Khanna tells CBS Trump–Xi must lead with an AI guardrails pact
On Face the Nation, Rep. Ro Khanna said Trump and Xi should put an enforceable international AI agreement first — banning recursive self-improving AI, blocking bio/nuclear misuse, and creating an independent FDA-style AI agency — and that Chinese firms have not answered his slowdown letters.
POLITICS desk — Khanna put enforceable AI limits on the Trump–Xi agenda before this week’s summit. The Sunday-show ask is the news. A signed pact is not.
Rep. Ro Khanna, a California Democrat and the top Democrat on the House Select Committee on the Chinese Communist Party, appeared Sunday on CBS’s Face the Nation with Margaret Brennan. He joined from Fremont. The CBS Face the Nation interview transcript of that 20 Sep 2026 airing is the filing event. These are Sunday-show remarks. They are not a treaty, not a signed Trump–Xi pact, not a new statute, and not a standing FDA-style AI agency.
Brennan asked whether any Chinese firms had answered his letters to DeepSeek, Alibaba, and other companies urging them to follow OpenAI and Anthropic and slow AI advancement until society builds guardrails. Khanna said: “They have not.” He added: “Winning an AI war is like winning a nuclear war. It is catastrophic for civilization.” He said the two dangers are “the loss of control or misuse,” and that “the Chinese firms have taken no actions on either of them.” File those lines as Khanna’s, via the CBS transcript. This desk did not send the letters or hear a company reply.
On FBI, NSA, and CISA warnings that China-based AI companies are copying U.S. models at industrial scale, Khanna said they are “engaging in distillation” — “copying American models.” Distillation here means training a new model by copying the behavior of an existing one. He said that is “one reason, frankly, to regulate American models,” because “some of the safeguards would protect the world.” He also said the U.S. still needs “an international agreement with Xi Jinping to have basic standards.” File distillation and the regulate-American-models line as Khanna’s. This desk did not audit a Chinese model.
That agreement, Khanna said, “should include international monitoring,” “checks through the data centers, where you can tell whether a model is being used for inference or training,” and “enforceable and verifiable tools.” Inference is when a finished model answers a query. Training is when a model is still being taught. He compared the verification job to arms control: “as Reagan said, you need to trust and verify.” File the monitoring / data-center / trust-but-verify picture as Khanna’s ask — not as a signed inspection regime.
His first ask: “Trump and Xi Jinping to come up with an international agreement.” Two core principles, as he named them: “banning any self-improving AI, recursive self-improving AI” — “AI that continues to improve itself and where we could lose control” — and “safeguards so that AI can’t be misused by people for biological nuclear weapons.” Recursive self-improving AI, in his wording, is a system that keeps rewriting itself without a human stop. File both principles as Khanna’s Sunday-show ask. They are not a White House proposal and not a finished communique.
He said the letters to Chinese companies “were simply seeking information about what they’re doing.” Quoted: “It wasn’t intended as a solution.” The point, he said, was to “highlight the need for the Chinese government and the U.S. government to come to an agreement, and that needs to be the first agenda item on the Trump-Xi Jinping summit.” File first-agenda-item as Khanna’s ask. Do not read this card as the summit schedule.
Asked about President Trump’s Saturday plan to create an AI Force and name an AI czar — already filed as trump-ai-force-czar — Khanna said he does not know Trump’s vision and does not want AI used as an “offensive capability.” What he wants: “an AI agency like the FDA.” FDA is the U.S. Food and Drug Administration, the independent federal body that reviews medicines and medical devices before they ship. He cited Geoffrey Hinton — “the godfather of AI” — saying AI can be regulated “like we regulate nuclear energy, electricity, aviation,” with “independent technical experts.” Quoted: “This is not complicated. We need to have an independent federal agency.” File the FDA-style agency as Khanna’s ask, not as a bill that passed and not as Trump’s AI Force.
Plain English for the rest of the card: Face the Nation = CBS’s Sunday interview show. recursive self-improving AI = Khanna’s label for a system that keeps improving itself, where people could lose control. distillation = copying an existing model’s behavior to train a new one. inference = using a finished model. training = teaching a model. guardrails = agreed limits meant to block loss of control or misuse. FDA-like AI agency = an independent federal reviewer, in Khanna’s analogy, not a military AI Force. trust but verify = Reagan’s arms-control line: do not take a counterpart’s word without checks. These are Sunday-show asks. A pact would have text, signatures, and inspections. None of those were published.
PRIMARY here: CBS’s 20 Sep 2026 Face the Nation interview transcript of Rep. Ro Khanna — Tier A PRIMARY, the originating CBS newsroom record of the Sunday airing — plus the same-day full-show Face the Nation transcript, same-event corroboration of the same remarks, not a second interview. The no-reply line, the nuclear-war analogy, the loss-of-control / misuse pair, the distillation / regulate-American-models line, the international-agreement / monitoring / data-center / trust-but-verify ask, the two core principles, the first-agenda-item line, and the FDA-like independent federal agency ask are Khanna-attributed via that CBS transcript. NOT claimed: a signed Trump–Xi pact, a treaty text, a new statute, that an FDA-style AI agency exists, that Chinese firms agreed to slow down, that this desk sat in the studio, a stock tip, or investment advice. Distinct from the already-filed bessent-he-ai-talks, trump-ai-force-czar, trump-rejects-ai-guardrails-amodei, and coxon-meet-the-press-ai-global-coordination.