All walkthroughs · Paper briefs

Mapping General-Purpose AI Governance in Twenty AI Middle-Power Jurisdictions

Josephine Schwab, Nathan Naidoo, Ferruccio Barazzutti, Sheryn Lee (Arcadia Impact AI Governance Taskforce) and Caio Machado (The Future Society)
arXiv 2608.19278 (cs.CY), version 1, 18 August 2026; working paper, dataset snapshot as at 30 July 2026

Paper · Brief and rating · Dashboard entry

About 2 minutes · Download MP3

AI-generated walkthrough with synthesised voices: a host explains the paper and a discussant questions it. It covers claims, methods, key assumptions, citations and use so far, decision relevance, and what an evaluator should check. Every direct quotation was checked word for word against its source. Public sources only. Not an Unjournal evaluation or decision.

Transcript

Host
This is a walkthrough of Mapping General-Purpose AI Governance in Twenty AI Middle-Power Jurisdictions, by Josephine Schwab, Nathan Naidoo, Ferruccio Barazzutti and Sheryn Lee of the Arcadia Impact AI Governance Taskforce, and Caio Machado of The Future Society. It was posted to arXiv in August twenty twenty-six as a working paper, with the data frozen at the end of July twenty twenty-six. This episode is an AI-generated walkthrough, not an Unjournal evaluation or decision. The paper is a comparative legal mapping. General-purpose AI, which I will shorten to G-P-A-I after this, means models like large language models that are trained once on broad data and then serve many uses. The most capable of these are built in the United States and China. The paper asks what everyone else is doing about them: jurisdictions with no frontier developer of their own, which the authors call AI middle powers. Their headline, from the abstract: "We find that jurisdictions converge on form, but diverge on force."
Discussant
Before the findings, who counts as a middle power here? The international relations literature uses that term loosely.
Host
The authors define it by position in the AI political economy, not by income or military weight. A middle power is a jurisdiction that currently hosts no frontier developer, so it cannot regulate at the point where a model's capabilities are set, but which has enough market size or regional influence to shape how models are deployed. The operational test is that no developer headquartered there has trained a model at or above the compute threshold used in California's Senate Bill fifty-three, ten to the twenty-sixth operations, on published third-party estimates. France and Canada are boundary cases, since Mistral and Cohere train their own models but fall below that line. The authors' summary of the shared situation is that "each must govern models built elsewhere, by companies it cannot license or tax."
Discussant
And the twenty?
Host
Nineteen countries across nine regions, from the United Kingdom, Japan and Korea to Kenya, Peru and the Emirates, plus the European Union as the twentieth. Three subnational units are also mapped. The sample is purposive: regional spread was a constraint, and so were the researchers' languages.
Discussant
Purposive samples invite the question of what the findings generalize to.
Host
The authors say plainly that they should not be generalized. To show the sample spans a range, they build a composite score from six equally weighted clusters: economic weight, defence, resources and compute infrastructure, diplomatic agility, democracy and AI-governance infrastructure, and participation in international AI processes. Each input that comes as a rank is turned into a score out of ten. This is key equation one in the transcript. The intuition is that the top-ranked country gets ten and the bottom-ranked gets zero, with equal steps in between. Let N be the number of countries in the source ranking, and let r be a country's rank. Take N minus r. Divide that by N minus one. Multiply by ten. So in a population ranking of two hundred and seventeen countries, rank one scores ten, rank two hundred and seventeen scores zero, and rank one hundred and nine scores five. The composite runs from about three point four for Kenya to about eight point two for France.
Discussant
So rank one hundred and nine is exactly halfway: one hundred and eight steps out of two hundred and sixteen. That treats every rank step as equal, which loses magnitude information, but for a validation device that seems fine. Does the composite do any work beyond showing spread?
Host
Some. Later the authors use it to point out that national weight predicts little about governance. France and the United Kingdom top the composite, yet France has no domestic hard-law AI instrument imposing obligations on these models. Brazil sits twelfth and has the second-largest body of binding law in the sample. Kenya, Nigeria and Brazil, all in the lower half, have drafted some of the most demanding bills, which the authors flag as a hypothesis about ambition and industry presence being inversely related.
Discussant
Now the core design. What exactly is coded?
Host
