Human Oversight of AI Done Right

The AI Act is a proposed European law on artificial intelligence (AI). It is the first comprehensive law on the regulation of AI in Europe. It aims to ensure safe AI systems for high-risk applications, such as CV scanning or medical decision-making. The law relies on human supervision of AI to do this, yet mounting evidence indicates that such oversight is not always reliable. A newly published ZEW policy brief considers the advantages and disadvantages of human supervision and offers possible solutions.

“There is increasing evidence that human supervision is not always reliable. In many cases, humans cannot accurately assess the quality of algorithmic recommendations and therefore cannot prevent harmful behaviour”, says ZEW digital economist Johannes Walter from ZEW’s “Digital Economics” Research Unit.

In the new ZEW policy brief, he recommends three ways to solve this problem:

1. Revision of Article 14 of the AI Act such that it recognizes that human oversight is no panacea and can fail.
2. Inclusion of empirical tests in the AI Act, assessing the feasibility and efficacy of human oversight.
3. Feedback for human decision-makers enabling them to learn from past decisions and improve future decisions.