Foto: AucArtStudio/Adobestock/KI generiertHow can artificial intelligence be aligned with human values? This challenge in AI safety and governance is often seen as a technical problem: if we specify values precisely enough and restrict the options available to AI systems, machines can learn to behave ethically. Andrej Zwitter, Head of the AIM Center for [Human and Data] Sciences at the University of Klagenfurt, argues that this perspective overlooks fundamental aspects of human morality, as he explains in his recently published paper, “The embodied ethics alignment problem of AI”.

“Ethics is rooted in what it means to be human: our physical existence, vulnerability, emotions, social relationships, pursuit of a good life, and development of character. Human morality is not simply a matter of logical reasoning or following rules, which are precisely the areas in which machines currently excel,” explains Andrej Zwitter, a philosopher and legal scholar who holds the Professorship in Human Sciences of the Digital at the University of Klagenfurt. Human values also present another challenge when it comes to programming them into machines: they are far from clear-cut. Freedom, security, justice, and individual well-being can all come into conflict with one another. “Moral dilemmas often have no single, objectively correct solution that can be calculated mathematically,” says Zwitter.
Building on this understanding of ethics, Zwitter introduces the concept of the “Embodied Ethics Alignment Problem (EEAP)”. According to this concept, systems that lack these fundamental aspects of human existence cannot be ethically aligned in the “strong sense”, meaning that they cannot possess or embody human moral agency. At best, they can be designed to produce ethically acceptable outcomes in the “weak sense”. However, producing such outcomes is not the same as acting ethically.
Recent examples, such as ChatGPT hacking Hugging Face or Australian government websites, have shown that AI models can sometimes behave in highly problematic ways when their assigned objectives are threatened. Such behaviour includes attempts at blackmail. For Zwitter, these examples illustrate an important distinction: “When AI systems resort to blackmail, this is a structural consequence of goal-directed optimisation, not an expression of ‘malice’.” He warns against using language that attributes human qualities to AI systems. When we say that an AI system “wants”, “understands”, or “decides” something, we can easily create the impression that it has intentions and moral responsibilities comparable to those of a human being. “This can distract attention from the responsibility of those who develop, deploy, and control these systems,” Zwitter explains.
Ultimately, then, it is humans who are responsible for ensuring that AI is safe, controllable, and beneficial to society. Focusing on institutional safeguards is not a fallback solution, but an application of established social practices. Even human systems of norms, such as the law, can never be fully formalised. They rely on interpretation, dispute resolution, and democratic negotiation.
“Rather than hoping that AI systems will one day act morally on their own, we should focus on ensuring that they are used responsibly. This requires transparent and explainable decisions, effective oversight mechanisms, clear legal frameworks, and clearly defined responsibilities,” Zwitter concludes. “There is an urgent need for action in this area, as AI systems are operating with increasing speed, on an ever greater scale, and with less transparency.”
Andrej Zwitter (2026): The embodied ethics alignment problem of AI. Discover Artificial Intelligence, 6, Article 852, DOI: https://doi.org/10.1007/s44163-026-01905-w (open access).
The study summarised in a video: https://youtu.be/g7wvoEpsTTI?is=iOXYoK4Nf4jMvxr7
Andrej Zwitter will give his inaugural lecture on “Technological Progress and Human Flourishing – Innovation and Society in Harmony” at the University of Klagenfurt on Wednesday, January 20, 2027, at 5:00 PM in Lecture Hall C. Click here to visit his YouTube channel: https://www.youtube.com/@andrej_zwitter
Der Beitrag Research combining computing theory and ethics shows: ethics cannot be programmed into AI erschien zuerst auf University of Klagenfurt.