Could as little as a one-hour workshop make AI safer?
We have a fairly good understanding of what makes AI trustworthy. Putting those principles into practice, however, is another matter. We spoke to RISE AI researchers Laurynas Adomaitis and Sepideh Pashami about the gap between ethics and technology development, and why making AI safer might start with something as simple as a conversation.
Just because an AI system works does not necessarily mean it can be trusted. It might perform exceptionally well under certain conditions but fail under others. And it can inspire confidence without necessarily deserving it.
So how do we build trustworthiness into AI systems from the outset, rather than trying to fix problems once they have already been developed? Laurynas Adomaitis answers the questions about ethics and design, while Sepideh Pashami contributes the perspective of technical robustness.
We've had guidelines for trustworthy AI for years. Why is it still so difficult to build systems we can trust?
– Researchers sometimes describe an operational gap between principles and practice. We can define what makes an AI system trustworthy and, to some extent, assess whether it meets those requirements. But that does not automatically tell us how to design and develop a system that delivers on them.
In 2019, the EU expert group on AI defined trustworthy AI as lawful, ethical and robust. Its ALTAI assessment tool helps organisations consider these principles, but cannot answer every practical design question.
This is exactly the challenge we are addressing in the European research project AIOLIA. We are exploring how ethical principles can inform development decisions, rather than only being assessed once a system is finished, says Laurynas Adomaitis.
Can an AI system seem more trustworthy than it actually is?
– Absolutely. Research into trust in automation suggests that our confidence is influenced both by how well a system performs and by how it presents itself. An AI system that is polite, offers detailed explanations or responds with confidence may appear more trustworthy even when its answers are incorrect.
This means we risk placing too much trust in systems that sound convincing, or too little in systems that actually perform well. The level of trust we place in a system needs to reflect what it can genuinely do.
– In practice, trustworthiness also depends on context. An AI system that works well in one situation may be less reliable in another. Understanding its limitations and what happens when conditions change is therefore essential, explains Sepideh Pashami.
How can we build trustworthiness into AI systems from the outset?
– One approach is known as ethics-by-design. It means integrating ethical considerations and safety into the development process, rather than trying to add them once the technology is finished.
An important part of this is anticipating potential risks, identifying ethical dilemmas and understanding the consequences of different technical choices early on.
One approach we are exploring is to bring together ethics experts, social scientists and technical developers for structured conversations. Looking at an AI system from different perspectives can reveal questions and risks that might otherwise go unnoticed. This may challenge assumptions and influence how the system is designed, continues Laurynas Adomaitis.
But isn't that complicated and time-consuming?
– Not necessarily! At RISE, we use methods that can be applied in workshops lasting just one hour. The approach is based on a structured dialogue between an ethics expert and technical experts, where they explore how ethical principles and guidelines apply to the particular AI system they are developing.
Most people today are aware that ethical principles for AI exist. The important thing is to encourage people to reflect on what these principles actually mean for their own work and the technical decisions they make. Since AI is used in so many different contexts, the questions also need to be adapted to each application.
– The best outcome for me is when someone from the development team or management says, "That's a good question, we hadn't thought of that!" That's when something new can emerge, an insight that actually influences how the system is designed, says Laurynas Adomaitis.
The aim is not to resolve every ethical question in an hour, but to create space for reflection and uncover issues that might otherwise be overlooked.
So perhaps the question is not just how much time we can afford to spend on ethics when developing AI, but what it might cost us not to.
AI ethics and trustworthy AI at RISE
At RISE, research in AI, ethics and technology development comes together to create AI systems that are safe, robust and responsible. Our work ranges from developing methods for identifying ethical risks and making better design decisions to creating trustworthy AI solutions for industry.
In the European research project AIOLIA, RISE explores how ethical principles can be turned into practical approaches throughout AI development. Sepideh Pashami and Jia Fu work within the SweWIN project on improving the resilience of generative AI models against adversarial attacks. Related research on DiffCAP is described in the RISE blog “We Are Making Large AI Models You Can Trust”:
