Do Citizens Agree with the EU AI Act? Public Perspectives on Risk and Regulation of AI Systems

Paper Title

Do Citizens Agree with the EU AI Act? Public Perspectives on Risk and Regulation of AI Systems

Publication Info

• Topic area: Public perceptions of AI risks and alignment with EU AI Act regulations.
• Keywords: AI regulation, EU AI Act, public opinion, risk-based approach, AI systems, fundamental rights, public policy, horizontal regulation, AI literacy, international governance.

Background and Problem

• Problem / challenge: The EU AI Act categorizes AI systems into three risk levels (unacceptable, high, minimal) and applies corresponding regulatory measures. However, it is unclear whether this framework aligns with public perceptions of AI risks and regulation.
• Significance: Understanding public opinion is critical for ensuring democratic legitimacy and fostering trust in AI systems, as well as for informing future amendments to the AI Act and other regulatory frameworks.
• Motivation and related work: Prior studies indicate public support for AI regulation but do not address alignment with specific regulatory frameworks like the AI Act. Concerns exist that the Act may prioritize corporate interests over public values, and its risk-based approach may be difficult to operationalize.

Solution

• Proposed approach: A multi-country study capturing public perceptions of the risks posed by 48 AI systems and opinions on their regulation, specifically in the context of the EU AI Act.
• Novelty:
  1. Direct comparison of public perceptions with the AI Act’s risk-based framework.
  2. Analysis of public support for horizontal regulation versus tiered risk-based regulation.
  3. Examination of the perceived importance of AI Act requirements across all risk levels.
  4. Cross-country comparison of public opinions in Germany, Spain, France, and the US.
• Procedure and key techniques:
  • Participants (N = 1,421) from four countries evaluated 48 AI systems categorized as minimal, high, or unacceptable risk under the AI Act.
  • Participants rated perceived risks to fundamental rights and values, suggested regulatory measures (ban, regulate, or no regulation), and assessed the importance of AI Act requirements.
  • Mixed-effects regression and ANOVA were used to analyze the data.

Results

• Concrete findings:
  • All 48 AI systems were perceived as posing moderate risks (mean ratings between 2–4 on a 7-point scale) to fundamental rights, democracy, the rule of law, and the environment.
  • Participants rated all AI Act requirements as moderately to highly important, regardless of the AI system's risk level.
• Advantage over baselines:
  • Public opinion diverged from the AI Act’s tiered framework, with support for broader regulation of all AI systems rather than selective mandatory requirements based on risk levels.
  • Minimal differences in perceived risks between high-risk and unacceptable-risk AI systems, challenging the Act’s categorization.
• Experiments / evaluation:
  • Conducted in Germany, Spain, France, and the US, with balanced samples (approximately 300 participants per country).
  • Participants evaluated risks to 13 fundamental rights and values, regulatory preferences, and the importance of 10 AI Act requirements.
• Limitations and future work:
  • Risk assessments were one-dimensional and may not capture the multidimensional nature of risk.
  • Limited to four countries and focused on the EU AI Act; future work could explore global perspectives and alternative regulatory approaches.
  • Scenarios for prohibited AI systems were abstract, potentially influencing responses.

Summary

This study reveals a disconnect between the EU AI Act’s tiered risk-based framework and public perceptions of AI risks and regulation. Participants viewed all AI systems as moderately risky and supported comprehensive regulation across all risk levels, rather than selective mandatory requirements. The findings suggest public preference for horizontal regulation that establishes minimal standards for all AI systems. Cross-country comparisons showed consistent opinions, indicating potential for unified international AI governance. These insights challenge the adequacy of the AI Act’s risk-based approach and highlight the need for AI literacy and transparency to align public expectations with regulatory frameworks.