Immediately and for Everyone
On 12 June 2026, a US export control directive suspended access to two models of an American frontier developer. It covered “any foreign national, whether inside or outside the United States”. As the company later explained, the order “took effect immediately”. The company “had no reliable way to verify nationality in real-time”. So it “suspended access to both models for all users”. The rule named a class of people. The system could not tell who belonged to that class at the speed the rule took effect. This piece argues that two things set how wide the cut is when availability is decided outside a company. One is how fast the decision takes effect. The other is whether the system can read the decision’s criterion at that speed.
The piece was developed with AI writing assistance under human verification and editorial control. A disclosure belongs beside that sentence. The June statements this piece reads were written by the company that makes one of the text tools named in the method note. The Italian order discussed below was addressed to the company that makes the other text tool and the image tool.
What the company wrote
The statement of 12 June gives the order and its effect. The directive cited “national security authorities” and covered the company’s own foreign-national employees. The company had to “abruptly disable” both models “for all our customers to ensure compliance”. That was the order’s “net effect”, in the company’s words.
The reason for that width came eighteen days later, on a page announcing the redeployment. The same page dates the way back. Access to one model returned “for a set of US organizations, following the US government’s approval on June 26”. The controls were lifted on 30 June. Both models were back on 1 July. The cut came the day the letter arrived. The return came in steps.
The clock ran the other way
Last week’s piece on this platform left a question for this week. A monitor built on a model depends on that model staying available. Availability can be decided outside the company. June is a documented case of the second sentence. I found no report that any monitor stopped. A monitor requiring live access to either model through the provider would have lost that access during the suspension.
In February, this series named cadence asymmetry. The essay on this platform put it in two sentences: “The execution moves in permanent time. The judgment remains in discrete time.” (When Speed Becomes Infrastructure, 18 February.) The usual worry runs in one direction. The system acts before the institution can look.
June ran the other way. The deciding institution, a state, acted at once. The system could not apply the rule selectively at that speed, because it could not verify the one fact the rule turned on. The reversal is my reading. The company does not use the term.
Speed alone does not explain the width. An earlier order, in another jurisdiction, also took effect at once.
A line a system can see
On 30 March 2023, Italy’s data protection authority, the Garante, acted against an American developer over its chatbot. It imposed a temporary limitation on the processing of personal data of people established in Italian territory. The limitation took effect on receipt. Its president signed it as an urgent measure, which the full board ratified on 8 April. The act lists its grounds, from missing information for users to the absence of any age verification.
On 11 April the authority set conditions, with the first seven due by 30 April. It suspended the limitation from the day they were met. On 28 April it reported that the developer had made the service accessible again to Italian users.
Both measures took effect immediately. They differed in what they required the companies to distinguish. The Italian order covered people established in Italian territory. Its April prescriptions also addressed users connecting from Italy. These are not identical categories. Connection location can serve as a technical proxy, but it does not establish where a person lives.
The June directive turned on nationality. The company says it could not verify that criterion in real time. My reading is narrower than a general comparison of the two enforcement systems. A decision’s speed matters, but so does the information required to apply it selectively. The public records do not establish how each Italian user was classified.
Where the door was
In July, this series located the point where authority reaches the model layer (No One Governs a Road by Talking to the Asphalt, 15 July). It is the door, where a deployment becomes subject to a jurisdiction’s authority. Italy and Brazil exercised authority over data processing connected to their respective territories.
The June directive used another criterion. It covered foreign nationals inside and outside the United States. Its operational effect reached users abroad through the supplier.
For a user in Brazil who lost access, the controlling decision originated in the supplier’s jurisdiction, not the user’s.
Get André Afonso Silva’s stories in your inbox
Join Medium for free to get updates from this writer.
Supervisors already have a category for this. A Financial Stability Institute paper published by the BIS on 9 September names concentration and sovereign access risks. It reports the Australian prudential regulator’s recognition that “regulatory decisions made overseas affecting globally significant AI providers could impair critical business processes”. The paper states its authors’ views, not necessarily the BIS’s. For the people it covered, Italy’s decision was made at home. For a user outside the United States, June’s was made overseas.
One word needs a fence. In this series, sovereignty has had two senses. One is temporal (the February working paper). The other is positional (the Framework’s Sovereignty of Position). The paper’s “sovereign” is a third sense, the national one. I keep the three apart.
