The Real AI Regulation Divide is not the US vs. the EU—It is Transparent vs. Opaque

The Real AI Regulation Divide is not the US vs. the EU—It is Transparent vs. Opaque

Author – Mohaned Bahr

Currently, the dominant concern in the AI regulatory discourse is how to decipher the code for enacting regulations that enable innovation while protecting against AI’s perils. In theory, this regulatory equation seems solvable. In practice, the accelerating AI race and the high capabilities that the new frontier models demonstrate suggest otherwise (Smuha, 2019). In addition, due to the highly transformative power of AI, governments have begun to view this technology as a national security asset, prompting them to adopt regulatory tools and design policies that serve their national interests rather than agreed shared standards. Thus, AI regulatory fragmentation is the ostensible outcome that spans binding vs. soft regulations and vertical vs. horizontal ones (Systemic Regulation of Artificial Intelligence AI Safety Legal Paper Series 12-24  56 Ariz. St. L. J. 545, n.d.). This fragmentation renders the underlying regulatory equation hard to solve, not easier.

This scenario is mostly visible across the Atlantic. For instance, the European Union (EU) has projected its AI Act as a template for the world, which regulators can use as a blueprint for their potential AI regulations (Almada & Radu, 2024). Meanwhile, the United States (US) framed the European approach as a drag on innovation and a sign that Europe is being phased out of the AI race (Bradford, 2025). This translates well in line with the current US AI policy. Since the current administration took office, the US AI regulatory philosophy has shifted towards “innovation first, regulate later,” and, in some contexts, it has embraced a “hands-off” regulatory approach. Such an approach aligns with the current US administration’s “America first” policy and gives AI developers the green light, without adequate guardrails to address the question of “But China” (Khan et al., 2026).

Simultaneously, other regulators elsewhere are observing this transatlantic contest and asking which model is the winner to copy and which is the loser to avoid. Arguably, that question got harder to answer in light of the recent US Executive Order 14409 (Orders, 2026), which presents a meaningful divergence from the administration’s own “hands-off” regulation, and ironically, lends credibility to the EU’s more precautionary regulatory model, which the very approach the administration preferred to dismiss as innovation-stifling.

In this regard, it is not being claimed that either the US or the EU model is simply right or wrong. The two models, similar to any other AI regulatory models in the world, are still experiments, subject to review and modifications to cope with the current rapid and unpredictable AI advances. Therefore, it is argued that the debate should not be centred on whether more or less regulation is better. In this opinion, it’s contended that the right axis to argue is not whether a jurisdiction regulates, but whether its regulation is sufficiently transparent to spur AI developers to innovate in a regulatory-certainty environment. Factors like public benchmarks, transparent review mechanisms and avenues for challenging contested decisions provide the AI industry with a well-designed road map of what is permitted and why. This map is what renders a regulatory regime capable of coping with and compatible with the agile approach of AI innovation, regardless of how strict or lenient it is. This blog post demonstrates that the recently issued Executive Order (EO) 14409 proves that point in reverse: it borrows the EU’s precautionary instinct, while discarding the transparency that made that instinct workable in the first place.

To test that claim, Section 3 of EO 14409, “Promoting Artificial Intelligence Innovation and Security”, is closely examined through a regulatory analytical lens, drawing on scholarly and press criticism of it in light of its closest EU counterpart, the AI Act.

EO 14409 vs. the EU AI Act: borrowing the impulse, losing the transparency

“More innovation, less regulation” has been the salient feature of current US AI policy, as is often said (Hine & Floridi, 2025). Such logic viewed the EU AI Act as a stumbling block for American AI technology and criticised it as a barrier to AI innovation. The US has called for its suspension, or temporary pause. (Rankin, 2025) Ironically, the Act itself has become a reference point within the US sub-state context. Several states in the US adapted and partially contextualised specific parts of the Act’s architecture to inform their own state AI regulations, in a rebellious response to federal silence on AI and to address AI-related harms. That sub-state AI regulatory momentum has nudged the current administration to direct its executive branch to challenge any state laws that conflict with the federal AI regulatory approach (Executive Order 14365—Ensuring a National Policy Framework for Artificial Intelligence The Executive Record, 2026), clearing the path for AI innovation without friction (Hine & Floridi, 2025).