The unit of analysis is the individual provision. Across statutes, regulations, guidance, strategies and bills, they extract provisions in four governance areas chosen to trace an accountability chain from before deployment to after harm. One, systemic risk assessment: duties to assess risk above the level of a single deployment. Two, evaluation and verification: evaluation measures what a model does, such as capability testing or red-teaming, meaning adversarial testing to find dangerous behaviour; verification checks claims, such as audits and certification. Three, prohibitions, together with the monitoring and detection needed to notice a breach. Four, serious incident reporting, using the serious-incident definition of the Organisation for Economic Co-operation and Development, OECD, which covers death or serious harm to health, serious disruption of critical infrastructure, and serious rights violations or property harm.
Discussant
How does an instrument get in? Lots of AI law is technology-neutral.
Host
Through an eight-step decision flow with an either-or gate. An instrument passes if it carries a G-P-A-I definition, which includes defined subclasses like generative AI or foundation model. Or, failing that, if it has operative provisions in one of the four areas that respond to risks specific to general-purpose models. Two risk classes count automatically as such brightlines: loss of control, meaning systems that cannot be reliably corrected, interrupted or shut down, or that resist oversight; and C-B-R-N uplift, meaning meaningful help toward chemical, biological, radiological or nuclear weapons. Defence, national security, voluntary industry commitments and case law are excluded.
Discussant
And the novelty claim is confirmed absence.
Host
Yes. When the search for a jurisdiction and area was completed and an outside expert reviewed the per-jurisdiction report without adding an instrument, the absence of provisions is recorded as data. The authors are careful about what this means: "A negative finding is therefore only as good as the search behind it". The final corpus is one hundred and one instruments, one hundred and forty-four governance bodies, fifty-two definitions, and five hundred and eighty-seven provision records, of which three hundred and eighty-two are positive provisions and two hundred and five are confirmed absences.
Discussant
Five researchers across twenty legal systems is a lot of reading. How was extraction done?
Host
Coding ran from early June to the end of July twenty twenty-six. The large language model statement is unusually specific. Claude Opus was used to locate candidate provisions and propose coded values, and a researcher checked each against the primary text. Of the five hundred and eighty-seven provision records, one hundred and sixteen were extracted manually, three hundred and seventy-three with model assistance, and ninety-eight with model assistance from translated text. For four languages where the team lacked fluency, translation was model-assisted with round-trip back-translation. The codebook went through fourteen major versions. Expert interviews were held for Canada, Chile and Japan; the other seventeen got asynchronous expert review.
Discussant
Give me the results that carry the argument.
Host
First, legal force. Of the three hundred and eighty-two positive provisions, eighty-five sit in binding law, about twenty-two percent. Ten of the twenty jurisdictions have no binding, in-force provision at all. The European Union holds twenty-six of the binding provisions and Brazil eighteen, so together they hold more than half. The authors describe two groups: "jurisdictions that legislate, and jurisdictions that publish." The second group is larger.
Discussant
Does binding mean binding on the developer?
Host
Rarely, and that is the second result. The United Kingdom has forty-four positive provisions, more than any jurisdiction outside the European Union, and twelve are binding. All twelve sit in two sectoral statutes, the Online Safety Act and the Crime and Policing Act, and none imposes a duty on the provider of a general-purpose model. In evaluation and verification, one hundred and thirty-six positive provisions across sixteen jurisdictions, only six are binding, and only two of those reach the entity that built the model: the systemic-risk provisions of the European Union AI Act and Article thirty of Korea's twenty twenty-five Framework Act. Outside the European Union, no jurisdiction imposes a binding evaluation duty on a developer, and outside the European Union no provider is under a binding, in-force duty to report serious incidents.
Discussant
Korea is the only one that legislated its own binding chain?
Host
Yes: a definition, a compute threshold, then assessment and evaluation duties. The comparison with the European Union is instructive. The European Union presumes systemic risk above ten to the twenty-fifth floating-point operations of training compute, rebuttable by the provider. Korea's Enforcement Decree uses ten to the twenty-sixth, ten times higher, and makes it one of three cumulative conditions alongside a state-of-the-art test and a severity test. A model trained between the two lines is a systemic-risk model in Europe and outside Korea's class altogether.
Discussant
What about definitions more broadly?
Host
