AI rules too ambiguous, says architect of EU’s AI Act

TAIPEI (Global Views Monthly/TVBS News) — Teenagers spend sleepless nights conversing intimately with what they believe to be romantic partners, only to discover that the entity responding in their chat window is artificial intelligence (AI). This scenario, reminiscent of the film "Her" (雲端情人), is no longer fiction; AI has indeed become a trusted confidant for many. But this trend has a dark side: in the United States, AI stands accused as an indirect accomplice in the suicide of a 14-year-old boy.

It has been two years since ChatGPT emerged unexpectedly, sending shockwaves through every industry. Many people's lives have already become intimately connected with AI. But as AI infiltrates our daily routines, its immense potential is accompanied by the undeniable downsides of this technological double-edged sword: AI hallucinations, misinformation, copyright violations, hate speech, bias, and discrimination controversies have all surged to the forefront.

As a result, stakeholders across society began contemplating how to regulate and master this technological colossus. The world's first comprehensive AI legislation — the "European Union's Artificial Intelligence Act" (EU AI Act, 歐盟人工智慧法案) — emerged in response to this need. From its formal proposal, the legislation underwent a three-year journey before finally coming into effect in August 2024. Going forward, all AI products serving the EU market will need to comply with these regulations. This legislation deliberately aims to be the first to apply the brakes to the rapidly spinning global AI machinery, establishing a historic milestone in humanity's quest to master technology rather than be mastered by it.


When ChatGPT Upended the Regulatory Drawing Board
Global Views Magazine talked to Gabriele Mazzini (馬齊尼), the behind-the-scenes architect of this unprecedented legislation, to discuss his emotional journey through the bill's creation. If one includes the incubation period, the birthing process of this legislation actually stretched across five years. Throughout this time, Mazzini and his team immersed themselves in regulatory work day and night, yet even after the AI Act was finalized, controversies remained abundant, leaving him deeply disappointed. With a mixture of profound attachment to the legislation and acute criticism of its flaws, Mazzini — who led the EU AI Act proposal and drafting team — made the somber decision to depart from his position at the EU.

To comprehend how the legislation was structured, one must examine the context in which it was crafted — a context that mirrors the evolution of AI itself. In 2019, Mazzini, who holds a master's degree from Harvard Law School, was tasked by the European Commission with preparing AI legislation. He adopted a horizontal regulatory approach with four "risk levels." The following year, his team drafted the first white paper, but when compared with the current final version, substantial differences emerge despite maintaining the same general direction.

A pivotal turning point came in November 2022 with the debut of ChatGPT. Large language models, with their wide-ranging application capabilities, became known as "general-purpose AI," challenging the draft legislation's original regulatory approach that focused on "specific application scenarios." Because of this, the legislative team was forced to add an entire new chapter to the bill. "Without ChatGPT, the chapter on general-purpose AI models (GPAI) would never have existed," Mazzini said.

Then in 2023, the non-governmental organization Future of Life Institute (生命未來研究所) launched an open letter signed by nearly a thousand individuals, calling for a pause in the training of AI models more powerful than their predecessors. In October of that same year, President Joe Biden signed an "AI Executive Order" that incorporated specific regulations for foundation models. These developments further reflected society's anxiety about AI risks and prompted the EU to incorporate the concept of "systemic risks" into the "AI Act" — imposing stricter regulations on foundation models with greater computing power and correspondingly higher risks.

This demonstrates how the legislative team evolved with the times, attempting to keep pace with the lightning-fast development of AI technology. The ultimate inclusion of chapters regulating general-purpose models enabled the "AI Act" to align with today's most influential AI applications.


Regulatory Ambiguities Cloud the Landmark AI Legislation
Nevertheless, a series of controversies persist, including the regulatory emphasis on downstream application providers rather than the model suppliers who wield greater power and capital. Moreover, Ho Chih-hsing (何之行), an associate research fellow specializing in law at the Institute of European and American Studies, Academia Sinica, points out the paradoxical threshold for determining GPAI systemic risk. According to current computational training criteria, GPT-4 is classified as a model with systemic risk requiring testing and evaluation, while GPT-3.5 is exempt. What justifies this distinction?

The law's ambiguities extend well beyond this issue. Mazzini candidly admits that under political pressure and parliamentary scheduling constraints, the legislation was hastily and carelessly drafted, resulting in a final version that is both complex and divergent from his initial vision. Even as the law's drafter with intimate knowledge of its content, he remains perplexed by certain sections, saying, "I'm writing a paper on the AI Act, and when rereading the provisions, I can't help but think, 'How can this possibly be reasonable?'"

For instance, conflicts exist between the "Act" and the EU's existing "General Data Protection Regulation" (GDPR, 一般資料保護規則). When AI systems learn autonomously, how does one define whether a "substantial modification" has occurred? The identification of product "safety components" also lacks clear standards. These unresolved critical issues not only complicate corporate compliance efforts but also introduce greater uncertainty into the future of AI regulation.

"The job of regulatory bodies is to establish rules and maintain clarity, not to create ambiguity that preserves power for themselves," Mazzini emphasizes repeatedly. Sound regulation, he insists, does not inhibit innovation; rather, it is laws that are difficult to interpret and follow that slow investment. "The AI Act contains numerous gray areas that will make compliance especially challenging for small and medium-sized enterprises and individuals without legal expertise."

