Updated October 19, 2026. This article follows China's July 2025 proposals, subsequent official reporting, and the separate United Nations process.
International AI policy sounds remote until you think about what a system might translate, recommend, refuse to explain, or get spectacularly wrong. Then the conference-room language starts following you home. I can appreciate a beautiful sentence about cooperation, but I also want to know what happens after the microphones are switched off and everyone has located the correct charging cable.
China's call for a global AI consensus belongs in that space between a promising principle and the arrangements needed to make it useful. It raises questions about who helps write standards, who can afford the technology, how systems are evaluated, and how much authority governments should exercise over their outputs. Those questions deserve attention beyond a contest between two national flags.
The original moment was July 2025, when China proposed a new international organization and the Shanghai AI conference released a governance action plan. The story has developed since then. Understanding it now means separating that proposal, the later founding agreement reported by China, and the UN's own institutions. Each has a different purpose, and a different claim to authority.
What the 2025 proposal actually offered
The Global AI Governance Action Plan was released on July 26, 2025, at the World AI Conference and High-Level Meeting on Global AI Governance. Its published text emphasizes international cooperation, national sovereignty, development, inclusion, and safety. It also advocates expanding computing infrastructure, supporting developing countries, and coordinating approaches to governance.
That is an agenda for expanding AI as well as managing its risks. Describing China's position simply as asking the world to slow down misses the development side of the proposal. The plan envisages wider applications and greater access, with cooperation intended to help countries participate in the technology's growth.
In accompanying remarks, China's foreign ministry described the proposed World AI Cooperation Organization, usually abbreviated WAICO, as a platform for innovation, capacity building, and coordination of rules and standards. Shanghai was being considered as its headquarters. The ministry also said the organization should complement the UN's role. These were China's stated objectives, rather than evidence that other governments had already accepted every arrangement.
I find it useful to read a diplomatic document at two levels. First comes the offer: what participants are being invited to do. Then come the arrangements: who would decide, who would contribute, who could disagree, and what decisions would actually require of members. A persuasive description of the destination leaves plenty to investigate about the route.
What changed in 2026
On July 20, 2026, China's foreign ministry reported that 29 countries had signed an agreement establishing WAICO during that year's Shanghai conference. It described the signatories as founding members from Asia, Africa, Latin America, and Europe. That official report moves the timeline beyond the original announcement of an idea.
The scope of the claim still matters. A reported founding agreement establishes a more concrete development than an invitation to discuss one. It does not, by itself, tell a reader every membership obligation, voting arrangement, budget commitment, or domestic implementation requirement. I would want the agreement and subsequent operational records before making claims about those details.
It also leaves room between international participation and worldwide consensus. Countries may cooperate on training or infrastructure while disagreeing about speech, security, market access, and state control. The practical question is where agreement has been documented and what it covers. Calling something global should begin a conversation about representation, rather than finish it.
Competition and cooperation can occupy the same room
The United States released its own AI Action Plan on July 23, 2025. The White House organized it around innovation, infrastructure, and international diplomacy and security. Its announcement included plans for exporting packages of American AI technology and standards to partner countries, accelerating infrastructure construction, and removing regulatory obstacles it considered burdensome.
Read alongside China's proposal, that gives us two public accounts of how influential governments wanted AI to develop. Both involved international relationships and the infrastructure behind the software. The documents differ in emphasis and language, but neither supports treating one country as interested only in commercial power and the other as interested only in generosity.
My interpretation is that cooperation can be useful to a government and increase its influence at the same time. A country offering training, technical standards, or access to infrastructure can address a genuine need while shaping future relationships. Assessing the offer means looking at its conditions and results. Assuming a noble motive or a sinister one in advance saves time mostly by skipping the evidence.
Sovereignty brings difficult questions with it
China's emphasis on sovereignty deserves close reading because international coordination still meets domestic rules. Its 2023 interim measures for generative AI services illustrate that tension. The measures cover services offered to the public within China, with an exemption for specified development and applications that are not offered to that public. They include requirements concerning socialist core values, prohibited content, privacy, and discrimination.
Those provisions show why broad promises about inclusion leave important questions open. A government may support wider access to AI while requiring particular content restrictions. For a reader concerned with political expression, a commitment to cooperation needs to be examined alongside the rules governing what a service may produce and what happens when it crosses a boundary.
I would ask the same practical questions of any proposed framework. Who defines the unacceptable content? Can a restriction be challenged? How are privacy and security claims assessed? How much explanation reaches the affected person? Those questions make the discussion more useful than sorting whole countries into a reassuring column and a frightening one.
The UN has its own process
A separate development came through the UN General Assembly's resolution 79/325, adopted in August 2025. It established an Independent International Scientific Panel on Artificial Intelligence and a Global Dialogue on AI Governance. The resolution describes the panel's role as producing scientific assessments and the dialogue's role as sharing experience and discussing cooperation. These are distinct from WAICO.
The inaugural annual dialogue opened in Geneva on July 6, 2026, with sessions on July 6 and 7. ITU's account describes participation by governments, companies, researchers, civil society, and the technical community. It also reports differing consultation priorities: governments ranked capacity building first, while most other stakeholder groups prioritized safety. Even within a shared process, participants brought different needs.
That difference is worth holding onto. It shows why agreement on the value of international discussion does not settle what should happen first. A research organization, a ministry seeking infrastructure, and a community affected by a deployment may all reasonably focus on different parts of the problem. A useful institution needs a way to make those priorities visible.
The UN describes the dialogue as a forum rather than a regulatory or enforcement body. That distinction helps set expectations. A forum can build understanding and inform policy, while enforcement requires an applicable authority and a concrete mechanism. Measuring a discussion platform by whether it can immediately compel every company on Earth would confuse its role with a different kind of institution.