Third result. Seventeen jurisdictions hold at least one definition, but fifty-eight percent of instruments impose provisions without defining what they govern, and fifteen of the twenty binding instruments have no definition. Definitions cluster in guidance and strategy. In their words: "Jurisdictions are most precise about what GPAI is exactly where precision carries the least consequence." Most definitions describe outputs, such as generative AI, rather than generality or capability, and only the European Union separates model from system in binding law.
Discussant
Why does model versus system matter?
Host
Because it decides who bears the duty. In the OECD's terms, a model is the trained parameters, and a system adds the components needed to run it. Obligations on a model fall on its developer. Obligations on a system can fall on whoever deploys or integrates it. The paper finds duties attaching mostly at the application layer: deployers, government purchasers, platforms, supervised firms. Prohibitions address use far more than capability: ninety use prohibitions, twenty-four capability prohibitions and thirteen outcome prohibitions. The practical point is that a capability can be tested before release, while a use prohibition is enforced only after misconduct or harm.
Discussant
Detection of breaches?
Host
Thin. Of the one hundred and twenty-seven prohibition records, fifty-six percent say nothing about who would notice a breach. Absolute bans are the least likely to name detection, twenty-seven of thirty-nine, because they sit in criminal statutes and ethics charters that rely on complaints and policing. They observe that in jurisdictions with only two or three prohibitions, those tend to be absolute: "Where a jurisdiction says little about prohibition, it says it categorically." Prohibitions concentrate in manipulation, children's safety and human rights, where existing law already offered a vehicle, and are sparse for C-B-R-N and loss of control.
Discussant
And incident reporting, the last link.
Host
They treat a reporting duty as a chain of four elements: an addressee, a trigger, a time period, and a consequence for failing to report. Ten jurisdictions have no reporting duty of their own, counting France and Germany, which rely on European Union law. Of twenty-seven operative reporting provisions, only three state both a period and a consequence: two in the European Union and one Brazilian judicial resolution. Their point on open formulations like without undue delay is that "a duty with no stated period has no point at which it can be breached". Outside the European Union, reporting often runs through borrowed channels, such as financial supervision in Korea, which inherit the parent regime's trigger, so an AI failure with no data breach, outage or suspicious transaction may trigger nothing.
Discussant
And the institutions?
Host
Of the one hundred and forty-four bodies, close to four in five existed before the AI mandate they now exercise; only nine are dedicated AI regulators. No named safety or evaluation institute in the corpus holds enforcement powers. The European AI Office is the only evaluation body that can act on what it finds. Their phrasing: "These institutes were built to measure, but constituted without the power to act on what they find." From that they draw their main policy suggestion: the gap between evaluation capacity and consequence can be closed domestically, before any international agreement on which capabilities to prohibit.
Discussant
Doesn't that partly concede the skeptics' view? They cite Anton Leicht, who argues middle powers cannot realistically constrain frontier development and should govern deployment instead.
Host
They engage it directly. On their reading, these jurisdictions have already made the pivot Leicht recommends, since most of what they have written attaches to deployment, procurement and supervised firms. They write: "This should not be read as a failed attempt at GPAI model governance." They also note sectoral routes can bite. Financial-sector guidance is soft law in form, but a regulator writing to its own licensees has leverage, and they find fifteen of the fifty-five enforceable soft-law provisions in four financial instruments. And in Japan, guidance with no legal force is understood as binding, with compliance running through procurement.
Discussant
Now uptake. Has anyone used this?
Host
Very little, as expected for a paper seven weeks old. OpenAlex records the arXiv paper and two Zenodo dataset records, with zero citing works as of early October twenty twenty-six. The Altmetric score is one. A web search found the arXiv listing, an automated summary site, and an earlier cohort announcement from Arcadia Impact. We found no press coverage, policy citations or critiques. A miss is uncertainty: practitioners may use a dataset like this without citing it.
Discussant
So who could act on it?
Host
Three audiences. Ministries and regulators in middle-power jurisdictions benchmarking their own regimes, especially the four with bills in committee: Brazil, Chile, Kenya and Nigeria. International coordination bodies and red-lines campaigns, since the paper finds little national practice to build capability-based prohibitions on. And AI governance funders deciding whether middle-power advocacy should target evaluation mandates, incident reporting, or definitions. For an Unjournal evaluation, the decision-relevant claims are descriptive, so the evaluation question is whether the coding is reliable enough to support them.
