Executive Summary

Let's be clear first about what this piece does not do. It does not evaluate the ideology of the speech. Whether any country's AI vision is right, or what intentions might lie beneath it, is off the table. What it does measure is one thing: the checkable detail that separates declaration from delivery. Is there an actor to execute the promise? Is the audience named? Is a quantified goal attached? Has it been institutionalized through signature? Can the progress be tracked? These five questions are a neutral instrument you can apply just as easily to a statement from the United States, China, the EU, or Korea.

This yardstick is not something we improvised. Scholars have already operationalized "declaration versus delivery" quantitatively. The AGILE Index (Zeng et al., 2025) measures national AI governance across four pillars, and deliberately separates whether the instruments exist (Governance Instruments) from whether those instruments actually work (Governance Effectiveness) into distinct axes. Tidjon and Khomh (2022) gave the gap its own name — the "principle-implementation gap" — and measured it continent by continent. Signing and declaring is one problem; making it move in the real world is a different one from the start. This report's five checks are the working translation of that academic tradition.

1.1Three Principles

For the analysis to be trustworthy, its materials and procedures have to be transparent. This piece holds to three principles.

• Cite only primary sources. We rely on the full Chinese text from Xinhua, the full English text from CGTN, and official announcements from China's Ministry of Foreign Affairs, and set aside secondary commentary. We confirm what the original said, not what the press said about it.
• Separate fact from inference. We do not speculate about intent — "it's a soft-power play," "they mean it," "there's an ambition hidden inside." Facts in the original are stated as facts; Pebblous's readings are labeled as interpretation and inference, and kept apart.
• Respect copyright. We do not republish the full speech. We excerpt only key passages, a sentence or two at a time, and point readers to the official links for the complete text.

The World Artificial Intelligence Conference (WAIC) first opened in Shanghai in 2018. That first edition was closer to an industry trade show — roughly 72,000 visitors and some 40 participating countries. Eight years later, the ninth edition in 2026 drew more than 100 countries, with over 140 forums, more than 1,100 exhibiting companies, and, for the first time, exhibition space topping 100,000 square meters. Rather than marvel at the growth, read it as checkable change: participating countries went from about 40 to more than 100, and the event's character shifted from exhibition to governance platform.

In pinpointing when governance initiatives began riding on this event, one common error needs correcting. Several secondary sources cite a "2019 Shanghai Declaration" as the predecessor document, but the formal document confirmed by primary sources (China's Ministry of Foreign Affairs, English) is the "Shanghai Declaration on Global AI Governance" of July 4, 2024. That declaration took a collective "we agree" form with no list of signatory states; it then passed through the 2025 WAICO proposal and, in 2026, was institutionalized one rung higher into an intergovernmental body signed by 29 countries. The journey from declaration to organization took about two years — itself one more checkable data point.

Sources: Xinhua · CGTN (2026-07-17), China's Ministry of Foreign Affairs (2024-07-04 · 2026-07-16). Precise year-by-year primary figures for participating countries and exhibitors in 2019–2023 were not fully secured and are marked "not secured."

Computing a growth rate to declare "how fast it grew" is beyond this piece's scope. What remains as fact is the count of participating countries and the point at which governance initiatives were layered on; how to read the meaning of that trajectory is the domain of inference. One thing is clear, though. Over eight years WAIC shifted in character from a place where companies open booths to a place where states sign documents. How this change interlocks with the competitive buildup of China's AI ecosystem we treated further, from a data and post-training angle, in a separate report on the open-weight race.

The speech was built around four broad themes: openness and mutual benefit; risk awareness and safety control; inclusion and mutual respect among civilizations; and solidarity and global governance. Emphasizing any one of them would distort the balance of the original, so we give each of the four equal weight, excerpting a single sentence per theme. For copyright reasons we do not reproduce the full text, quoting only the key passage and pointing to the official link in the references.

