Politics, Technology, and Trade: Impacts on Negotiations over the Global Governance of AI (a data-driven exercise)
By Javier Surasky
For
instance, at the United Nations—still the core of multilateral life despite its
current pronounced weakness—some delegations co-sponsor resolutions, push for
ambitious normative language, and participate in technical negotiations.
Others, by contrast, are virtually absent from these dialogues.
Understanding
why this difference exists helps clarify the extent to which global AI
governance will be genuinely inclusive or remain concentrated in the hands of a
few. Furthermore, this difference reveals how a feedback loop in the digital
divide among countries has already begun to take shape.
To achieve
this, we built an original database covering all 193 UN Member States, then
tested several machine-learning models to detect consistent patterns in
diplomatic participation in debates on the future of AI.
Our main
challenge was the lack of an indicator for AI governance participation. We
therefore used UN Member States’ actions in General Assembly debates on the two
relevant resolutions: Safe, Secure and Trustworthy Artificial Intelligence
Systems for Sustainable Development (A/RES/78/265)
and Artificial Intelligence in the Military Domain and its Implications for
International Peace and Security (A/RES/79/239).
From these, we built an engagement indicator by rating each country based on
its role in these resolutions (presenting the draft = 1; co-sponsoring the
draft = 0.5; adding co-sponsorship at adoption = 0.25; neither = 0). The mix of
these roles yields values from 0 to 1 in 0.25 increments, reflecting levels of
engagement across both resolutions.
First point
to highlight: no State co-authored both resolutions.
To work
with this conceptualization, and given the limited sample (193 cases), we
reduced dispersion by applying a binary scheme: those who barely participate (0
or 0.25) and those who show more significant involvement (0.5, 0.75, or 1).
This distinction reflects a familiar political logic in diplomacy: “joining for
formality” is not the same as signing onto a resolution that stakes out a
position in an emerging field.
With this
dependent variable defined, we then constructed a set of predictors to
represent dimensions which, according to our expert intuition, could influence
diplomatic engagement: technological capabilities, institutional quality,
international insertion, trade profiles, and membership in global coalitions.
To capture
technological capabilities, we used indicators such as the level of
digitalization, the e-government index, and the AI Readiness score.
For the
institutional dimension, we included the democracy index, the rule-of-law
strength, and moving averages of civil-liberty indices.
For
international insertion, we incorporated each country’s level of trade
dependence on the three major actors shaping AI governance (the United States,
China, and the European Union), along with variables reflecting membership in
groups such as the G77, BRICS, the G20, the OECD, or the European Union. The
rationale is that AI governance is shaped by both domestic capacity and
external dynamics that condition political incentives.
After a
technical cleaning of the generated database, we assigned label roles to mark
the expected outputs for the models. In an initial run, the country variable
was marked as an identifier, and the participation variable related to the
above-mentioned resolutions was the label. All other attributes were treated as
regular features. The original co-sponsorship value was excluded to avoid “data
leakage”, a situation in which the model reaches an artificially easy solution
by relying on a single variable closely aligned with the target classification,
thereby failing to use other variables.
We first
used a decision tree for its interpretability, setting the Gini criterion and
limiting depth for clarity. The resulting tree yielded 77.6% accuracy, and the
variables with the greatest weight—those consistently appearing as main
branches—were digitalization, democracy, rule of law, and AI-readiness metrics.
In other words, the tree showed that countries participating more actively in
AI governance tend to have strong digital infrastructure and high-quality
institutional systems. Multilateral diplomacy is filtered through domestic
political, governance, and technical capacities.
To contrast
this interpretive structure, we ran a Random Forest model, which combines the
statistical robustness of many trees with the ability to avoid overfitting.
Accuracy remained high (74.1%), but the most relevant contribution was the
improved observation of variable hierarchies. The five most influential
predictors were digitalization, the e-government index, democracy, rule of law,
and the Global AI Readiness Index. These reflect three distinct capability
dimensions—digital, institutional, and technological—that converge toward a
common tendency.
Trade
dependencies and international bloc memberships played only secondary roles.
This is important because it challenges simple, geopolitically driven views.
Membership in G20, BRICS, or G7 alone does not predict which countries are
active in AI governance.
As a more
extreme contrast, we ran a decision tree using only levels of trade dependence
on the three actors leading debates on AI governance. We found a strong
relationship between trade with the EU and the likelihood of participation, but
the model’s predictive capacity fell to 62%. This decrease is itself highly
telling: comparing the three models shows that trade with one of the three
competing governance poles has limited influence on the likelihood of active
involvement in AI debates at the multilateral level, and global coalitions add
little to predicting how engaged a State will be in these discussions.
What most
differentiates participants from nonparticipants is their domestic capacity to
understand, regulate, and leverage advanced technologies. If a State lacks
digital infrastructure, strong institutions, or a minimal base of technical
preparedness, its diplomats are less likely to play an active role in UN
negotiations on AI.
This
finding has deep political implications. It tells us that the future of global
AI governance is being shaped in a space where countries with strong
technological and institutional foundations dominate. The digital divide is not
only technological-economic; it is regulatory, governance-related, and
democratic. If the international community aspires for AI to be governed
inclusively and for emerging standards to genuinely represent the diversity of
global positions—which is far from being the case—it is as urgent as it is
indispensable to strengthen these multiple capacities in countries with low
involvement.
This
analysis also shows that machine-learning models can help study international
politics. They do not replace theory but reveal patterns that might go
unnoticed. Here, trees and Random Forests empirically show that diplomatic
participation in technological issues depends on domestic capacities to
integrate those technologies. Participation relates more to governance,
democratic strength, and political structures than to technical matters. Thus,
AI not only reshapes the economy, but it also reorganizes the geography of
normative power, reinforcing earlier arguments about the wider implications of
these governance patterns. The results of our work are not a final verdict—nor
were they intended to be—but rather a suggestion of alternative avenues to
explore and generate evidence as debates on AI governance progress within the
United Nations. They show how critical it is to pay attention to who is in the
room—and who is not, tying back to the original concerns about inclusivity and
participation.
These
models clearly show that inclusion will not occur automatically as AI advances.
It will require strategic investments in digital capabilities, credible
institutions, and technical literacy within States.
Today, we
see the most marginalized, least powerful, and institutionally weakest excluded
from debates, even when they form coalitions. This exclusion widens the digital
divide, shaped by governance logics and marked by silenced voices.
And the
silences of today will echo in new forms of domination, vulnerability, and lack
of protection for entire populations in the future, unless we act in the
present.
Note:
Metadata Dictionary: https://drive.google.com/file/d/18puchzw3RIOf_z1robzuIEQTq5aQuEQi/view?usp=sharing