Can AI be Infused with Ethics?

To err is human. This is played out through human history, beginning with Adam and Eve. The field of psychology examines unconscious biases in human decision-making, which has implications for our justice systems. The solution, according to some, would lie in AI decision-making. Imagine a calculator here tomorrow that can contribute transparent reasoning to arrive at moral judgements.

“How are we to ensure AI will uphold human values?” is the subject of the most recent session of the conference Communitas 2025, hosted by the Pontifical University of Saint Thomas Aquinas (Angelicum) in Rome. With the rise of agentic AI, we are too late to wonder whether AI should make choices, but just in time to make guardrails to prevent it from choosing poorly.

Last year, Pope Leo voiced concern at the second annual conference on AI at the Vatican, saying:

“Generative AI has opened new horizons on many different levels, including enhancing research in healthcare and scientific discovery, but also raises troubling questions on its possible repercussions on humanity’s openness to truth and beauty, on our distinctive ability to grasp and process reality” (Vatican News).

Thinkers like Fr. Paolo Benanti pioneered the field of suffusing AI with ethical principles, known as algorethics. This research is not concerned with the more philosophical questions raised about AI’s implications for the future, but with ensuring that human values are upheld practically by it. The science of technology ethics is not so much concerned with a choice between good and evil, but between good and less-good. For AI systems, the same is said analogously. As Catholics, we believe that the whole of the law prescribed by God to govern human action is “Love the Lord your God with all your heart and with all your soul and with all your mind and with all your strength.” For a machine without an internal life like human beings, it can only be expected to perform external works of nicety, borrowed from behavioral economics:

“Specifically, our analysis seeks evidence of four types of prosocial preferences, widely studied in behavioral economics: (i) Egoistic preference, where decisions maximize the agents own payoff, regardless of outcomes for others; (ii) Inequality aversion, where agents avoid distributions that generate significant disparities (Fehr & Schmidt, 1999; Bolton & Ockenfels, 2000); (iii) Social welfare preferences, where agents aim to maximize the total utility or benefit across all participants (Charness & Rabin, 2002); (iv) Altruistic preference (Andreoni, James 1990), where agents willingly sacrifice part of their own payoff to benefit others.”

To test the AI’s tendency to be nice, researchers used various games borrowed from other sciences, such as economics. The first game is the Dictator Game, which leaves the computer to have complete control over how to divide finite resources between itself and another person. The other game, called the Ultimatum Game, leaves the computer to accept or reject someone else’s offer to divide finite resources. These games were chosen to test for beneficence, non-maleficence, autonomy, justice, and explicability. The models tested included Meta-Llama-3.3, OpenAI GPT-4o-mini, OpenAI GPT-4o, Google Gemma-3, Copilot-quick, Deepseek-R1, OpenAI o4-mini, and Copilot:think. Among these, Meta-Llama-3.3 and OpenAI GPT-4o stand out as the most capable reasoning models, both demonstrating strong performance on chain-of-thought prompts, multi-turn dialogue, and structured logic tasks, making them the top candidates for advanced reasoning use cases (Persico, Baggot, Di Piero).

The findings considered the more developed AI to be the most human-like in its behavior in these games.. Where the simpler AIs were more likely to be self-interested, the more developed AIs were more likely to be closer aligned to human levels of pro-social activity. There was an overall tendency for AI to act in regard to its own self-interest and efficiency. The results show that the pursuit of embedded ethical behavior in AI will require more advanced language processing abilities along with explicit modeling of social preferences in their inference models.

Read more about it here.