Unpredictable decisions, Unpredictable results
Alfredo González Briseño
7/28/20269 min read
It was all over the news. If you use Claude, most likely you were affected from the "kill switch" created by the government regulation that took everyone by surprise. All regulatory decisions have an explanation. Some are rational, others are not. Some are transparent, while others are very opaque. "How" these rules are made matters a lot—often more than final decisions. The export control directive on Anthropic is a clear example of important regulatory decisions around the world not following good practices, resulting in less trustworthy and unpredictable business environments. Let's analyze and talk about what happened.
Summary of what happened
If you heard the news or know the details, bear with me for a second while I share the story with others who haven't.
On June 12, a Friday, the Bureau of Industry Security of the US Department of Commerce issued an export control directive requiring Anthropic to get a license for any export, reexport or transfer of its latest models Fable 5 and Mythos 5. Anthropic reported receiving the directive at 5:21pm ET.
The immediate consequence was that Anthropic suspended access to foreign nationals—living abroad and in the US, including their own foreign employees— to the two models targeted by the new regulation.
The alleged reason for this regulatory measure was national security concerns—although it remains unclear as the directive notified as an Is-Informed Letter (IIL) was not published.
Two weeks after on June 26, also a Friday, another letter was sent providing licensing exemption to certain foreign nationals—including those employed by Anthropic—, only applicable to the use of the model Mythos 5.
Forbes, Mayer Brown and others reported on this—although the second letter (IIL) was also not publicly disclosed by the government.
On June 30, 18 days after the initial decision, the export control on the use of model Fable 5 was lifted too, as announced by Anthropic, media and social media—there was no public letter or official government communication regarding this decision.
Anthropic announced also its collaboration with the US Government and how its models became available after the June 26 and 30 decisions.
Regulatory policy lenses to look at this decision
Decisions like this happen all the time around the world. In some countries or administrations, they are the normal way of creating government rules. A "way" of making decisions that to many in the public—including those directly affected—seem opaque, unpredictable, disproportionate, targeted, unreasonable and unfair.
I have heard this type of complains from business associations and firms in many continents. Complains about "weekend rules" that are enacted late on Friday and take everyone by surprise on Monday.
When decisions are made under "national security concerns"—like in the case of Anthropic's models— it is more difficult to analyze them from a policy perspective. Many of the underlying details and rationale are not publicly available for others to review. Most of what is available are media reports filling out the information gaps.
However, this does not mean that this type of regulatory decisions with significant impact cannot be analyzed using a framework or set of "policy lenses."
When those lenses are built around the principles of good regulatory practices (GPR) and due process in decision-making, it is possible to clear out some of the noise from what is said in the news, and start seeing decisions as they are or should be.
To contextualize more, GRPs have been advocated for years by many countries—including the US—in international trade fora, organizations and agreements.
I will be using a simple framework grounded on GRPs, to analyze the export control directive on Anthropic's models. The graphic below shows the 5 pillars of my framework as well as the related questions to test them. I call this the "Better Regulation Test" or BRT Framework.
The Better Regulation Test has nothing to do with the actual text of the regulation. Rather it focuses on the "how" or "process" leading to final regulatory decisions.
The "how to" regulate is often more important than the rule itself, as it is the foundation for creating trust in government and future compliance. You can read and learn on the concept and empirical evidence of "procedural fairness" backing up this idea.
Before getting started, it is important to notice that governments worldwide face an enormous challenge from regulating rapidly evolving and disruptive technologies and innovations like AI. This makes more relevant the need to adopt good regulatory practices consistently, including those that promote innovation, while protecting the public good.
Source: Self, based on international good regulatory practices.
What the BRT Framework tells us
As there is no official information regarding the decision and rationale to impose licensing requirements on Anthropic, through an export control directive, I will be answering the Better Regulation Test questions based on online public information. I invite you to run the same analysis with the information and insights you have.
1. Problem and Evidence. A good regulation starts with a clearly identified problem it is trying to solve, and evidence about it. Otherwise, decisions tend to be based on perceptions, ideology, wrong ideas and even good intentions.
In the case of Anthropic the alleged concern was national security. Because the directive, notified as an Is-Informed Letter (IIL) was not published, it is difficult to assess the actual problem and concern the government saw and communicated in this case.
From Anthropic's Announcement on June 12, we know the following:
"The letter did not provide specific details of its national security concern. Our understanding is that the government believes it has become aware of a method of bypassing, or “jailbreaking” Fable 5."
"To date, the government has only given us verbal evidence of a potential narrow, non-universal jailbreak, which essentially consists of asking the model to read a specific codebase and fix any software flaws. Our understanding is that one potential jailbreak was shared with the government. We have reviewed a report that we believe is the basis of the government's directive and validated that the level of capability displayed there is widely available from other models (including OpenAI’s GPT-5.5), and is used every day by the defenders who keep systems safe."
An article by Politico on the 24 hours leading to the decision reports that:
"On Thursday, two days after the model’s public release, Amazon CEO Andy Jassy raised concerns to the White House about the ability to bypass the model's guardrails, according to the two administration officials and the senior White House official. (Amazon, which is an investor in Anthropic, was responding to an administration request for feedback, said a person familiar with Amazon’s discussions.)"
Politico continues describing the discussions between the federal government and Anthropic, and mentions that:
“Export controls were a last resort after begging them for hours to work with us,” the senior White House official said. “This was not something we wanted to do, but our hands were tied.”
After publication, one of the people close to Anthropic disputed that the company was given a choice to voluntarily work with the administration.
“The White House gave 90 minutes to take the models down, with no details on the actual threat," the person said. "There was never any begging — or asking — for them to work with us, just a declared 90 minute deadline."
