Artificial intelligence is transforming how Irish businesses operate — from chatbots handling customer queries to AI tools that draft contracts, manage inventory, and analyse financial data. But as AI takes on more responsibility, a crucial question arises: how do we make sure it behaves safely?
A new research paper from scientists studying AI safety, titled “Learning Safe Agent Behaviour from Human Preferences and Justifications via World Models,” offers a compelling answer. It turns out the key to keeping AI under control is something we already have plenty of: human judgment.
The Problem with Letting AI Learn on Its Own
Teaching an AI system to make good decisions is harder than it sounds. The traditional approach involves giving the AI a “reward function” — a mathematical formula that tells it what outcomes are good and what are bad. But in the real world, defining “good” behaviour is surprisingly difficult.
Take a customer service chatbot, for example. You want it to be helpful, but you also want it to stay within your company’s policies, avoid making promises it can’t keep, and never share confidential information. Writing a mathematical formula that captures all of these constraints is nearly impossible. Something always gets missed.
This is where the new research makes a breakthrough. Instead of trying to program every rule into the AI, the researchers developed a method called DROPJ that lets humans teach AI what safe behaviour looks like — through simple preferences and explanations.
How It Works: Teaching AI Through Preferences
The approach is surprisingly straightforward. First, the AI explores a simulated version of its environment — a practice run where it can try different approaches without real-world consequences. A human then reviews the AI’s simulated actions and simply says which ones they prefer and, crucially, why.
For instance, if you were training an AI that handles customer complaints, you might prefer it to offer a refund before escalating to a manager. But more importantly, you’d explain why — because refunds resolve issues faster and keep customers satisfied. The AI learns not just “what to do” but “why it’s the right thing to do.”
The researchers found that this approach had two major benefits. First, it significantly reduced the computational resources needed to train the AI, because the human’s simulated practice runs were more efficient than random trial-and-error. Second, and more importantly, the AI’s actual performance improved when it learned from human preferences rather than hard-coded rules.
Why Safety Justifications Matter
The most interesting finding from the study is the role of “safety justifications.” When humans not only expressed a preference but also explained their reasoning about safety, the AI became noticeably better at avoiding risky behaviour.
This matters a great deal for Irish businesses. The EU AI Act, which came fully into force in 2024, places strict requirements on how AI systems are trained and deployed. Systems that can demonstrate they were trained with human safety oversight will be in a much stronger compliance position than those that were trained entirely automatically.
For example, if your business uses an AI tool to screen job applications, you need to be confident it isn’t introducing bias. Training the AI with human preferences and justifications around fairness gives you a documented audit trail — evidence that you took reasonable steps to ensure your AI behaves responsibly.
Practical Steps for Irish Business Owners
You don’t need to become an AI researcher to benefit from this approach. Here’s what it means for you:
1. Choose AI tools that involve human oversight. When evaluating software vendors, ask how their AI is trained. Tools that use human feedback in their training loop are likely to be safer and more reliable.
2. Document your human review process. Under the EU AI Act, demonstrating that a human was “in the loop” during AI training and deployment can be a significant compliance advantage.
3. Train your team to spot AI mistakes. The best safety system is a well-trained human who can recognise when an AI has gone off track and intervene.
4. Start with low-risk applications. Use human-supervised AI for tasks like drafting emails or summarising reports before moving to higher-risk areas like hiring or financial decisions.
The research is clear: the safest AI is the one that listens to humans. For Irish businesses navigating the new world of AI regulation, that’s a reassuring message.