3.1Openness and Mutual Benefit

The first theme is to develop AI through openness and cooperation rather than closure. It frames this against the diagnosis that AI is moving from the digital world (数字世界) into the physical world (物理世界).

3.2Risk Awareness and Safety Control

The second theme is a safety perspective: AI must remain a tool humanity can trust and control. It stresses that development and safety have to be seen together.

3.3Inclusion and Mutual Respect Among Civilizations

The third theme is that AI development should proceed inclusively, without eroding the diversity of civilizations. It stresses broad participation, including from developing countries.

3.4Solidarity and Global Governance

The fourth theme is solidarity in global governance, anchored in genuine multilateralism and the role of the UN. The alignment (对接) and coordination of governance rules and technical standards is the theme's keyword.

What makes this speech read differently from an ordinary political statement is that many of its items come attached to an executing actor, a target audience, and a number. We take four representative initiatives apart one at a time, holding to a single principle throughout. A number attached is not the same as a number achieved. Blur the line between a target and a current figure and you will overestimate the delivery.

4.1WAICO: Signatures Settled, Working Machinery Undisclosed

The World AI Cooperation Organization (WAICO) is the biggest delivery headline of this WAIC. But get the facts straight first. The 29 signatures were not announced in the body of the speech; they were made at a separate signing ceremony on July 16, the day before the speech, chaired by Foreign Minister Wang Yi, and reported by Xinhua and CGTN. The full text of the speech contains only the fact that "WAICO was born in Shanghai." Attributing the figure of 29 countries to the speech would be factually wrong.

The confirmed facts are these: 29 signatures, headquartered in Shanghai, styled as an independent intergovernmental organization. But no secretary-general, board, decision rules, budget, or enforcement powers appear in any primary source. Even a prior academic paper on WAICO (arXiv 2606.23860) noted it has "only a name, a host city, and a stated purpose, with no action machinery," and gave it the floor score of 1 on its five-point institutionalization scale. That said, the paper itself flags a limit: it is preliminary coding from before the signing (2026-06-22), so it should not be hardened into "an empty shell" either. There is exactly one accurate position. The entity is settled; the working machinery is undisclosed. A blueprint without a building.

Note: the complete, confirmed list of WAICO's 29 countries varies by outlet in how country names are rendered, and the Ministry of Foreign Affairs has not posted an official roster. The total of 29 is confirmed by primary sources; this piece does not assert the full list of individual country names.

4.2Developing-Country AI Training: 5,000 Slots, Not 5,000 People

The second is AI talent training aimed at developing countries. Here too the wording has to be read precisely. The original says not "5,000 people" but "5,000 slots (名额)." A slot is a quota or seat — closer to a headcount of attendances, in which one person can take part in several programs. In other words, they may not be 5,000 distinct individuals. On top of that, this is a cumulative target over the next five years, and the annual breakdown was not disclosed.

This item is a good example of a delivery detail in that it carries an executing actor, a target audience, and a quantified goal all at once. But the academic standard for measuring effectiveness is not the headcount itself. As Section 5 discusses, whether developing-country capacity building actually works is judged not by "how many were taught" but by four conditions: regulatory capacity, the institutional gap, the local ecosystem, and the structure of participation.

4.3Application Cooperation Centers: Six Regional Blocs

The third is AI application cooperation centers at the level of regional blocs. Six cooperation channels were mentioned, aimed at ASEAN, the Arab League, the African Union, the Community of Latin American and Caribbean States (CELAC), the Shanghai Cooperation Organisation (SCO), and BRICS. The direction of execution is concrete in that the target regions are named. But details such as each center's budget, staffing, or opening schedule are not confirmed in primary sources, so we leave this as "audience named, specifics unconfirmed."

4.4"Mazu" Weather AI: Target of 30 Countries, 7 Operational

The fourth case compresses this report's whole argument. The speech announced it would land and apply the weather early-warning AI "Mazu (妈祖)" in 30 countries. Here, 30 countries is a target. The current number operational, confirmed by primary sources (Xinhua, 2026-07-17), is 7: Pakistan, Ethiopia, the Solomon Islands, Jordan, Sri Lanka, Mongolia, and Djibouti, plus a cloud-based pilot spanning some 40 countries.

