Executive Summary
Word came out that President Trump and President Xi had agreed not to leave large-scale military operations to AI alone. The person who said so was not either leader but Scott Bessent, the US Treasury Secretary, and the setting was not the summit room but a CNBC interview on November 20. The two had met the day before at the APEC summit in Shenzhen, China.
The joint statement issued at that same meeting contains nothing of the kind. What reporters described in it is UN reform, expanded fuel supply, and generalities about nuclear risk. The clause placing decisions on nuclear weapons under human control, which Biden and Xi put into an official White House readout at the 2024 APEC summit in Lima, is also absent this time. Less is on paper in 2026 than in 2024.
Sections 1 through 3 report what the coverage establishes. Section 4 is the Pebblous reading. A claim that a person was involved can be confirmed only when that involvement leaves a record, and the condition holds far beyond a summit room.
Key Figures
Three numbers summarize this case. How much of the joint statement went to military AI, the last year the nuclear clause made it into an official text, and the number of actions a Senate bill names as human responsibility.
Sources: The Epoch Times, SCMP, NPR (2024 Lima APEC), govinfo (text of S.4682).
Zero lines
Military AI language in the joint statement
No document to point to, only the broadcast remark of a cabinet member who sat in the room
2024 only
The year human control of nuclear decisions appeared in a document
Written into an official White House readout, then absent from the 2026 Shenzhen statement
Three
Actions a Senate bill uses to define human responsibility
The ability to supervise, intervene in, and terminate, set down in statutory language
The Agreement Surfaced in a CNBC Interview
Scott Bessent, the US Treasury Secretary, was the first to make this public. On November 20 he told CNBC that the two leaders had agreed to keep a person inside every military decision, and not to let AI run a large-scale military action on its own. He said "SI" rather than AI. It is the Trump administration's shorthand for super intelligence, and what it points to here is AI in the ordinary sense.
"I think it was very important that President Trump and President Xi agreed that humans should always be involved in any military decisions that SI makes, and the SI should not be allowed to be autonomous for major military actions," Bessent said. Two promises sit in that one sentence, joined by a single conjunction: that people stay involved, and that major military action is not handed over. He presented both as something the two leaders agreed to.
The leaders had met a day earlier. The APEC summit ran from November 18 to 19 in Shenzhen, China, and the US-China bilateral on the 19th was followed by a 90-minute three-way meeting that added Russian President Putin. It was the third time Trump and Xi had met face to face in 2026.
The route of the announcement sets the character of this agreement. It was not a joint press conference by the two leaders, and it was not an official White House readout. It was a remark made on television the next day by one cabinet member who sat in the room. Which military systems are covered, what human involvement means in concrete terms, and how anyone would find out that the promise had been broken: none of that came from either side. Reuters, Bloomberg and other outlets noted the same gaps.
A joint statement did come out of the three-way meeting. The contents reported from it carry no AI language. They cover UN reform, expanded fuel production and supply, immunity from the International Criminal Court for each country's own officials, and a general pledge to reduce nuclear risk. This article could not open the statement itself, and the list of items rests on reporting that relayed it.
In 2024 It Was Written Into a White House Document
Two years ago the same kind of promise was written down. Biden and Xi met at the APEC summit in Lima, Peru in November 2024, and the official White House readout carried a line saying that decisions on the use of nuclear weapons should stay under human control rather than with AI. It was recorded as the first case of the Chinese side publicly agreeing to such a statement.
That clause is not in this year's statement. Language about reducing nuclear risk survived, but the sentence pinning the decision to people is gone. What 2024 put on paper has dropped out of the 2026 text, and the military AI agreement reported this time never passed through a document at all.
There are four places to compare the two years. Who met, who the document says holds nuclear decisions, where the human-involvement promise on military AI was recorded, and whether a way exists to confirm that the promise is kept. The first three moved over two years, and the last one is empty in both.
| 2024 APEC, Lima | 2026 APEC, Shenzhen | |
|---|---|---|
| Leaders at the table | Biden, Xi | Trump, Xi, Putin (three-way meeting) |
| Human control of nuclear decisions | Stated in the official White House readout | No such sentence in the joint statement |
| Human involvement in military AI | No separate language | Relayed only in the Treasury Secretary's broadcast remark |
| Way to confirm compliance | Not defined | Not defined |
▲ Pebblous original compilation. The 2024 entries draw on NPR's reporting of the White House readout, and the 2026 entries on coverage of the Shenzhen meeting by several outlets.
The last row, empty across both years, is worth a separate look. Even on the 2024 clause, the US policy group Center for AI Policy pointed out that it was not binding and that a follow-up formal agreement was not guaranteed. Writing something down and confirming that it was honored are separate steps, and both summits stopped at the first one. In 2024 at least one line got written; in 2026 not even that.
