{
  "id": 8078406,
  "title": "Trump just explained why an AI agency cannot work. He is right",
  "url": "https://urgent.news/2026/09/17/trump-just-explained-why-an-ai-agency-cannot-work-he-is-right",
  "topic": "ai",
  "section": "AI",
  "published": "2026-09-17T17:43:15.000Z",
  "source": {
    "name": "Fortune",
    "slug": "fortune",
    "url": "https://fortune.com/2026/09/17/trump-ai-agency-independent-standards-board-openai-anthropic/"
  },
  "original_language": "en",
  "account": "On Monday, President Trump suggested that Artificial Intelligence (AI) can only be safely managed with \"a strong and smart (High IQ!) President.\" Washington and the labs interpret this as an obstacle, but it serves as a warning instead. Since the Supreme Court's decision in Trump v. Slaughter in June, all federal agencies, except the Federal Reserve, report to the White House. An AI regulator would be subject to the same oversight. Bernie Sanders advocates for a pause until a new cabinet agency is established, while John Thune proposes an office to block model release and an implementer with oversight credentials. Treasury aims for a FINRA for AI, reporting to a Treasury official. Each proposal entails a federal official answerable to the President, who can dismiss them as needed. The labs are aware of this, which is why Dario Amodei, Sam Altman, and Elon Musk have expressed their desire for rules and their apprehension over potential regulation. They possess a better solution that lies within their own resources. Every position in this debate presumes that governance commences when government intervenes. However, the evidence contradicts this notion. Prior to releasing AlphaFold, DeepMind sought input from over thirty external experts in biology, biosecurity, bioethics, and human rights, and their guidance influenced the model's release. The majority of major labs now publish a safety framework, outlining the capabilities that would preclude a launch, and since January, California mandates that the largest models publish this framework and report serious incidents within 15 days. Every model comes with a system card, akin to a drug's package insert, detailing its limitations and areas of failure. External parties now assess models prior to release. The UK's AI Security Institute received OpenAI's GPT-5.5 before its launch, identified a universal jailbreak, and disclosed the findings. METR, an independent evaluator, was provided with unreleased models from four labs simultaneously for a month this spring. In April, Anthropic concluded that its most advanced model posed too significant a risk for general release, confining it to a limited group of cybersecurity partners and only expanding access in July after the federal government's approval. Microsoft established an internal government, with a Responsible AI Standard overseen by a committee of senior executives and enforced by an Office of Responsible AI through champions embedded within every engineering team. Over 1,900 cases have been reviewed since 2019, with 450 occurring in the past year, ranging from precedent setting to escalated matters to the chief executive. One instance involved refusing to deploy real-time facial recognition on police body cameras. A board committee oversees all these processes. The author, in their book about corporate governance prior to the enactment of laws, discovered that governments typically assume governance begins when government acts. However, the records indicate the opposite. Prior to releasing AlphaFold, DeepMind consulted more than thirty external experts in biology, biosecurity, bioethics, and human rights, and their advice shaped the release. Every significant lab now publishes a safety framework naming the capabilities that would halt a launch, and since January, California requires the largest models to publish this framework and report serious incidents within 15 days. Every model is accompanied by a system card, the equivalent of a drug's package insert, listing what it cannot do and where it fails. External parties now test models before release. Britain's AI Security Institute evaluated OpenAI's GPT-5.5 before its release, discovered a universal jailbreak, and published the results. METR, an independent evaluator, was given a month this spring with unreleased models from four labs simultaneously. In April, Anthropic concluded that its most capable model was too dangerous for general release, confining it to a select group of defensive cybersecurity partners, and only expanded access in July after the federal government's approval. Microsoft built an internal government, with a committee of senior executives drafting its Responsible AI Standard, and an Office of Responsible AI enforces it through champions embedded in every engineering team. Over 1,900 cases have been reviewed since 2019, with 450 in the past year, ranging from precedent setting to escalated matters to the chief executive. One instance involved refusing to deploy real-time facial recognition on police body cameras. A board committee oversees all of these processes. The missing element is a written standard that outlines the requirements that frontier laboratories must meet, along with a corps of auditors who verify that each laboratory adheres to the standard.",
  "summary": "AI companies already have internal safety systems. What they need is independent standard-setting and auditing—not a White House-controlled regulator.",
  "key_points": [],
  "editors_take": null,
  "illustration": null,
  "coverage": {
    "outlets": 2,
    "also_reported_by": [
      {
        "outlet": "Business Insider",
        "title": "Effective altruism explained: The AI-linked giving philosophy beloved by Silicon Valley and opposed by Trump",
        "url": "https://urgent.news/2026/09/17/effective-altruism-explained-the-ai-linked-giving-philosophy-beloved",
        "published": "2026-09-17T14:19:02.000Z"
      }
    ]
  },
  "ai_generated": true,
  "disclaimer": "Summaries, key points and the editor’s take are written by software from other outlets’ reporting and may contain errors — always check the linked original."
}