Yes, the EU AI Act Covers Agentic AI—Here’s How – Unite.AI

0
3


There is a strong narrative right now that agentic AI and AI agents fall into some regulatory gap – that the EU AI Act is ill-equipped to govern this. We have read this in reputable Norwegian news outlets (Morgenbladet and Dagensnæringsliv), and repeated by leading academics (University of Oslo). We see two fundamental flaws in these claims… and we think they are based on fears, not on fact.

First, agentic AI is regulated under the EU AI Act.

Second, such claims undermine existing regulations. They reinforce the notion that AI companies are better positioned to manage the risks associated with their own systems – this is a common argument that leads to regulatory capture.

We work quite closely with the AI Act, as appointed members of Standards Norway’s national AI committee. We also sit on the European committees that develop harmonized standards for providers of high-risk AI systems; these standards help companies meet the AI Act’s legal requirements.

Agentic AI Is Regulated in Several Ways Under the EU AI Act

The European Commission has published several plain-language, easy-to-read FAQs regarding “agents” and “agentic AI”, and have clarified: yes, they fall within the scope of the regulation, quite simply because they meet the larger definition of an AI system (How are AI agents addressed within the AI Act? | AI Act Service Desk). Some more examples:

  • Agentic AI is frequently built upon large language models, which are regulated as general-purpose AI models.
  • When an agentic AI system interacts with humans, the regulation’s transparency requirements are also triggered.
  • If an AI model is classified as posing a systemic risk, additional requirements apply to the provider. These requirements include, among other things, assessments of the model’s capabilities – including its autonomous and agentic features.

The AI Act doesn’t mention the word “agent”, but this is because it does not regulate specific types of technology. Rather, it regulates what the technology is designed to do and the risks associated with those functions.

The AI Act Intends to Protect Our Fundamental Rights

A lot of AI Act criticism from those wanting more regulation, comes from the structural impossibility of perfect risk removal. But it would have been impossible for an international regulatory framework to anticipate every future risk.

Instead, the AI Act focuses on risks posed by AI systems that could affect health, safety, and fundamental rights – a point stated in the very first sentence of the first article.

What we have not highlighted in this first sentence is also crucial: the AI Act is market-oriented. It shall encourage the use of AI. It shall support innovation. The AI Act, in other words, is pro-AI. Within this context, certain applications and domains are regulated.

Certain Areas Are Pre-Classified as High-Risk

Some examples are AI systems used in:

  • recruitment
  • the allocation of essential services
  • medical devices
  • systems integrated into critical infrastructure where safety is paramount.

Providers of AI systems (whether agentic or otherwise) that could impact our rights are required to establish comprehensive risk management systems, including assessing how users may misuse the AI system.

Regulatory Capture: When Providers Propose Self-Regulation

If we listen to Anthropic, OpenAI, and Meta, as well as certain voices within the Norwegian debate, one might get the impression that AI and agentic AI are too complex for the AI Act. Too advanced for the EU. Too dangerous to be regulated by anyone other than the companies themselves.

We should be cautious when the world’s most powerful technology companies – which have lobbied hard for a less stringent implementation of the AI Act – now claim that the regulation is inadequate for governing their most potent systems. The solution they propose is to “regulate” themselves.

This is a classic mechanism of regulatory capture, a process where law and regulations are gradually shaped by the very entities they are intended to govern, for their own benefit.

The United States provides a clear example of this (Inaugurating The Era Of Super Intelligence – The White House). The White House Accord on Superintelligence is a “voluntary pact” with leaders from X, Meta, Anthropic, OpenAI, and Google, in which these companies promise to self-regulate. The document notes that laws might be appropriate, at some hypothetical point in the future: “Over time, it may make sense to codify these steps into laws or regulations.” Over time. Perhaps. Regulations may be reasonable to consider.

For now, however, Americans are expected to trust that these major tech companies will regulate themselves.

Unlike the United States, Europe Has a Binding Law: The AI Act.

And yet regulatory capture remains a risk, including through the erosion of confidence in this law. Because if tech giants can convince people that the AI Act isn’t good enough, we will not support its enforcement any more. Claims that agentic AI is imply too advanced/smart/dangerous for the AI Act to handle, are arguments that weaken our support for regulation.

Once regulatory capture takes hold, it is difficult to reverse.

We share the concern that some AI systems may lead to consequences that are difficult to foresee. Precisely for this reason, we need risk management procedures – as the AI Act also requires. (Dr. Muller works on the committee that is almost finished with developing the harmonized standard for risk management, 18228.)

The EU AI Act is not perfect, and it will likely need to be further developed. However, portraying agentic AI as an area outside of the EU AI Act only builds uncertainty, not safety.



Source link

LEAVE A REPLY

Please enter your comment!
Please enter your name here