Access means being able to use the technology
The case for broader participation becomes easier to understand when we imagine an actual project. Suppose a small education organization wants a system that helps teachers prepare lessons in a local language. Access would need to include a usable service, appropriate learning materials, people who can evaluate the results, and a way to keep the project running. An invitation to a conference would address only one part of that picture.
That example is hypothetical, but it helps me think about promises of capacity building. I would look for a named partner, a clear purpose, training that reaches the intended users, and terms the organization can understand. I would also want to know what happens when funding ends or the original provider changes its service. The ability to continue matters alongside the ability to begin.
There is a practical distinction between receiving a tool and helping decide how it should work. If teachers identify problems in the output, their findings should have somewhere to go. If a public institution cannot afford an update, that should be visible in the project's assessment. Participation becomes more meaningful when the people doing the work can influence the next decision.
Evaluation should follow the intended use
NIST's AI Risk Management Framework offers a useful example of guidance already available. It is intended for voluntary use and treats trustworthiness as something to consider throughout design, development, use, and evaluation. Its 2024 generative AI profile addresses risks including fabricated information, privacy, harmful bias, and misleading content. NIST's framework webpage also notes that the original framework is being revised, a reminder to check versions.
For me, the next step is to ask what evidence an evaluation would supply. A system used to draft a restaurant's promotional copy has different consequences from one used to support a consequential decision about an individual. I would want the testing conditions, intended users, relevant limitations, and response to failures explained in terms appropriate to that use.
Imagine a pilot in which an education tool handles routine classroom questions well but struggles with local place names. An overall score could conceal a problem that matters a great deal to its users. Asking for the test results behind the score, including the situations that caused difficulty, would make the findings more useful. This is an example of what I would want disclosed, rather than a report of a particular pilot.
Language belongs inside the assessment
NIST's generative AI profile also discusses how weaker performance in languages with less available data can affect accessibility and language preservation. That gives substance to a concern that can otherwise stay abstract. A system described as multilingual needs to be evaluated for the languages and communities expected to use it, with the limits of that evaluation made clear.
I would want fluent speakers involved in designing that assessment and interpreting the results. For an educational application, they could help examine whether an explanation is accurate, whether the vocabulary fits its audience, and whether local examples make sense. Those are questions I would put to a project, not capabilities I would assume from its marketing.
Working with language has taught me to appreciate how much meaning can live in a small choice of words. Wine descriptions alone can turn one person's useful explanation into another person's elaborate grocery list. The stakes are different in education or public services, but the underlying lesson travels: an answer should make sense to the person who needs it.
What I would count as progress
There are shared principles to build on. UNESCO adopted its Recommendation on the Ethics of Artificial Intelligence in November 2021, emphasizing rights, dignity, fairness, transparency, and accountability. The existence of that framework is useful context for a newer proposal. International AI governance did not begin with a single announcement in Shanghai, Washington, or Geneva.
What I would look for now is documented follow-through. A training initiative could publish its purpose, participants, resources, and assessment. An evaluation effort could explain its methods and report limitations. A governance forum could make the scope of its decisions and the route for public input easier to understand. Each would let people compare a stated commitment with a result.
I would also pay attention to what changes after criticism. A process becomes more credible when participants can point to a problem, get an explanation, and see an appropriate response. That takes time and work, but it gives words such as accountability and inclusion a practical meaning. The most encouraging update might be a clear record of a problem being fixed.
A future worth explaining clearly
China's initiative deserves scrutiny because the questions it raises are consequential and because a proposal can develop into institutions with lasting influence. The same scrutiny belongs with American initiatives, UN mechanisms, and the companies supplying the systems. I want to know what each can do, who participates, and which commitments a reader can actually check.
I definately want international cooperation to produce something useful. I want smaller organizations to understand the tools they adopt, more communities to help shape the technology they use, and ordinary people to have somewhere to turn when an automated answer affects their lives. Those are outcomes I can care about without pretending that every participant agrees on everything.
The new code is being discussed in conference rooms, but its consequences will arrive in much more familiar places. I want the conversation to reach those places too, in language people can understand and with evidence they can examine. That would give consensus a purpose beyond looking handsome in a press release.
Related Reads You Might Enjoy
How AI Is Learning to Feel Pain and What That Means for Humanity
Artificial Photosynthesis Could Power the Future, And It’s Closer Than You Think
AI is Already Outperforming Humans in Image Analysis, Here’s What That Means for All of Us
The AI That Writes Its Own Rules: Inside DeepMind’s New Era of Algorithmic Creation
When AI Eats the Grid: Why Artificial Intelligence Might Outconsume Bitcoin by 2025
Digital DNA: Are We Building Online Clones of Ourselves Without Realizing It?
Claude 4 Begged for Its Life: AI Blackmail, Desperation, and the Line Between Code and Consciousness
Sources and further reading
China: 2025 explanation of the WAICO proposal and its intended relationship with the UN
China: July 20, 2026 official report of the WAICO founding agreement
White House: America’s AI Action Plan announcement, July 23, 2025
UN mandate registry: resolution 79/325, Scientific Panel and Global Dialogue
ITU: opening of the UN Global Dialogue, July 6, 2026, and consultation priorities
NIST: voluntary AI Risk Management Framework and revision status
NIST: 2024 Generative Artificial Intelligence Profile, including language and information risks
UNESCO: Recommendation on the Ethics of Artificial Intelligence
The conference room, computing facility, and classroom images are AI-generated editorial illustrations. They do not depict an actual meeting, named institution, or documented cooperation project.