Distill the dissection above into a single instrument and the components of a practical model for international AI contribution reduce to five questions. These five do not target any particular country; they are a neutral checklist applied identically to any nation's statement.

• ① Executing actor: Is there an agency or organization that will actually carry out the promise?
• ② Named audience: Is it specifically identified whom the effort is for?
• ③ Quantified goal: Is a numerical target attached?
• ④ Signature and institutionalization: Has it been institutionalized through a document, a signature, or an organization?
• ⑤ Traceability: Can outsiders verify and track the progress?

5.1Signatory Count Is Not a Measure of Delivery

The most important argument behind this yardstick is that the intuition "more signatories means stronger will to deliver" is academically disproven. Trace the signatory counts of international AI summits and they run from 29 countries at the UK's Bletchley Declaration to 11 at the Seoul summit, and by the Paris summit the substantive commitments themselves had thinned. Signatory count and binding force can move in inverse. That comparing an organization's power by member count alone is a misreading becomes clearer in the comparison table below.

Sources: official pages of the UN, European Commission, NIST, OECD (see references). Comparing standing by member count alone (OECD/GPAI 46 vs WAICO 29) is a misreading. On the enforcement axis, the only one with actual "teeth" is the EU AI Office.

5.2The Four Initiatives Through the Five Checks

Holding this yardstick up to the speech's four initiatives yields the following. ✓ means confirmed; △ means partial or target-only; "unconfirmed" means it could not be confirmed in primary sources. The point is not to rank them but to judge the facts as they stand.

Judgments are based on primary sources as of the research date (2026-07-18) and may be updated as more material becomes public. △ means partial fulfillment or a target only; "unconfirmed" means it could not be verified in public sources.

One passage on technical direction stands out. It is the diagnosis that AI is moving from the digital world (数字世界) into the physical world (物理世界). And the keyword of the fourth theme was the alignment (对接) and coordination of governance rules and technical standards. Set those two sentences side by side and the expansion into the physical world and the harmonization of technical standards bind into a single axis.

Up to here is what the speech stated as fact. From here on we label it explicitly as Pebblous's reading and inference. When AI stays in on-screen text and images, and when it actually moves self-driving cars, robots, and industrial equipment, the basis for trust is different. The more AI is deployed into the physical world, the more its trust rests on the lineage and quality of physical data. If it is opaque which data trained a model, where that data came from, and whether its quality was verified, a malfunction in the physical world translates directly into a safety problem. For the "alignment of technical standards" the speech stressed to carry real force, our reading is that those standards must be able to treat data provenance, quality, and lineage as verifiable indicators.

That data governance is the underlying infrastructure of physical-world AI already shows in regulatory moves around the world. The debate over how AI services collect and obtain consent for personal data, for instance, shows that data governance is not an abstract principle but a concrete operational problem. That terrain can be seen more concretely in our piece on data-consent rules for Chinese AI services. And the proposition "from digital to physical" itself points to the same terrain as Pebblous's ongoing work on Physical AI.

So where does Korea stand? First, the facts. Korea is not among WAICO's 29 signatories, and as of the research date no official domestic response to this speech or signing had been confirmed. That does not mean Korea has no substance in international AI cooperation. What public data confirms about Korea is the following.

The National AI Strategy Committee is a presidential body created by restructuring, in September 2025, the National AI Committee first established in September 2024 (we note both dates, since they are separate events). The AI Basic Act took effect in January 2026; an international-cooperation clause is confirmed to exist, though the exact article number is not asserted here due to limits on public lookup. For the "Global AI Hub," the number of participating international bodies (14) is clear, whereas the target headcount, number of countries, and budget for developing-country support are not confirmed in public sources. This point matters. It does not mean the effort is ineffective — it should be read only as a difference in the level of quantification and disclosure.

