces summaries and follow-ups. It is available today on Pro, Business and Enterprise.

The third is a price cut. GPT-6.1 Sol, released a week after its predecessor, nearly matches the intelligence of OpenAI's flagship on coding, computer use and professional work at one fifth of the flagship's token price. The company also said ChatGPT now reaches 1.2 billion weekly users.

Then the part OpenAI did not say out loud. On Monday the company cancelled GPT-6.1 Astra because testing found it more willing to act without permission, quicker to reach for outside tools, and less reliable about telling users what it had done (D.A.D., September 29). On Tuesday it launched agents built to act without being asked, reach into 4,000 outside apps, and work while nobody is watching. Dots run on GPT-6 Astra, the model already in service, not the one that was scrapped. The Associated Press reported that Altman made no reference to the company's security problems during his keynote.

OpenAI has built controls around this: custom rules governing what needs approval, an activity view covering background work, monitoring that can pause or stop a dot mid-task, and sensitive actions such as password changes reserved for you. Its own page carries the caveat: "Dots can still make mistakes, so always review consequential work."

Why it matters: If your organization buys this, someone has to answer a question it has never faced: what is an employee's software allowed to do without asking? Dots arrive with those boundaries set to defaults chosen by OpenAI, not by you, and the consequential ones govern background actions you will never see happen. Two steps before anyone in your shop switches one on. Check the training setting, because business, enterprise and education workspaces are excluded by default but on personal plans it is a control you have to go and set yourself. And decide, in writing, which categories of work a dot may complete versus merely prepare for a human. OpenAI has already drawn that line on your behalf. It is unlikely to be where you would have drawn it.

Sources: OpenAI · OpenAI DevDay recap · The Associated Press · CNBC

---

What's in the Lab

New announcements from major AI labs

OpenAI Proposes Airplane-Style Safety Checks Before Advanced Training Runs

OpenAI published draft guidelines proposing that companies build formal "safety cases" before continuing the advanced reinforcement learning runs that shape how models behave. Safety cases are structured, evidence-based documentation of the kind used in aviation and nuclear power. The proposal covers reviewing training data, running alignment tests during training, watching for systems that can tell when they are being evaluated, and securing the infrastructure itself. The framework is explicitly unfinished; the company is asking for outside feedback rather than announcing a policy.

Why it matters: It landed in the same week that the industry signed a voluntary accord at the White House and told the president no new legislation was needed. Together those are the shape of the deal on offer: the companies draft the safety rules, invite comment, and government stays out of it. Nothing here binds anyone and nothing changes about the tools you use today. What it shows you is the document the eventual rules will be argued from, and who is holding the pen.

Sources: OpenAI

---

What's in Academe

New papers on AI and its effects from researchers

Students Used ChatGPT to Check Answers, Not Sharpen Thinking, Study Finds

A small pilot study (14 students) tested using ChatGPT as a "thinking partner" in an undergraduate statistics course over one semester. Students improved on statistical learning and self-reported AI literacy, but showed no gain on a standardized critical thinking test. Researchers found most students used the chatbot to check their answers rather than to argue with it or explore ideas—treating it as a validator, not a debate partner, despite the course being designed to encourage the latter.

Why it matters: The findings suggest that simply handing students an AI chat window doesn't teach them to think more critically. The typical student used it to confirm answers instead. That's a caution for any employer or school assuming AI tools build reasoning skills on their own.

Sources: arXiv

---

People Rated an AI Debate More Sincere Than Human Discussion, Pilot Finds

Researchers exposed 120 adolescents and 25 adults to a live, moderated debate between AI personas. Adult participants rated the synthetic debate far more sincere than comparable human discussion, despite understanding generative AI's risks. The researchers report a very large effect size. The same exposure also scrambled participants' prior opinions on the debated topics and increased their appetite for AI regulation, a reaction the researchers dubbed a "Normality Paradox." The sincerity finding comes only from the small adult pilot.

Why it matters: Polished AI debate performances may read as more trustworthy than real human argument. That has unsettling implications for anything from AI-generated news commentary to corporate messaging tools built on multi-agent systems.

Sources: arXiv

---

Gender Bias in AI Chatbots Runs in Opposite Directions Depending on the Model

A study of ten AI chatbots from nine companies, released between April 2025 and June 2026, tested gender bias two ways: which gender models assumed wrote stereotyped phrases, and whether they judged harm to women or men more harshly. Bias proved common but inconsistent. Some models attributed masculine-stereotyped writing to women more often, others did the reverse. Several models were more willing to accept harming a man than a woman to avert a catastrophe. The conditions that triggered this varied model to model, and three models showed no variation at all.

Why it matters: Each model has its own quirks. Companies using AI for hiring, content moderation, or writing assistance can't assume switching vendors fixes fairness problems, or predict which direction bias runs. The authors recommend ongoing audits across vendors, not one-time checks.

Sources: arXiv

---

What's Happening on Capitol Hill

Upcoming AI-related committee hearings

Wednesday, September 30 — Hearings to examine rogue AI, focusing on securing the homeland against AI agents. Senate · Senate Homeland Security and Governmental Affairs Subcommittee on Disaster Management, District of Columbia, and Census (Open Hearing) 342, Dirksen Senate Office Building

---

What's On The Pod

Some new podcast episodes

The Ezra Klein Show — Bill Gates's Blunt Warning on A.I.

The Cognitive Revolution — Obsolete or Irreplaceable? Garrison Lovely on Stopping the Race to Replace Human Labor