What the documents let you check
Monday’s newsletter, in Portuguese, proposed a documentary test for whether an interruption can be contested. Identify the authority, the scope, the reasons given to the addressee and the route to ask for review. Then find what a review could change and by when. The three cases return three different answers.
Italy. The act identifies the authority, scope and grounds. It also identifies a judicial route for challenging the measure: an action before the ordinary court within thirty days, or sixty for an appellant residing abroad. That is a documented route for challenging the administrative decision. It does not establish what procedural standing each affected user had.
The subsequent record is also dated. The authority set conditions on 11 April. The service was accessible again on 28 April.
Brazil. On 2 July 2024, Brazil’s data protection authority, the ANPD, published a preventive measure requiring the immediate suspension of certain personal-data processing for generative AI training. The measure also covered data from people who did not use the platform. It established a daily fine and five working days to demonstrate compliance.
The company requested reconsideration. On 10 July, the authority left the measure in force, extended the compliance deadline and deferred examination of the request’s merits. On 30 August, it approved a compliance plan and suspended the measure.
This is a limited analogue. The order restricted data processing, not access to a model. But the record shows a formal challenge, an interim decision and a later change in the measure, each with a date.
June. The company’s statements name “the US government” and no agency. The scope is a nationality. The letter, the company wrote, “did not provide specific details of its national security concern”. I found no route to review described in either statement. The company’s pages date what changed. They do not describe the procedure behind it.
The discount
Three cases, chosen because they are documented, not sampled. Two concern data protection and one concerns export control, under different laws and for different purposes. The documents differ as well. For Italy and Brazil I read the authorities’ own acts. For June I read the addressee’s two statements and a supervisory paper. The directive itself is not public, as far as I found. Part of the difference may sit in what was published rather than in what existed. Nothing here assesses the merits of any of the three decisions.
In June the decision was a state’s and the execution was the company’s. In an inventory of dependencies, the column “who controls” takes at least two entries for a frontier model. One is the company that holds the position. The other is the state that holds jurisdiction over the company. Thursday’s piece starts there.
One reading of June is tempting: it was only a tooling problem. That reading is incomplete.
On the company’s account, reliable real-time nationality verification could have narrowed the suspension. It would not have changed who issued the directive.
References
Anthropic. (2026, June 12). Statement on the US government directive to suspend access to Fable 5 and Mythos 5. Official Anthropic publication.
Anthropic. (2026, June 30; update July 1). Redeploying Fable 5. Official Anthropic publication.
Crisanto, J. C., Currat, A., & Yong, J. (2026, September 9). When machines attack: frontier AI cyber threats and policy responses in the financial sector (FSI Occasional Paper No 28). Bank for International Settlements.
Garante per la protezione dei dati personali. (2023, March 30). Provvedimento del 30 marzo 2023 [doc. web n. 9870832].
Garante per la protezione dei dati personali. (2023, April 11). Provvedimento dell’11 aprile 2023 [doc. web n. 9874702].
Garante per la protezione dei dati personali. (2023, April 28). Comunicato del 28 aprile 2023 [doc. web n. 9881490].
Autoridade Nacional de Proteção de Dados. (2024). Despachos Decisórios nº 20/2024/PR/ANPD (2 July), nº 24/2024/PR/ANPD (10 July) and nº 33/2024/PR/ANPD (30 August). Diário Oficial da União.
Silva, A. A. (2026, February 18). When Speed Becomes Infrastructure. Medium.
Silva, A. A. (2026, October 12). A porta de saída. Essência AI Insights, №50. LinkedIn Newsletter.
On method
Written with the support of Claude (Anthropic) and ChatGPT (OpenAI) for text and ChatGPT (OpenAI) for the image. The full cognitive signature (intent, constraints, judgment under uncertainty, veto and responsibility) is mine, and so are the errors. Facts are verified against primary sources before they enter, and where verification was not possible the text says so.
I publish this note every week because this series argues that a record produced because someone asked for it is worth less than a record produced in the ordinary course of work. Nobody asked me for this note.
Essência AI Insights is a weekly series on governance, systemic risk and institutional infrastructure for emerging technologies. Written by André A. Silva