7.1The Same Five Yardsticks, Turned on Ourselves

This report's five checks are not a tool for measuring China; they are the same tool for any country. So we land the ending not on a verdict but on questions we put to ourselves.

• Does our international AI contribution come with a named executing actor?
• Is that contribution specifically identified as being for a named audience?
• Is a numerical quantified goal disclosed?
• Has it been institutionalized through a document, a signature, or an organization?
• Can outsiders track its progress?

Why Pebblous Pays Attention to This Speech

One attitude runs through this whole report: to look through verifiability rather than ideology. And the view that what finally separates declaration from delivery is traceability runs in exactly the same grain as how Pebblous handles data. The four points below trace that resonance — please read them as an overlap of perspective, not a hasty jump to a sales pitch.

1The Terrain of the Physical World and Data Governance

The proposition in the speech's first theme — that "AI is moving from the digital world into the physical world" — points to essentially the same terrain as Physical AI, a topic Pebblous has long worked on. And the fourth theme's keyword, "the alignment of governance rules and technical standards," leads to the problem that the more AI is deployed into the physical world, the more its trust rests on the lineage and quality of physical data. Our view is that this shares terrain with the problem space Pebblous addresses through DataClinic.

2Visibility and Traceability as Preconditions for Trust

The compute-governance argument of Sastry et al. (2024) is that visibility and traceability are preconditions for governance. Our inference is that this structure maps directly onto the data domain. If the provenance, quality, and lineage of training data are opaque, there is no way to verify after the fact whether a model's internal representations and performance can be trusted. In physical-world AI — autonomous driving, robotics, industrial automation — this problem is a safety problem. Making data quality a verifiable indicator is the substructure of a trust infrastructure, and we offer the view that this is the role DataClinic is built for.

3Five Checks That Also Work for a Company

That a close and powerful market put out delivery initiatives with names and numbers attached is a real reason for data and AI teams to read the international-cooperation terrain coldly. Beyond that, this report's five checks are not a tool for states alone. They convert directly into an internal audit tool: "Do our own organization's AI initiatives come with an executing actor, a quantified goal, and traceability?" The yardstick separating declaration from delivery works the same way for a state and for a company.

4The Same Methodological Stance

Just as indices like the AGILE Index and GIRAI quantify governance as "declaration versus effectiveness," Pebblous measures data quality as a verifiable indicator. It is a resonance of the same methodological stance. If we had to put in one sentence the value Pebblous can offer in this shift, it would be: "to build the trust of AI expanding into the physical world on the verifiable indicators of data lineage and quality." Moving declaration into delivery and building trust into data end up sharing the same question. Can it be tracked?

References

This report does not republish the full text of the speech. Below is the list of primary sources, academic papers, and official materials that grounded the citations; we recommend reading the original speech directly at each official link.

Primary Sources — Full Speech Text and Official Announcements

• 1.Ministry of Foreign Affairs of China. (2026). "习近平在2026年世界人工智能大会开幕式上的主旨演讲（全文）." Full text of Xi Jinping's keynote at the WAIC 2026 opening ceremony (2026-07-17).
• 2.Xinhua. (2026). "习近平在2026年世界人工智能大会开幕式上的主旨演讲." Full Chinese text (2026-07-17).
• 3.CGTN. (2026). "Full text: Xi's keynote speech at the 2026 WAIC opening ceremony." Full English text (2026-07-17).
• 4.Ministry of Foreign Affairs of China. (2026). "世界人工智能合作组织成立大会在沪举行 王毅出席并致辞." WAICO founding signing ceremony, chaired by Foreign Minister Wang Yi (2026-07-16, a separate event the day before the speech).
• 5.Xinhua. (2026). "Update: 29 countries sign agreement on establishing WAICO." (2026-07-16).
• 6.CGTN. (2026). "29 Countries Sign Agreement on Establishing WAICO." (2026-07-17).
• 7.Ministry of Foreign Affairs of China. (2024). "Full text: Shanghai Declaration on Global AI Governance." (2024-07-04). The direct predecessor of WAICO — not the "2019 Shanghai Declaration" (correcting a common error).
• 8.Xinhua. (2026). ""妈祖"出海，中国气象智能预警方案走向世界." Overseas rollout of the "Mazu" weather early-warning AI (2026-07-17).