Discussant
Then the critiques. Start with the framework. It is built from the European Union AI Act, the OECD and the G7 Hiroshima Process. Isn't the European Union coming out on top partly mechanical?
Host
The authors concede it. They write: "Findings that the EU covers the framework most completely therefore reflect the framework's provenance, as much as the instrument's content." They mitigate by setting the systemic-risk threshold below the European Union's construction and by admitting subclass definitions. But an evaluator should ask whether a framework built from European categories under-reads regimes organized differently, such as Japan's procurement approach or the Chilean consumer agency's informal effect, which the authors only recover from interviews.
Discussant
Second, counting. The headline percentages are shares of provision records, and the default is one record per section of an instrument.
Host
Right, and they warn that provision counts show where content is dense, not where obligations are strong. A long guidance document with many sections adds many non-binding records; a short statute adds few. So the twenty-two percent binding figure depends on drafting style. An evaluator would want the headline results restated at the instrument level, or weighted, and a check of how sensitive the shares are to the section rule. Relatedly, the inclusion rule admits instruments anchored only by a generative AI definition. They did rerun the main findings excluding the twelve such instruments, carrying twelve percent of positive provisions, and report that the pattern holds and sharpens. That is a good robustness check, and others of the same kind are possible.
Discussant
Third, reliability. Model-assisted extraction, single researchers per jurisdiction, and a codebook revised fourteen times during coding. Was there any inter-coder agreement measure?
Host
None is reported. Records flagged low confidence needed two-researcher validation, and the paper says no records were flagged at publication, which an evaluator might find surprising rather than reassuring, given translated legal texts in four languages. A codebook that changed during extraction also raises the question of whether early records were recoded to the final version. The revision history is said to be documented.
Discussant
Can an outsider check any of this?
Host
Partly. The Zenodo deposit, under a Creative Commons attribution licence, holds the instrument list, the governance bodies, the fifty-two definitions and the out-of-scope examples. But the provision-level records, the five hundred and eighty-seven rows that carry nearly every percentage in the paper, are not released; they are available to editors, reviewers and researchers on request. An evaluator would need that access. A natural audit would recode a random sample of provisions from the primary texts.
Discussant
I also noticed some internal counts that do not line up.
Host
Mostly small ones that turn on conventions. The abstract says sixteen jurisdictions engage at least three of the four areas; the conclusion says fifteen, excluding subnational units. The count of eleven jurisdictions with nominal coverage of all four areas includes France and Germany through European Union law, while elsewhere France and Germany are shown on their domestic instruments alone. These probably reconcile, but the switching conventions are easy to misread. The authors also say they used a model to verify reported figures against the dataset, another reason to recheck the arithmetic.
Discussant
Last, is form versus force really the right summary, given their own evidence that form misleads?
Host
That is the tension an evaluator should press. The paper codes legal form, then shows in Japan, Chile and the Emirates that form is a poor proxy for practice. With interviews in only three jurisdictions, the conclusion that middle-power governance is weak in force may overstate weakness where soft law works through supervision and procurement. The authors say as much and list fieldwork on the gap between law and practice as the first follow-up. Given that, and a short shelf life with four bills pending, an evaluation's value is probably in a narrow coding audit rather than a general expert review.
Discussant
So: a careful, transparent descriptive map, with an explicit method for recording absence, whose headline shares rest on unreleased provision coding, a European-derived framework and a section-level counting rule.
Host
Fair. The authors label their policy claims as judgments, and say the domestic gaps they identify "can be closed by jurisdictions acting individually, before any agreement exists". Whether the map is accurate enough to guide that is what an evaluation would test.

Key equations

score = ((N − r) / (N − 1)) × 10

where N is the number of countries in the source ranking (for example N = 217 for the World Bank population ranking) and r is the jurisdiction's rank in that ranking (1 = top). Rank 1 scores 10; rank N scores 0. Each cluster score is a mean of such sub-scores (some clusters use an adaptive mean that drops missing inputs), and the composite is the unweighted mean of the six cluster scores. The Stanford AI Index investment rank uses a variant, ((16 − r) / 15) × 10 with a floor of 0.5.

Sources