2. Policy Alternative. Once the problem at stake is clearly identified, a good regulatory practice is to assess whether the proposed decision or alternatives—which may include other regulations, non-regulatory measures or simply doing nothing—is the best approach to tackle the problem at stake.
It is clear that the option of doing nothing was not followed. Again, it is difficult to determine whether alternative solutions were considered, based on the lack of public government information. Yet, the short period of time, the description of Politico of how those 24 hours from Thursday to Friday evolved, and the government lifting up the licensing requirement 2 weeks after, are simply not a good sign.
3. Predictable decisions. In most corners of the world, firms, investors, entrepreneurs and stakeholders affected by regulation value predictable decisions. Predictability in rulemaking implies a few things: (i) planning and announcing regulatory changes; (ii) establishing technical dialogue to understand the problems at stake and alternative solutions; (iii) intra-governmental and public consultations not to reach consensus, but to respectfully listen to what others have to say and give them a legitimate channel to influence final government decisions; (iv) be transparent about final decisions and how feedback received was considered or not.
In the case of Anthropic, despite having a national security concern, the decision was as unpredictable as it could be. While there was engagement with the affected party—as reported by Politico—the timeframe, evidence, decision announced on a late Friday afternoon, the fact that the directive letters (IILs) have not been publicly disclosed by the government, and the reversal of the "switch off" a few weeks after are all indicative of a not well planned and unpredictable decision.
4. Leading institutions. Regulatory decisions need to consider not only the "how" to regulate but also the "who" will be responsible for implementing final decisions. Who will monitor compliance? Who will enforce it? These are not questions about political leaders. These are questions about institutions that should remain and survive despite changes in administration.
In the case of Anthropic, current issues of institutional capacity were fully uncovered. Not because of the Bureau of Industry Security itself. Rather, because AI is moving way faster than any government and regulator can or will ever be. But this pacing problem—although very tangible— is not always the main problem. Innovation is and will always be ahead of regulation.
The real problems include: bad regulatory practices; regulators being asked to do more with less resources; and the absence of regulatory tools and solutions oriented to address issues from disruptive technologies and innovation.
The case of Anthropic in the US raises an important questions for other countries worldwide: Can you succeed and get better results from AI without a government institution or set of institutions responsible for coordinating AI governance and regulations across government? Food for thought.
5. Future review. One of the possible regulatory solutions oriented to challenges from innovations that change rapidly is the frequent review of regulatory decisions, either through sunset clauses or similar practices. Beyond their name, the important function of these policy tools is to set a defined timeframe to review the regulatory decision 6, 12, 18 or 24 months after implementation. The goal: to assess if its original purpose remains valid or whether the rule needs to be updated or eliminated.
In the case of Anthropic, collaboration and dialogue between the tech firm and the government appeared to be more a reaction than a concrete measure for future review of the decision and its effectiveness. Even when the export control was lifted, there was not indication of a framework to be followed in future decisions.
A summary of the Better Regulation Test Assessment
Below is a graphic showing a summary of the analysis above. As mentioned before, you may have more elements and information—either you are in the government, industry, media, academia or somewhere else. I invite you to use the framework and do your own analysis.
Source: Self, based on international good regulatory practices.
Run your own analysis and score your Better Regulation Test
I also created an interactive test to score each pillar. Feel free to use the QR code below or this link to score your BRT using a scale from 1 (lowest) to 7 (highest).
The problems not many people are talking about
It is unclear whether the export control directive on Anthropic will bring legal battles to the federal government, besides the lawsuit by the Canadian legal technology startup, which ended once the access to Anthropic's models was restored—as reported by Reuters.
However, there are two topics that have not received a lot of attention, so far:
(1) The cost created by the decisions happening within a 18-day period.
(2) Whether the decisions made in those 18 days followed a due process.
Topic (1) requires an economic analysis. Unpredictable regulatory decisions often create unnecessary costs, some of them significant to all parties involved—private sector, citizens and the government itself.
For the private sector, this decision obviously imposed direct economic costs on Anthropic and the users of its models Fable 5 and Mythos 5—overall estimated costs yet to be accounted for.
From a government perspective, the cost was more political than financial. Unpredictable decisions undermine trust in government. When these decisions become the norm, they are negatively reflected in the business environment and investment climate.
Topic (2) requires a more detailed legal analysis. At least in the US, regulatory decisions need to follow a due process mandated by the Administrative Procedures Act (APA). For those interested, this is the legal requirement from the APA for rulemaking under regular circumstances.
Because the export control directive requiring Anthropic to apply for a license fell under circumstance of military or foreign affairs government functions, different procedures are followed.
These different procedures for the case of export controls are governed by the Export Control Reform Act (ECRA) and the Export Administration Regulations (EAR).
At first glance 50 U.S.C. 4813(a)(15), 4815, 4821(a), 4819(c) and 4843(c); EAR 744.11, 744.22 and Supplement No.5 to Part 744; as well as the references they make—including those to sections of the APA—are key to understand the due process in this case. To get clarity, a deeper legal analysis is needed.
What's next?
From my end, the intention was to show you how regulatory decisions can be analyzed with a simple framework or "policy lenses"—grounded in good practices available for governments when regulating AI or any other field. This is an agenda that has been there for decades.
You can use this analytical framework whether you are in government, industry, a business association, media, think tank, academia or other roles. As I mentioned, you may have more information and insights, which will make your analysis richer.
I will be posting more on the Better Regulation Test and other tools you can use to clear out the noise, see things as they are or should be, and better understand regulatory decisions—including those that are unpredictable.
