My Manifesto for Human-Centered AI
This post outlines a vision for human-centered AI and human-AI interaction. AI is, or will be, a general purpose technology that will have a profound impact on the world. If we understand how to apply it. It is not good enough to ask AI itself, how we should apply itself. We as humans decide this.
But who do we decide what is “better”?
Ben Shneiderman has a great framework — the aim is to get as close as possible, to a system that has high levels of automation and high levels of human control. You decide the direction and can dial in how much automation you are need.
That is the vision I have for what AI could be. A system that is tuned in to how I work. A system I direct. Amplifying my abilities, while keeping what makes me unique as an individual.
AGI or ASI sound very mystical and magical — but if no one can actually use it, then the system is rather an anti-climax. The mission is to make AI useful by applying research from HCAI and HAI.
To this end, there are a few principles that I feel aid in this mission:
Principles
Human-AI Collaboration
I believe that the power to unlock AI’s potential, be in the form of LLMs or whatever comes after, is to create interaction patterns around collaboration between the user and the system. This would mean being able to dial in or dial back the usage of AI. It would adapt to your unique style. Automation might form a part of that, it might not. AI might not be a part of that solution at all. The goal is to have a system adapt to your unique style of doing things.
Human-Centered Benchmarks
Benchmarks guide development. Language models are compared on common benchmarks for different tasks. Developers train and evaluate the progress their model is making on these benchmarks. They are very task-oriented. That is not necessarily a bad thing. It does make for a rather one-sided picture though. By incorporating more collaborative measures that highlight not only outcome, but the journey to get there, we could create more meaningful interactions.
Open-Source & Open Research
Open-source software has shown how we can demonstrate value by sharing them. Solving problems, testing them and getting feedback is invaluable. Sharing research around the different areas that make up human-centered AI and human-AI interaction is key to distributing these ideas. Building up a library of design patterns and interaction designs that form the foundation of future products and services.
Safe, Reliable & Trustworthy
Ben Shneiderman put these 3 words down as the goals for developing these autonomous systems. Building safe systems means not only ensuring they do not cause harm, but also means considering the impact it has. Not just on the individual user. But also the team, the organisation and society at large. Reliability and trustworthiness tie into this. They influence the impact, good or bad, a system can have. Transparency and auditing these systems allows for safe, reliable and trustworthy deployment.
As we continue to integrate LLMs into our products and workflows, new and interesting interaction patterns will emerge. A strong UX foundation, built on HCI & HAI research will be the key to unlocking AI’s potential.
Thanks for reading!
Where do you see design playing a role?
Let me know what you think in down below.
P.S. Follow me on LinkedIn for more on UX and AI.