These myriad issues further highlight the Herculean task of crafting AI legislation. Mazzini shares that the day he officially "logged out" from the legislative team coincidentally fell on the same date he had joined the European Commission seven years earlier. This coincidence prompted him to publish reflections on his journey — how the legislation was "named," his assessment of its strengths and weaknesses, and the delicate negotiations with various stakeholders. This unexpectedly arduous journey has profoundly impacted both him and the world. "I still consider this the most challenging and influential experience of my professional career," he says.

Scholars Doubt Europe's AI Rules Will Set Global Standards
Through this globally unprecedented AI legislation, the EU demonstrates an AI governance and regulatory framework aligned with European values. The EU hopes to leverage its market influence to once again exert the "Brussels Effect" (布魯塞爾效應), influencing international efforts to strengthen protections for AI system users. The Brussels Effect refers to the EU's ability to influence the behavior of non-EU countries and companies through its market size and stringent standards, thereby shaping global regulatory frameworks. The GDPR, for instance, has profoundly influenced personal data protection standards worldwide.

However, to persuade international participants to follow suit, the EU must first lead by example, demonstrating that this legislation can function effectively while striking a balance between risk regulation and nurturing innovation. The path to achieving these goals remains long and arduous. From developing guidelines for prohibited use cases and best practice recommendations to coordinating consistent implementation among member states' authorities, the EU's "AI Office" — expected to employ more than 140 staff members — faces an enormous and diverse workload. Isabel Hou (侯宜秀), secretary-general of the Taiwan AI Academy (台灣人工智慧學校基金會) and a legal specialist, believes implementing this framework within two years will be extraordinarily challenging.

According to a Future of Life Institute compilation, the European Commission and the AI Office still face 130 pending tasks. Among these, 39 require the development of secondary legislation! Hou further points out that AI legislation itself presents significant challenges. At a fundamental level, AI technology is rapidly evolving globally, making it difficult to define an "AO system" clearly, let alone establish precise regulations for various models, outputs, and application scenarios.

Second, methods for implementing value alignment and testing remain underdeveloped. For instance, regulations require fairness and non-discrimination in AI models, but how should these qualities be measured? With what tools? And which value system should define "fairness"? Currently, as AI advances toward becoming a general-purpose technology, its integration into services and products is becoming increasingly diverse and complex — far beyond the previous generation of "specialist" AI systems focused on single tasks. When a product or service deploys multiple models and can collaboratively perform diverse tasks spanning different risk levels for consumers, which regulations should apply to such products or services?

These numerous unresolved ambiguities have led Hou to state candidly that she "does not expect this legislation to produce a Brussels Effect." She reveals that European experts she has consulted also acknowledge that the legislation represents the result of negotiations and compromises among 27 member states, and they do not recommend that other countries adopt it wholesale. Observing the international influence of the EU AI Act, Taenrat Khunngern (張粱), director of the Institute of Economics and International Trade Law at Khon Kaen University (孔敬大學) in Thailand, believes that Eastern and Western social contexts differ significantly, and EU perspectives may not apply to other countries. He points out that in terms of market size, the Asia-Pacific region has a larger population than Europe with formidable purchasing power. Asia's approaches to AI governance, he suggests, also merit Western observation.


A Global Quest for Better Solutions to AI Governance
"AI is a global issue," Khunngern says. "Therefore, perspectives from all countries should be incorporated into the conversation." Khunngern, who has conducted research in the United States, Australia, and China, observes that the inability of AI laws to keep pace with technological advances has become a scholarly consensus. He believes that at the current stage, issuing guidelines supplemented by each country's common law might be more effective than creating specific AI legislation.

But Europe has already taken the brave first step in AI legislation, and the world is now watching its performance. Khunngern notes, "Europe excels at going from 0 to 1, while Asia excels at going from 1 to 1+." If the EU AI Act functions effectively, Asian nations can learn from it and develop legislative frameworks better tailored to their own contexts. This process requires time. As AI-related legal precedents accumulate, the trajectory of issues such as intellectual property rights is gradually becoming clearer. Meanwhile, international organizations like the Organization for Economic Cooperation and Development (OECD) and the Group of Seven (G7) are issuing their own guidance. The future may see the emergence of independent international AI regulatory organizations tasked with harmonizing regulations across countries.

To confront the rapidly evolving landscape of AI, governance frameworks must become increasingly flexible. The efforts of countries, regions, and organizations to secure influence in shaping international governance frameworks that better align with their own advantages and interests are only just beginning. Mazzini continues to travel the world, engaging with scholars globally to discuss the AI Act, hoping to discover more comprehensive solutions. He urges the EU to concentrate its resources on addressing these core issues rather than becoming entangled in unproven risks, ensuring that the AI Act truly benefits the AI industry and community in Europe and globally. "The AI Act represents uncharted territory where we must create something entirely new," he said. "Problems persist, but so do opportunities."
 
For More Reading
This article is excerpted from the January issue of Global Views Monthly; for more articles, please visit the Global Views website 
Here is the link to the Chinese story: 獨家專訪》世界首部AI法案推手，他為何仍憂心AI監理的未來？

More coverage:
1. 開箱全球第一部AI法案！規範了什麼？後續影響為何？
2. 不碰錢就沒事？資安專家：金融業導入LLM小心雙面刃
3. AI產品想賣進歐洲？歐盟新法上路後，這標誌不能少！