Expert assessments point the same way. Zhao Tong of the Carnegie Endowment for International Peace said the contrast between the ambition Trump has declared for nuclear arms control talks and the thin content of the statement could not be starker, and that the outcome fell far short of expectations. On North Korea's nuclear expansion and on non-proliferation questions around Iran, the statement carries no specific commitment either.
What could not be confirmed belongs here too. The text of the three-way joint statement and the full articles from some outlets could not be opened directly, so the list of statement items and Zhao's assessment were cross-checked against other outlets reporting the same content. Bessent's remark was carried in identical wording by several outlets.
The Senate Has a Bill That Tries to Write a Checkable Standard
An attempt to move the same problem into statutory language is under way in the US Congress. Senator Mark Kelly, a member of the Senate Armed Services Committee, introduced two bills on June 5, 2026. One covers ultimate human responsibility in defense systems (S.4682), and the other, the WARP Act (S.4683), covers the military's readiness to use AI.
What the first bill sets out to do is define one phrase. For autonomous weapons systems, it treats the ability of commanders and operators to supervise, intervene in, and terminate as ultimate human responsibility. Unlike a spoken promise that humans are involved, the bill writes down in three actions what someone must be able to do for involvement to count. Both bills are pending in the Armed Services Committee.
The difference has less to do with binding force than with whether anything can be checked. Whether a system can be supervised, intervened in, and terminated is something you settle by inspecting it. A single sentence saying that a person was involved does not even settle what to examine. Whether the bills pass is a separate matter, and what is worth noting here is that a method for turning the same promise into a checkable form has already been proposed.
Why Pebblous Is Watching This Case
A promise between states and an approval step inside a company differ in scale, but they snag on the same point. Is there a way to confirm afterward that a person was involved? The gap Bessent's remark leaves is not in what was agreed; it is in the absence of any record that could confirm it.
Inside companies the same question now comes up in daily operations, as AI agents begin placing orders, approving refunds, and changing account permissions. When an audit or an incident review opens, the first explanation offered is that a person made the call, not the AI. The question that follows is always the same. Who approved it, when, and on the basis of what they saw?
Answering that requires deciding in advance what gets written at the moment of approval. The industry converges on four items.
- The identity of the approver and the time of approval. If several people share one account, this field is as good as blank.
- What the agent proposed, along with the screen or summary the approver actually saw. Without that, an approval survives as nothing more than a click.
- The policy and the model version in force at that moment. The same action can be permitted under one rule and refused under another.
- Evidence that the record was not edited afterward. A log that can be changed after the fact does not serve as evidence in a dispute.
The separation principle the industry calls maker-checker rests on this record as well, since the separation holds only when the record shows that the party proposing and the party approving were different. Without the record, what remains is a sentence. A statement saying that people stay in control, and an internal policy saying that every automated approval passes a reviewer, are the same kind of sentence once there is no way to trace either one back.
Put the two routes side by side and the point where they diverge becomes visible. A promise carried by speech stops at the memory and the phrasing of the person who carried it, and an approval kept as a record can be opened again by someone else months later.
What this case shows is that even an agreement at the highest level falls outside anything checkable once it skips the record. Confirming this inside an organization is simple. Take one item an AI agent handled last month that is marked as approved by a person, and trace whose screen that approval happened on and at what time. If the record does not reach that far, that organization's human control sits in the same place as a sentence in a joint statement.
References
News Coverage
- 1.The Epoch Times. (2026). "Trump and Xi Agree Military AI Decisions Should Require Human Involvement, Bessent Says."
- 2.Deccan Herald. (2026). "US, China Agree AI Must Not Make Military Decisions Without Human Oversight."
- 3.South China Morning Post. (2026). "Concrete Aims or Just Symbolism? Trump, Xi and Putin Talk Nuclear Risk but Controls Stall."
- 4.South China Morning Post. (2026). "Xi, Trump Agree Military AI Decisions Must Involve Humans, Bessent Says."
- 5.Bloomberg. (2026). "AI Military Use Restricted by Trump, Xi Pact, US Treasury's Bessent Says."
- 6.NPR. (2024). "Biden and Xi Take a First Step to Limit AI and Nuclear Decisions at Their Last Meeting."
Legislation & Policy
- 7.Center for AI Policy. (2024). "Biden and Xi's Statement on AI and Nuclear Is Just the Tip of the Iceberg."
- 8.Office of Senator Mark Kelly. (2026). "Kelly Introduces Bills to Smartly Integrate AI into U.S. Military." Press release.
- 9.United States Congress. (2026). "Ultimate Human Responsibility in Defense Systems Act of 2026." S.4682, 119th Congress.