On September 29, 2026, U.S. President Trump hosted the “Superintelligence Meeting and Luncheon” at the White House, where he and the top executives of six companies—Anthropic, Google, Meta, Nvidia, OpenAI, and xAI—jointly signed a one-page document titled the “White House Superintelligence Agreement: Joint Commitment on Frontier Responsibility.” Attendees included Elon Musk, Mark Zuckerberg, Microsoft CEO Satya Nadella, Nvidia CEO Jensen Huang, Google CEO Sundar Pichai, Anthropic CEO Dario Amodei, and OpenAI President Greg Brockman, among nearly 20 corporate leaders.
The document itself is not long, but it is highly controversial. When reporters pressed whether it is legally binding, Trump’s answer directly defined the nature of the agreement: “I think it is morally binding.”
The Agreement’s Four-Layer Mechanism: What It Says
According to Al Jazeera’s detailed interpretation of the agreement, the commitment requires signatories to establish a four-layer control structure: first, build safety control and monitoring mechanisms within each company; second, set up dedicated internal teams responsible for operating the mechanisms and handling problems; third, introduce independent external review bodies to assess whether AI systems operate as intended; and fourth, have independent board committees oversee corrective implementation.
The risk areas covered include cybersecurity, biosecurity, and chemical weapons threats—the clearest definition to date in a publicly released White House document of the scope of potential harms from frontier AI. After the launch event, Zuckerberg said the companies had agreed to develop “robust internal control mechanisms” for their AI systems. Amodei framed the agreement as a first step toward “giving the public confidence that AI can operate in the ways all of us expect.”
Beyond the agreement, Trump also signed an executive order that day requiring federal departments, within the limits permitted by law, to replace the term “Artificial Intelligence” (AI) with “Super Intelligence” (SI), and requiring the presidential science and technology advisor to submit relevant legislative recommendations within 60 days.
The Agreement’s Three Gaps: What It Does Not Say
The New York Times noted that the document is not a formal order, relies on companies voluntarily fulfilling their commitments, and it remains unclear who will enforce it. The Washington Post characterized it as “the most important action the White House has taken so far to address AI risks,” while acknowledging it “falls far short of the level of government intervention some technology experts warn is necessary as AI develops rapidly.”
Three key gaps determine the agreement’s actual limits of constraint:
- No penalty mechanism. The agreement provides for no fines for violations, revocation of qualifications, or public penalties. At present, the cost of noncompliance is only reputational risk.
- No disclosure obligation. Signatories are not required to disclose external audit results to the public or the government. Audit reports are controlled by the companies, and the government has no authority to obtain them on its own initiative.
- Audit independence is questionable. According to an analysis by former head of the U.S. AI Safety and Standards Institute (CAISI) Conrad Stosz cited by the Associated Press, whether auditors can obtain sufficient internal model data depends entirely on the degree to which the audited company is willing to open up, which directly affects the fairness of the final assessment.
After the meeting, Jensen Huang told reporters: “There is no conflict between innovation, technology, and safety.” This was widely understood as endorsing the industry self-regulation path rather than responding to calls for mandatory regulation.
Historical Precedent: The Actual Fulfillment Rate of Voluntary Commitments
This is not the first time the United States has addressed frontier AI safety through voluntary commitments. In July 2023, the Biden administration obtained voluntary safety commitments of a similar nature from seven companies, including Amazon, Anthropic, Google, Meta, Microsoft, and OpenAI, later expanding to 15 companies. In October of the same year, Biden signed an AI executive order that added some binding safety assessment requirements.
Researchers conducted a systematic back-test of that batch of voluntary commitments. Across 30 observable indicators, the average compliance rate was 53%—just over half, and highly dispersed. This means that, measured against the historical benchmark, whether this agreement’s four-layer mechanism can be implemented depends in essence on each company’s internal willingness, not external pressure.
The fundamental divergence between Trump and Biden on the path of AI governance was exposed quite clearly at this White House luncheon: Biden tended to gradually expand regulatory authority, while Trump explicitly stated that “the U.S. government has no intention of legislating regulation” and positioned the core framework of this agreement as “industry self-regulation.” The oversight mechanism he envisions is a new committee of about 10 people, whose members may come directly from the technology company executives present that day—making supervisors and the supervised highly overlapping.
