What does Generative UI mean for HCI Practice?

Source: Microsoft Research / UCSD / MBZUAI / Apple / MIT (CHI EA '26) · 2026-04

The increasing capability of AI models to generate user interfaces has the potential to transform HCI and design practice. We invite researchers, designers, developers, and practitioners to explore how generative UI – interfaces created by AI models – will reshape design methods, workflows, and user experiences. Our goals are to (i) envision how generative UI can underpin innovative humancentric experiences, and (ii) reflect on how HCI and design practice could and should evolve to meet the opportunities and challenges this presents. This will be an interactive and discussion-oriented workshop, featuring a pop-up panel, creative ideation exercises, and collaborative artefact development. Artefacts produced through the workshop will be shared online afterwards and will, we hope, result in an Interactions or CACM article. We will welcome submissions from scholars and practitioners working on dynamic or generative UI, as well as those with expertise in related areas. To keep participation broad, participants will be asked to submit a two-page position paper (in ACM single column format), a two-page pictorial, or a two-minute video at the workshop website. We expect approximately 35 participants to register and attend, including the organizers.

Introduction. Artificial Intelligence (AI) models are demonstrating a growing capacity to generate user interfaces (UIs), while also being endowed with greater flexibility and control to do this in real time for end-users. This allows interfaces to be created and adjusted dynamically by AI models at point of use, with the potential to support context-specific and personalized user experiences. This development presents considerable opportunities but also significant implications for HCI practice, as it suggests a trajectory where AI systems design and deliver UI directly to users. In contrast, HCI has built its foundation on developing tools and methods to support the crafting of interfaces by people. Practitioners seek to understand the needs and values of intended users, engaging in processes of ideation, testing, and iterative prototyping. They participate in teams that build their ideas into features or products, in the context of relatively stable software ecosystems. They evaluate those features or products and use that to inform further design. The technologies that derive from this process are generally consistent and predictable (indeed, these are core values within HCI), and people are involved in their production at every step of the way. In this workshop we will explore how HCI practice could and should change as AI models become increasingly capable of generating UI. Arguably, this growing capability will introduce a radical shift in the design roles we in the HCI community play, including the methods we use to influence the development of interactive systems. We ask: How will we, as HCI scholars and practitioners, continue to shape interactive experiences in ways that are humancentric? And, how will the workflows that produce technologies shift to incorporate UI generated by AI? In particular, if UI is generated and delivered directly to users by AI models, many of the traditional approaches HCI has adopted to ensure good design may need to be rethought. Processes of ideation, testing, and iteration may need to focus on what an AI model will generate at run-time, rather than on what a stable design that is to be ‘shipped’ should be. These new processes may need to complement broader changes to technology production workflows, which could become increasingly ‘model-centric’. Evaluation too may need to center AI models, focusing on the identification of metrics that support their development so that interactions between them and their users improve. We suggest that this calls for a need to both (i) understand what innovative AI-enabled experiences that meet users’ needs and values could be as models become more capable of generating UI, and (ii) reflect on and evolve our own practice. This workshop sets out to address these needs by providing a working space for design, discussion, and the production of artefacts, which will ground the building and sharing of knowledge. Attendees will include researchers, practitioners, designers, and developers. We will build on existing HCI knowledge as well as providing space for imagining future possibilities. We will envision how UI that is generated by AI models at run-time can change the user experience, and will consider how to conduct human-centric research and practice in a model-centric ecosystem. We begin this proposal by describing relevant work and proposing some working definitions1 that will ground questions for the HCI community. We then link this to the workshop’s goals and explain how the activities we plan will address them.

Related work. While UI generated by AI models garners attention, HCI already has a body of knowledge that explores ‘dynamic UI’, interfaces that are rendered to suit a user’s specific needs or to fit their context. Here, adjustments to the user interface can either be driven by the system (for this we use the term ‘adaptive UI’) or controlled by the user (as in ‘malleable software’). Adaptive UIs, which are tailored by the system for the user, their context, and/or their activity, have been explored in domains such as adaptive hypermedia, context-aware mobile apps, and intelligent tutoring systems. Early examples include SUPPLE and SUPPLE++, which optimized the interface for device constraints and users’ capabilities [16, 17, 18], PUC [31], a ‘Portable Universal Controller’ that supported consistency across different appliances, UNIFORM [32], which produced remote control interfaces by drawing on users’ previous experiences, and HUDDLE [33], which generated task-based interfaces across multiple appliances. In mixed reality, research has explored how to create interfaces that take into account user preferences [38], cognitive load, task, and environment [27], and environmental and social cues [26]. More recently, bio-physical signals have been leveraged to adapt UI, with the aim of supporting emotional wellbeing [24] and reducing cognitive load [1, 14, 15, 21]. Malleable software supports user-led customization at point of use. More specifically, Litt et al. [28] describe this as a software ecosystem where anyone can adapt their tools to their needs with minimal friction, giving users agency as co-creators. Examples include Bespoke [44], which enables users to generate GUIs to suit their own complex workflows, NL2INTERFACE [8], which allows users to generate a multi-visualization interface for data analysis, and LIDA [11], which supports the authoring of data visualizations and infographics. In DynaVis [43] dynamic widgets are generated and can be edited through natural language. There are lessons to be learnt from dynamic UI for the generation of UI by AI models. By personalizing content and workflow, dynamic UI can reduce cognitive load and cater for diverse needs. However, it can also introduce complexity, reduce predictability, require the collection of potentially sensitive data, and create challenges for users’ awareness of, and control over, possibilities for interaction. Additionally, its development can be methodologically complex. Machine learning (ML) technologies are increasingly being used in the design of dynamic UI (e.g. [42]), including to support the generation of new UI variations [45], to plan sequences of adaptations to the interface [41], and to ground adaptation decisions in specific contexts [19].