However, against this backdrop, the EO 14409 has been interpreted as a shift in the current US-AI policy tone, since it intended to subject cover (frontier) models with specious cybersecurity and privacy implications to the government’s oversight, which is classified and lacks transparent procedures. This blog post is not criticising the EO as a whole; however, the criticism is centred on its Section 3 in specific. If the EU’s approach is precautionary regulation, Section 3 is precautionary regulation too, but dressed in a different rhetoric. Section 3 requires developers of highly capable models to subject their developed models, before their release, to government vetting on cybersecurity grounds. Such a requirement has been framed as voluntary, but critics argue that it serves as a de facto mandate that sits awkwardly next to the administration’s own “less regulation” doctrine (PhD, 2026; Tillipman, 2026a).

What does Section 3 of the EO 14409 require?

Section 3(a) directs the creation and development of a classified benchmarking process to evaluate a model’s advanced cyber capabilities and determine whether it counts as a covered model. Section 3 (b) (i)-(iii) go further, and mandate that developers must submit qualifying models for a 30-day government review before release, and require developers to engage the government in the selection process of a trusted partner who gets early access to the pilot version of the designated model (Controlling Advanced Artificial Intelligence, n.d.; Orders, 2026).

EO 14409 focuses on oversight design, but lacks the transparency feature

Critiques of this EO’s design framework are nearly unanimous and are convincing for several reasons. Firstly, ambiguity by design: the first criticism directed at Section 3(a) was that a classified reviewing and benchmarking process provides no visibility for AI developers to learn about the thresholds their models will be reviewed against. No developer is expected to invest capital, resources and time in developing a model that will ultimately be vetted in a process in which its criteria are not explicitly and unequivocally communicated. Secondly, no due process: the EO language reads as if the vetting decisions are not subject to any independent review or appeal procedure to review the government’s decision against the qualified models. Such an approach runs against a basic expectation of regulatory certainty that investors and developers alike rely on. (Tillipman, 2026b). Thirdly, discretionary control over market access: the government’s role in selecting which trusted partner gains early access to a frontier model has been criticised as a form of government censorship resembling China’s approach to managing its domestic AI sector (PhD, 2026).

In this regard, the AI Act, which was in the firing line of the current US administration, remains considerably more transparent, although it is subject to criticism for its flaws and contains open-textured language (Ebers, 2024). However, the EU AI Act lands as far more legible than the EO. The Act is explicitly transparent about the AI practices it prohibits, providing developers with a clear map of what to avoid. In addition, it sets out clear obligations for high-risk AI systems and specifies criteria for classifying general-purpose AI models as posing systemic risk. Besides, the Act is not silent on the appeal mechanism, under which developers can lodge appeals against contested decisions, separate from the bodies that issued them (Jozak, 2025).

This is the irony that is worth highlighting; the administration framed the EU’s model as a stranglehold on innovation and then issued its own precautionary framework with fewer safeguards than the one it was criticising. If either regime deserves “the innovation-stifling” label, the evidence points here toward the one without the appeal rights—not the one with them.

However, the above does not imply that the Act is free of criticism. The Act has undergone a review to simplify regulations (EU Digital Omnibus on AI Enters Into Force, n.d.). Therefore, AI regulation will need to stay agile rather than fossilise around today’s assumptions.

Conclusion: Not asking “how much regulation”, and start asking “how much transparency”

The claim that regulation, as such, stifles innovation does not hold up once you look past the slogan. EO 14409 makes the point inadvertently: even a self-described “hand-off” regime ends up regulating once the risks get real; however, it did so opaquely, through classified benchmarks and unreviewable executive discretion, rather than through published rules and appeal rights. This is an unfavorable scenario for developers; rather than a stricter but transparent regime, it is a tacit concession that the precautionary instinct behind the EU AI Act was never the real problem.

Therefore, for future regulators building an AI framework from scratch, including in nascent regimes, the lesson is not to copy the “Brussels” or “Washington” effect, but to test any proposed framework against three concrete markers and ensure its integration. The markers are: public and published benchmarks and thresholds, an independent body to challenge adverse decisions and market access is never left to unchecked executive discretion. Failing to meet these three concrete components will hamper investment regardless of any innovation-friendly claims. A regime that passes them can be demanding, and still remain a credible foundation for innovation. That is one of the standards that nascent AI jurisdictions should borrow—not a label, but a checklist.