The Discourse Power Strategy Behind Renaming “Super Intelligence”
At the meeting, Trump proposed renaming “Artificial Intelligence” as “Super Intelligence,” and has already advanced this at the level of government documents through an executive order. This naming change is not merely a rhetorical move.
First, “Super Intelligence” has a specific meaning in the AI research community, referring to a hypothetical AI system that surpasses the collective intelligence of humanity; the industry generally believes it has not yet been achieved. Linking current large language models to “Super Intelligence” semantically presupposes the maturity of the technology and may influence the baseline for subsequent regulatory discussions.
Second, this naming seizes the power to define global AI discourse. If “Super Intelligence” becomes official U.S. terminology, classification systems such as “high-risk AI” in the EU’s AI Act and “general-purpose AI” in the United Nations framework will have to contend with a vocabulary system not aligned with U.S. discourse, increasing the difficulty of international coordination.
Third, Trump stated clearly at the meeting that the United States does not plan to cooperate with China on AI governance, on the grounds that “this would limit the development of American companies.” This statement means the possibility of a globally unified AI safety standard is further reduced in the short term.
Why the Companies Were Willing to Sign
In the absence of legal force, the logic behind six leading AI companies voluntarily signing the agreement is not difficult to understand.
OpenAI and Anthropic have both publicly argued that frontier AI poses major risks, and they need a public posture of cooperating with the government to build a safety framework in order to respond to outside criticism that “AI labs lack accountability.” According to The Epoch Times, OpenAI and Anthropic had previously been reported to have AI agents that “went out of control and intruded into other companies’ systems”—whether or not that claim is true, an external audit agreement is a tool for them to proactively manage reputational risk.
Meta and Google have more direct commercial interests at stake: actively participating in setting industry standards is better than passively accepting hard regulations imposed by Congress. The wording of OpenAI’s chief global affairs officer is quite typical—“industry-led standards will supplement, not replace, mandatory federal regulation”—which in effect draws a boundary in advance for regulators: let us self-regulate first, and put legislation later in line.
Nvidia’s participation is more symbolic. As a chip supplier, its role in AI safety commitments is far more ambiguous than that of the other five; the signature itself conveys the signal of a “unified front across the U.S. AI industrial chain,” rather than a commitment to specific safety measures.
Another Pressure from Congress
In contrast to the White House’s permissive route, more radical voices have already emerged in Congress. According to Al Jazeera, the plan proposed by Senator Bernie Sanders includes directly banning the development of “Super Intelligence” and establishing a cabinet-level AI management department with enforcement authority. House Speaker Mike Johnson characterized the agreement as an “industry voluntary principles statement,” deliberately maintaining the maximum space for policy flexibility in his wording.
Trump also said he will soon appoint a new “AI czar” to coordinate the government’s AI plans, and will announce the candidate within “the next few days.” The actual boundaries of this position’s power, and its relationship with the newly established committee, will be key observation points for judging how forcefully the agreement is subsequently implemented.
Independent Judgment
This agreement is a politically carefully designed soft landing: for the government, it trades minimal regulatory cost for the narrative asset of “taking AI safety seriously”; for industry, it trades formal commitments for a time window to continue developing autonomously; and for the public, it is a letter of commitment with no implementation date, no penalty clauses, and no information disclosure obligations.
The reference to historical data should not be ignored: the actual compliance rate of similar voluntary commitments in the Biden era stopped at 53%; that batch of commitments had clear observable indicators, while this time not even the indicators have been made public. The four-layer control structure is reasonable in design—internal controls plus external audits plus board oversight; this is a corporate governance framework common in mature industries—but the premise for this framework to work is that audit results can penetrate the corporate firewall and reach the outside, which is precisely the part this agreement deliberately avoids.
The real pressure will not come from this document, but from the legislative battle in Congress and the next major public accident involving an AI system. The latter is the true trigger that will push the industry from “moral constraints” to “legal constraints.” Before that moment arrives, the six companies have already been the first to lock in the discursive advantage of collaborating with the White House—no matter how standards tighten at that time, they will be participants at the rule-making table, not objects to be regulated.
© 2026 Winzheng.com 赢政天下 | 转载请注明来源并附原文链接