Artificial intelligence is becoming more versatile every day. From planning the fastest route to translating text, AI has become a standard tool in our lives. But there’s a challenge: it’s not enough for AI to just deliver useful results. It also needs to follow legal, ethical, and social norms the same way humans do. The question is: how can we teach machines to respect such norms?

Researchers at TU Wien (Vienna University of Technology) have developed a new approach that addresses this problem. By combining machine learning with logic, they have found a way to train autonomous agents not only to achieve their goals but also to follow predefined rules. Even more, they can establish a hierarchy of norms where some rules are considered more important than others. This groundbreaking work earned a Distinguished Paper Award at IJCAI 2025, a leading AI conference in Montreal, Canada.

Trial and Error: The Traditional Way

Teaching AI often works like teaching a pet:

• Reward when it does the right thing
• Punish when it does the wrong thing

This is known as reinforcement learning, a method where the AI learns through trial and error to maximize rewards.

For simple rules like avoiding dangerous actions this works well. But when it comes to more complex rules, such as conditional norms (“Do A, but only if B happens”), reinforcement learning falls short.

As Prof. Agata Ciabattoni from TU Wien explains:

“If the agent finds a way to earn rewards easily, it might ignore or delay its main task just to collect more points. That’s not real compliance with norms.”

Norms as Logical Formulas

To overcome this, the TU Wien team used a different approach. They drew inspiration from philosophy and logic. Instead of rewarding compliance, they represented norms as logical formulas—and applied punishments for breaking them.

For example:

• “You must not exceed the speed limit” → “If you exceed the speed limit, you get a penalty of X.”

Crucially, each norm is treated as its own objective, separate from the main task.

First author Emery Neufeld explains:

“The agent may have a primary goal, like finding the best route to a list of destinations. But alongside that, we define additional norms it must follow. By treating each norm as an independent objective, we can calculate how to balance them against the main goal.”

This allows the system to handle even complex sets of rules:

• Rules that apply only under certain conditions
• Rules that depend on the violation of other rules

Flexible Norms: Adapting Without Starting Over

One of the most exciting aspects of this approach is its flexibility.

If the set of norms changes, the AI doesn’t need to start its training from scratch. Instead, researchers can simply adjust or reorder the rules declaring some to be more important than others.

Prof. Ciabattoni highlights this advantage:

“We have a system that learns to comply with norms but we can later modify these norms or adjust their importance without redoing all the training.”

This research opens the door to creating trustworthy AI systems that don’t just perform tasks efficiently, but also act in ways that respect laws, ethics, and human values.

• In traffic systems, AI could prioritize safety rules over speed.
• In healthcare, AI could follow ethical standards alongside delivering accurate diagnoses.
• In robotics, machines could adapt to new social norms without retraining from zero.

The TU Wien team has shown that it’s possible to teach AI how to follow rules logically and flexibly a crucial step in making AI both useful and reliable in our daily lives.

For More information: Combining MORL with Restraining Bolts to Learn Normative Behaviour