What is human-centred AI?
No one likes systems that are rigid and force you to work in just one way. When such systems make counterintuitive decisions, people naturally avoid them or find ways to manipulate them – because they simply don’t trust them. So, how can an AI system be truly helpful? And what kind of AI systems do people want to rely on? David Geerts lists the three conditions for placing humans at the heart of any AI system: human control, transparency, and trust.
In 2017, two librarians of the East Lake Library in Florida (USA) created a fictional member of their library to check out books that were rarely borrowed. An AI system had recently been implemented to remove low-popularity books from the library, and the librarians wanted to prevent having to repurchase books that could become more popular again at a later time. Instead of being able to exert some control over the automated system, they had to resort to creative, but ultimately unauthorised, practices to keep those books in stock – something they got suspended for.
Unfortunately, there are a growing number of such examples where an automated AI system takes over control, at the expense of human agency. If we want to retain our humanity in the age of artificial intelligence, we need to design human-centred artificial intelligence based on three key principles: agency, transparency, and trust.
Human control & agency
As first principle of human-centred artificial intelligence, AI systems should offer at least some degree of human control and agency. As the example above shows, AI-based or automated systems are still prone to making errors, and likely will continue to be so in the future. How comfortable would you be in a fully automated car that wouldn’t allow you to intervene – even if you would be heading towards a wall? Even so, a high level of automation can – and will be – very desirable in many cases. It is possible to design AI systems with a high level of automation while still including human control, e.g. in the form of supervisory control, where many tasks are performed by AI, but humans can supervise and intervene if necessary.
Transparency
A second principle of human-centred AI is transparency. AI systems are often described as a ‘black box’, of which no one seems to know what is happening between giving input and receiving output. This way of describing AI seems to absolve the developers of any responsibility, as even the developers do not know what’s going on under the hood. Nevertheless, whether it is for movie recommendations (where the stakes are relatively low) or for a more high-stakes medical diagnosis, AI system transparency is a key element to making clear that its decisions and outcomes benefit its users. Transparency can be implemented at various levels. The most obvious, but difficult approach is to show how the system processes the input to reach an output, but that is often where AI systems become indecipherable. Transparency can also be achieved by, for example:
- showing how the system was trained,
- which data was used during its training,
- what its general operating principles are,
- or by sharing the level of certainty or uncertainty of the outcomes.
An AI system that explains its input, actions or output will ultimately be more human-centred.
Trust
Finally, a human-centred AI system is one that can be trusted. This third principle is multi-faceted and a result of complex interplay between different elements. For example, AI systems that are explained well and offer high levels of control, will be more trustworthy than AI systems with only one of these characteristics. But there is more at play in determining if we trust an AI system or not, depending on the type of system. Besides giving us control and being transparent, we trust AI systems that:
- do not harm us,
- are not biased,
- can be held accountable if something goes wrong,
- are implemented in a responsible way,
- with respect for our privacy and security.
Designing a trustworthy AI system therefore means paying careful attention to various elements at the level of the algorithm itself, the data that is being used as input, and the way it is implemented in practice.
Although these principles seem complex, they are not so difficult to apply, at least if developers are willing to involve users early in the development of AI systems. Through the practice of human-centred design, designers can discover user needs for specific AI-based applications, while creating and testing prototypes that are easy to control, explain how they work, and are ultimately trustworthy, as is also requested within the EU AI Act. Only then will we enter a future where AI is not seen as a threat to human activities, but as assistants that will augment our intelligence, rather than replace it.
Learn more at:
Human-centered AI
webinar - online - PUC KU Leuven Continue, VAIA
David Geerts
Dr. David Geerts is a freelance proposal manager, human-centered designer and trainer. He was Senior Research Manager of the KU Leuven Digital Society Institute (DigiSoc) in Belgium for three years, for which he organized and managed various events, facilitated internal and external collaborations, supported the submission of project proposals, and created national and international visibility. Before, he managed the Meaningful Interactions Lab (Mintlab) at KU Leuven for 17 years, while specializing in human-centered design and user experience of new technologies, including social media and interactive TV. David has published in various international conferences and journals and regularly gives presentations, tutorials and guest lectures on topics related to Human-Computer Interaction. In 2007, he co-founded u-sentric, an SME that provides CX and UX consultancy services.
Related courses
HPC Café
Seminarie - Online, Karlsruhe - KIT, HammerHAI, Training geselecteerd in de Belgische AI Factory Antenna (AIFA)
SAIAR Summer Studio: From latent to physical space
Zomerschool - Kortrijk - Howest Hogeschool, SAIARlab