Exploring AI as a design material
To design for AI, we need new tools and new ways of working. In AI Studio, we’ve made a design tool to help us explore new interactions. We’re calling it our Chat ‘Playground’.
It allows our interdisciplinary team of designers, product managers, and data scientists to collaborate and explore alternative AI interactions without writing code. The team can quickly model new behaviours, uncover interaction patterns, and observe how an AI system behaves in different situations.
It also helps us find implementation challenges earlier, and brings multi-disciplinary teams together around a shared ideation space.
Using AI as a design material
Conversational AI is a design problem as much as it is an engineering one. Design choices matter as much as technical ones.
Working with dynamic prototypes helps us see where interactions work, where they break down, and what new questions they raise.
For example:
- what are the trade-offs between generic and personalised experiences?
- when is general information irrelevant, or when does personalisation require too much input from the user?
- how might we allow free exploration and still work towards task completion?
To explore these behaviours responsibly, we are designing by building. Then, as a team, we can discuss the pros and cons of the patterns that emerge. Things like, how might we offer different levels of participation for those who want it? Or how should UI and dialogue work together?
Our Playground is a shared space for us to make tangible design decisions and try them out. These then become ‘seed’ ideas for deeper exploration and testing.
Playground v0.1
Our Playground is a custom model that includes relevant content. It uses system instructions to manage conversations by understanding and keeping track of the user’s goals — which it records on its own internal notepad or ‘scratchpad’.
It also has access to various resources that give it additional ‘skills’ when it needs them, for example transactional flows, templates and guidelines.
The Playground is a chat assistant that helps people access information and complete simple tasks.
The AI makes notes about the goals of the conversation, and when relevant, it can choose to follow certain rules like step-by-step instructions (flows) or structured responses (templates).
By building our Playground in this modular way we can easily add experimental functionality, like new flows, templates or dialogue patterns, without needing to touch any code.
Plus, we can easily see how the AI is behaving by analysing its responses and checking its internal ‘scratchpad’.
This prototype is our first attempt at building a Playground. We’re now working on a more permanent prototype with access to more UI components and functionality.
We’ve started running sessions with data scientists to see where this design thinking can inform the system architecture. For example, we’ve looked at how the system can know which is the most relevant service or how to hand off to a human at the right time.
We’ll be working with departments across government to develop common behaviours and interactions that we can use to provide reusable templates for implementation and evaluation.
Moving from answers to actions
This experimental design work supports GDS’s wider aim to create a new generation of personalised, joined-up public services.
We’re currently using the prompts, context, flows and guidelines from our Playground as a starting point for a number of different projects across GOV.UK AI, including GOV.UK Chat.
Some things we’re exploring are:
- completing simple transactions in Chat using step-by-step instructions
- generating dynamic personalised journeys based on Chat context
- handing off to the right departmental support when needed
- modelling data-rich transactions where privacy is especially important
We also want these interactions to feel natural, transparent and consistent across departments. Our Playground is a place for us to start shaping these common patterns.
Get in touch
We’re looking for departmental partners to collaborate on AI experimentation. If you’d be interested in working with us, or to discuss our explorations so far, you can email us at govuk-ai@dsit.gov.uk.