References

Almada, M., & Radu, A. (2024). The Brussels Side-Effect: How the AI Act Can Reduce the Global Reach of EU Policy. German Law Journal, 25(4), 646–663. https://doi.org/10.1017/glj.2023.108.

Bradford, A. (2025, August 18). Opinion | Trump Wants to Let A.I. Run Wild. This Might Stop Him. The New York Times. https://www.nytimes.com/2025/08/18/opinion/trump-ai-europe.html.

Controlling Advanced Artificial Intelligence: Executive Order 14409 Explained. (n.d.). [Legislation]. Retrieved September 22, 2026, from https://www.congress.gov/crs-product/IF13268.

Ebers, M. (2024). Truly Risk-Based Regulation of Artificial Intelligence—How to Implement the EU’s AI Act. SSRN. https://doi.org/10.2139/ssrn.4870387.

EU Digital Omnibus on AI Enters Into Force. (n.d.). Retrieved September 22, 2026, from https://www.hunton.com/privacy-and-cybersecurity-law-blog/eu-digital-omnibus-on-ai-enters-into-force.

Executive Order 14365—Ensuring a National Policy Framework for Artificial Intelligence · The Executive Record. (2026, August 19). The Executive Record. https://theexecutiverecord.com.

Hine, E., & Floridi, L. (2025). From No-Win to No-Lose: Journal of AI Law and Regulation, 2(3), 196–211. https://doi.org/10.21552/aire/2025/3/4.

Jozak, T. (2025). 2024 EU AI Act: A detailed analysis. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.5196856.

Khan, M. S., Poyer, A., & Khan, M. J. (2026). The institutional stack: Realigning U.S. AI governance. Telecommunications Policy, 50(9), 103294. https://doi.org/10.1016/j.telpol.2026.103294.

Orders, E. (2026, June 2). Promoting Advanced Artificial Intelligence Innovation and Security. The White House. https://www.whitehouse.gov/presidential-actions/2026/06/promoting-advanced-artificial-intelligence-innovation-and-security/.

PhD, A. M. (2026, July 13). Voluntary, With an Asterisk*. Medium. https://medium.com/@adnanmasood/voluntary-with-an-asterisk-dace7518a59e.

Rankin, J. (2025, November 7). EU could water down AI Act amid pressure from Trump and big tech. The Guardian. https://www.theguardian.com/world/2025/nov/07/european-commission-ai-artificial-intelligence-act-trump-administration-tech-business.

Smuha, N. A. (2019). From a “Race to AI” to a ’Race to AI Regulation’—Regulatory Competition for Artificial Intelligence. SSRN Electronic Journal. https://doi.org/10.2139/ssrn.3501410.

Systemic Regulation of Artificial Intelligence AI Safety Legal Paper Series 12-24  56 Ariz. St. L. J. 545. (n.d.).

Tillipman, J. (2026a). ‘Voluntary’ Until the Government Is Your Customer. Lawfare. https://www.lawfaremedia.org/article/voluntary–until-the-government-is-your-customer.

Tillipman, J. (2026b). ‘Voluntary’ Until the Government Is Your Customer. Lawfare. https://www.lawfaremedia.org/article/voluntary–until-the-government-is-your-customer.

Image Attribution

Generated by: ChatGPT 5.6

Date: 2026

Prompt: (Prompt was created by Claude Sonnet 5 Medium) Editorial photograph, conceptual, for a legal/policy blog header. A single vintage brass balance scale sits on a dark wooden desk, but one side holds a stack of clearly labelled, open legal documents catching warm light, while the other side holds a sealed black box with a padlock, in shadow. Muted, serious color palette—deep navy and charcoal tones with warm amber lighting on the documents side. Shallow depth of field, soft directional light from the left, shot on a 50mm lens, subtle film grain, no text overlays, no people, no visible national flags or logos. Aspect ratio 16:9.

Contact Us

FIll out the form below and we will contact you as soon as possible