Method. A number of dimensions of dynamic and generative UI emerge from this overview of prior work, which could ground discussions of how genUI could impact HCI and design practice. These include:

1. Is the generated UI for a designer/developer or for the end-user? A key distinction for the HCI community lies between UI that is created to support the software design process, as reported by Lee [25] and Chen et al. [7], and UI that is produced in the general course of human-AI interaction. In the former, the designer/developer explicitly asks the model to produce UI, which feeds into a broader process whereby generated UIs are refined by practitioners before being experienced by endusers. Alternatively, AI can generate functional UIs directly at run-time, which are interacted with by end-users as they are generated. These point to the opportunity and need to alter HCI and design practice. 2. Is the UI constant or does it change? Prior work suggests variations ranging from static genUI (e.g. wireframes and visuals produced as part of the design process [25]); to interactive genUI (e.g., in dynamic prompt middleware [12, 34], genUI takes the form of interactive radio buttons that do not change once generated); to dynamic genUI, which can be changed by the user, or can change itself, during the interaction. 3. If the interface can change, are changes user-driven and/or system-driven? I.e., is the user in control of changes to the UI, as in the case of malleable software, does the UI adapt to the user and/or their context, or are both possible? 4. How long is the interaction timeframe? Is the generated UI produced to support an ephemeral interaction, where UI is generated and interacted with on the fly before the user moves on, or does it support a longer-lasting activity?

In this workshop we will consider the implications of genUI for HCI design and practice. Our goals are twofold: (i) to explore and imagine what sorts of experiences genUI should underpin to meet people’s needs and values; and (ii) to contemplate how HCI practice could and should change in order to support this. In relation to (i) and as indicated in the previous section, genUI can enable ephemeral experiences in which a UI is experienced only once, it can underpin much longer interactions through which the UI can change, it can give greater control to users to make those changes, and it can use complex reasoning to support systemdriven adaptations. The workshop will be a vehicle to ask: What sorts of experiences should genUI enable? How are those experiences related to existing HCI knowledge? And what challenges and research questions do these suggest? Critical and speculative design methods point to ways of envisioning what new user experiences could be [2, 13, 23]. Examples of existing relevant knowledge include that cited above, as well as related approaches including instrumental interaction [4, 35], dynamic abstractions [39], activity-centric computing [3], end-user programming [30], the use of design heuristics, guidelines, and patterns (e.g. [5]), and research into how to automatically transfer these [10] and make them malleable [29]. In relation to (ii), we aim to build a shared understanding of how genUI creates opportunities and challenges for HCI practice, how we can apply existing techniques to the design of genUI, and whether new techniques are needed. We will reflect upon:

Discussion. • The inclusion of genUI in the design process. This tends to augment existing design approaches through the generation of interfaces that feed into a design process led by skilled designers [7, 25], and raises questions about the opportunities genUI presents for design and HCI practitioners. • The democratization of UI generation, whereby the design and creation of UI is increasingly accessible to non-experts [7]. This ranges from people involved in software development, such as product managers, software engineers, and entrepreneurs, to end-users who could be increasingly capable of customizing and developing their own tools. This raises questions about how the software development process could be reimagined, and what roles HCI and design practitioners will play. • The potential for increased malleability of software. Together with the above, this suggests the possibility of workflows that deliver apps and tools intended to be customized by end-users, and raises questions about how HCI practice can ensure those customized experiences are valuable. • Its capacity to be central to human-AI interaction, as AI models are given the flexibility to generate novel interactive UI at point of use. This raises questions about how HCI practice can inform UI generated directly by AI models, be this through the application of design patterns and guidelines, or through ML techniques. An example of the latter is WHAM [22], a World and Human Action Model developed to support creative practice in gameplay design. To discuss these questions and address the workshop goals, we will encourage participation from diverse attendees, including researchers, practitioners, designers, and developers, those working on dynamic and genUI, and those with expertise in related areas.