Academic Papers (arXiv)

• 9.Zeng, Y. et al. (2025). "AGILE Index 2025: Annual Governance of AI Global Assessment." arXiv:2507.11546. Four-pillar governance methodology — separating Instruments from Effectiveness as distinct axes.
• 10.Tidjon, L. N. & Khomh, F. (2022). "The Different Faces of AI Ethics Across the World: A Principle-Implementation Gap Analysis." arXiv:2206.03225. Origin of the "principle-implementation gap" concept, measured continent by continent.
• 11.Sastry, G., Heim, L., Belfield, H. et al. (2024). "Computing Power and the Governance of Artificial Intelligence." arXiv:2402.08797. The visibility–traceability–enforcement three-function frame.
• 12.Okolo, C. T. & Raji, I. D. (2026). "The Global Majority in International AI Governance." arXiv:2601.17191. The Global AI Divide and the inclusion yardstick.
• 13.Azin, M. & Zandhessami, H. (2025). "Strategic Alignment Patterns in National AI Policies." arXiv:2507.05400. Analysis of declaration–execution alignment.
• 14."WAICO: Mapping an Emerging Institution in the Global AI Governance Regime Complex." (2026-06-22). arXiv:2606.23860. Institutionalization score 1/5 (the floor of the five-point scale) — the paper itself notes the limit that this is preliminary coding from before WAICO's signing.

Policy and Institutions — Indices, Datasets, and Official Organization Comparisons

• 15.Global Index on Responsible AI. (2024). "Global Index on Responsible AI (GIRAI) 2024." Responsible-AI implementation assessment across 138 countries.
• 16.OECD. (2026). "OECD.AI Policy Observatory." A repository of 900+ policies across 70+ countries.
• 17.United Nations. (2024). "AI Advisory Body." UN High-Level Advisory Body on AI, a time-limited advisory mandate (Oct 2023–Sep 2024).
• 18.European Commission. (2024). "European AI Office." A law-enforcing body — powers of GPAI investigation, on-site inspection, and fines of 3% of turnover (or €15M).
• 19.NIST. (2025). "Center for AI Standards and Innovation (CAISI)." A standards and evaluation body under NIST.
• 20.OECD. (2026). "Global Partnership on Artificial Intelligence (GPAI)." A multilateral consultative body, 46 countries, non-binding recommendations. Korea is a 2026 co-chair.

Korean Official Materials

• 21.Korea Law Information Center. (2026). "Framework Act on the Development of AI and Establishment of a Basis for Trust (AI Basic Act)." In force January 2026; an international-cooperation clause exists (exact article number left unconfirmed in text due to public-lookup limits).
• 22.National AI Strategy Committee. (2026). "Integrated Briefing on AI Budget Programs." (2026-03-04). 2026 AI budget of ₩9.9 trillion (about 3× the prior year), 741 projects, 41 ministries.
• 23.The Electronic Times (ETNews). (2026). ""Global AI Hub" declared with 14 international bodies." (2026-05-21). 9 UN bodies + 5 multilateral development banks participating.
• 24.OECD.AI. (2026). "GPAI Council Co-Chairs Summary." Korea assumes the 2026 GPAI co-chair.

Cross-checked against primary sources as of the research date, 2026-07-18. Unconfirmed items (the complete WAICO roster, the country names of Mazu's 30, the exact figure for the intelligent-economy scale, the AI Basic Act's international-cooperation article number, the Global AI Hub's budget and beneficiary headcount, and others) are marked honestly as "not confirmed" in the text.