AI learns to follow predefined norms through a combination of logic and machine learning

Artificial intelligence is becoming increasingly versatile—from route planning to text translation, it has long become a standard tool. But it is not enough for AI to simply deliver useful results: it is becoming ever more important that it also complies with legal, ethical, and social norms. But how can such norms be taught to a machine?

At TU Wien, a new approach has now been developed. By combining machine learning and logic, autonomous agents can be trained to follow predefined norms. It is even possible to establish a hierarchy of these norms—declaring some to be more important than others. At IJCAI 2025, an AI conference held this year in Montreal, Canada, this work was recognized with the Distinguished Paper Award.

Trial and error

Teaching AI new abilities sometimes works a bit like teaching tricks to a pet: reward if the task is performed correctly, punishment if the response is wrong. The AI tries out different behaviors and, through trial and error, learns how to maximize its reward. This method is called reinforcement learning and plays a key role in AI research.

“One could try to teach AI certain rules by rewarding the agent for following norms. This technique works well in the case of safety constraints,” says Prof. Agata Ciabattoni from the Institute of Logic and Computation at TU Wien. “But this wouldn’t work, for instance, with conditional norms (‘do A under condition B’). If the agent finds a way to earn a reward, it might delay finishing its actual job on purpose, to have more time for scoring easy points.”

Norms as logical formulas

The TU Wien team chose a fundamentally different path, inspired by old philosophical works: norms are still represented as logical formulas, but agents get a punishment when they do not comply with them. For example, “you must not exceed the speed limit” is translated as “if you exceed the speed limit you get a punishment of X.” Most importantly, each norm is treated as an independent objective.

“The artificial agent is given a goal to pursue—for example, to find the best route to a list of destinations. At the same time, we also define additional rules and norms that it must observe along the way,” explains Emery Neufeld, the first author of the paper. “The fact that each norm is treated as a different objective allows us to algorithmically compute the relative weight that we have to assign to these objectives in order to get a good overall result.”

With this technique, it becomes possible to encode even complicated sets of rules—for instance, norms that apply only under certain conditions, or norms that depend on the violation of other norms.

Flexible norms

“The great thing is that when the norms change, the training does not have to start all over again,” says Agata Ciabattoni. “We have a system that learns to comply with norms—but we can then still adjust these norms afterwards, or change their relative importance, declaring one rule to be more important than another.”

In their paper, Ciabattoni and her team were able to show that this technique allows a wide range of norms to be imposed, while the AI continues to pursue its primary goals.